Predictive Artificial Intelligence Models for Non-Communicable Disease Burden Forecasting in Africa

doi:10.21203/rs.3.rs-8065115/v1

Background:

Non-communicable diseases (NCDs) are rapidly emerging as the dominant cause of death in Africa, projected to surpass infectious diseases by 2030. Artificial intelligence (AI) and machine learning (ML) tools offer significant potential for forecasting disease burden, enabling early interventions and data-driven health system responses. However, the landscape of AI-based predictive modeling for NCD forecasting in Africa remains poorly defined. This study systematically mapped existing literature on AI and ML models used for NCD burden prediction across the continent to identify methodological trends, geographical coverage, and research gaps.

Results:

A total of 127 studies met inclusion criteria from 8,547 screened records. Most studies (68.5%) developed individual-level risk prediction models rather than population-level burden forecasts. Cardiovascular diseases (42.5%) and diabetes mellitus (38.6%) dominated the evidence base, while cancer (10.2%) and chronic respiratory diseases (8.7%) were underrepresented. Research activity was highly concentrated in South Africa, Kenya, and Nigeria, with 32 African countries unrepresented. The majority of models used supervised learning techniques such as random forests, support vector machines, and gradient boosting methods. Only 23.6% of studies applied temporal or longitudinal forecasting approaches, and external validation was reported in just 18.1% of cases. Most datasets were hospital-based (51.2%), with limited integration of multiple data sources to improve generalizability.

Conclusions:

AI-driven predictive modelling for NCD forecasting in Africa is expanding but remains methodologically fragmented and geographically uneven. The dominance of cross-sectional prediction over temporal forecasting limits its value for long-term planning. There is an urgent need for collaborative, multicounty research that incorporates diverse African data, standard validation frameworks, and integration into health system infrastructures. Strengthening local capacity for AI development and governance will be essential to ensure equitable and context-appropriate digital health innovation across the continent.

Artificial intelligence

machine learning

non-communicable diseases

disease forecasting

predictive modelling

Africa

digital health

health informatics

1.1 Background and Rationale

Non-communicable diseases have emerged as one of the most formidable public health challenges confronting the African continent in the twenty-first century. Historically, the disease burden profile in sub-Saharan Africa (SSA) has been predominantly characterized by communicable diseases, maternal and child health conditions, and nutritional deficiencies [1, 2]. However, a profound epidemiological transition is currently underway, driven by rapid urbanization, demographic shifts toward aging populations, and widespread adoption of lifestyle patterns associated with increased NCD risk factors [2, 3, 4]. Recent evidence from the Global Burden of Disease Study 2017 documented a 67.0% increase in disability-adjusted life-years (DALYs) attributable to NCDs in sub-Saharan Africa between 1990 and 2017, with NCDs now accounting for 29.8% of the total disease burden [5].

The four principal NCD categories, cardiovascular diseases, cancers, chronic respiratory diseases, and diabetes mellitus, collectively account for more than 80% of all premature NCD deaths globally [6]. Within the African context, cardiovascular diseases constitute the second leading cause of NCD burden, responsible for 22.9 million DALYs in 2017, representing 15.1% of the total NCD burden [14]. Diabetes mellitus presents a particularly concerning trajectory, with late diagnosis and inadequate glycaemic control frequently culminating in severe complications [7]. Mental health and substance use disorders, chronic kidney disease, and other NCDs contribute substantially to morbidity, yet remain inadequately addressed by existing health systems [8, 9].

The implications of this epidemiological transition extend far beyond health outcomes, threatening to undermine socioeconomic development gains across the continent. NCDs impose catastrophic healthcare expenditures on households, disproportionately affecting vulnerable and socially disadvantaged populations with limited access to quality health services [10, 11]. If current trajectories persist unabated, NCDs are projected to surpass communicable, maternal, neonatal, and nutritional diseases combined as the primary cause of mortality in sub-Saharan Africa by 2030 [2, 6]. This prospective shift necessitates a fundamental reorientation of health systems that have historically prioritized infectious disease control and maternal-child health interventions [5, 6].

1.2 The Role of Artificial Intelligence in Health Systems Strengthening

Artificial intelligence and machine learning technologies have demonstrated transformative potential across multiple domains of healthcare delivery, from diagnostic enhancement and treatment optimization to resource allocation and health system management [1, 3, 10]. AI encompasses computational systems capable of performing tasks traditionally requiring human intelligence, including pattern recognition, predictive analytics, natural language processing, and decision support [12]. Within the healthcare context, machine learning algorithms, which enable systems to learn from data without explicit programming, have proven particularly valuable for analysing complex, high-dimensional datasets to identify patterns and generate predictions [13].

The application of AI technologies to NCD management and prevention is expanding rapidly across Africa. Recent initiatives have demonstrated the utility of AI-powered diagnostic tools for early detection of conditions including tuberculosis, malaria, and diabetic retinopathy [3, 10]. In Kenya and Nigeria, AI-enhanced systems have improved disease detection accuracy, significantly reducing mortality rates [14]. Agricultural applications, including AI-driven precision farming and weather forecasting, have addressed food security challenges affecting over 60% of Africa's population, with implications for nutrition-related NCD risk factors [15].

The African AI market, valued at approximately US$1.2 billion in 2023, is projected to reach US$7 billion by 2030, reflecting accelerating investment and growing technological capacity [3]. Countries including South Africa, Kenya, and Nigeria are leading AI adoption, with South Africa's National Artificial Intelligence Strategy providing a comprehensive framework for innovation addressing socioeconomic challenges [3]. However, the potential economic contribution of AI to Africa's economy, estimated at US$2.9 trillion by 2030, remains contingent upon addressing persistent infrastructural, regulatory, and capacity-building challenges [3, 6, 8].

1.3 Predictive Modelling for NCD Burden Forecasting

Predictive modelling for disease burden forecasting represents a specialized application of AI/ML technologies focused on estimating future disease incidence, prevalence, mortality, and associated health system impacts at population levels. Unlike clinical prediction models that assess individual patient risk, burden forecasting models aim to inform strategic health planning, resource allocation, and policy development [12]. These models integrate diverse data sources, including demographic projections, epidemiological trends, risk factor prevalence, socioeconomic indicators, and health system capacity metrics, to generate actionable intelligence for decision-makers.

The complexity of NCD epidemiology, characterized by multifactorial aetiology, prolonged latency periods, interaction effects among risk factors, and socioeconomic determinants, renders traditional statistical forecasting approaches limited in predictive accuracy [5, 12]. Machine learning techniques, particularly deep learning architectures and ensemble methods, offer enhanced capacity to capture non-linear relationships, temporal dependencies, and complex interaction patterns inherent in NCD data [11]. Recent advances in time-series forecasting, including recurrent neural networks and transformer architectures, have demonstrated superior performance in predicting disease trajectories compared to conventional regression-based approaches.

1.4 Knowledge Gaps and Review Objectives

Despite the proliferation of AI applications in healthcare and growing recognition of the NCD crisis in Africa, comprehensive synthesis of evidence regarding predictive models specifically addressing NCD burden forecasting in African contexts remains absent from the literature. Existing reviews have examined AI applications in specific disease categories, focused on clinical risk prediction rather than population-level forecasting, or addressed global contexts without specific consideration of African implementation challenges and opportunities [11, 13, 16]. The unique sociodemographic characteristics, health system configurations, data infrastructure limitations, and cultural contexts of African settings necessitate dedicated examination of AI/ML applications in this geographical region.

Understanding the current state of evidence regarding AI-driven NCD forecasting in Africa is essential for multiple stakeholder groups. Policymakers require evidence to inform strategic investments in digital health infrastructure and AI capacity building. Researchers need comprehensive mapping of existing work to identify priority areas for future investigation and avoid duplication of efforts. Health system managers must understand available tools and their validation status to guide adoption decisions. International development partners seek evidence to direct technical and financial support effectively.

1.5 Review Aims and Research Questions

This scoping review was designed to systematically map and synthesize the existing body of literature on predictive AI and ML models for NCD burden forecasting in Africa. Specifically, the review addressed the following research questions:

What is the extent, nature, and distribution of published literature on AI/ML predictive models for NCD burden forecasting in African populations?

Which specific NCD categories (cardiovascular diseases, diabetes, cancer, chronic respiratory diseases, others) have been addressed through predictive modelling studies?

What geographical distribution characterizes existing research across African countries and regions?

What AI/ML methodological approaches and algorithms have been employed for NCD burden prediction?

What data sources, including their quality, accessibility, and comprehensiveness, have been utilized for model development and validation?

What temporal dimensions (cross-sectional risk prediction versus longitudinal burden forecasting) characterize existing models?

What validation strategies and performance metrics have been employed to assess model accuracy and generalizability?

What critical knowledge gaps exist in the current evidence base, and what are the priority areas for future research?

A scoping review methodology was selected as most appropriate for addressing these research questions, given the objective of mapping a broad and heterogeneous literature base, identifying knowledge gaps, and examining the nature of evidence rather than evaluating intervention effectiveness [16, 17, 19]. This approach enabled comprehensive inclusion of diverse study designs, methodological approaches, and publication types while maintaining methodological rigor through systematic search strategies and transparent reporting.

2.1 Protocol Development and Registration

This scoping review was conducted following the methodological framework established by the Joanna Briggs Institute (JBI) for scoping reviews and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines [20, 21, 22]. The review protocol was developed a priori, specifying all methodological procedures including search strategies, eligibility criteria, screening processes, data extraction frameworks, and synthesis approaches. The PRISMA-ScR checklist was completed to ensure comprehensive reporting of all recommended items [23].

2.2 Eligibility Criteria

Eligibility criteria were established using the Population, Concept, and Context (PCC) framework recommended for scoping reviews, as articulated in Table 1.

Table 1
Inclusion and Exclusion Criteria using PCC Framework
Criterion	Inclusion	Exclusion
Population	Studies examining NCD burden in African populations (all 54 African countries); Studies addressing any age group; Studies including major NCDs (cardiovascular diseases, diabetes, cancer, chronic respiratory diseases) and other NCDs recognized by WHO	Studies exclusively examining non-African populations; Studies addressing only communicable diseases, maternal health, or nutritional conditions
Concept	Studies employing artificial intelligence, machine learning, deep learning, or other computational predictive modeling approaches; Studies focused on forecasting, predicting, or estimating future NCD burden (incidence, prevalence, mortality, DALYs); Studies developing, validating, or implementing predictive models	Studies employing only traditional statistical methods without AI/ML components; Descriptive epidemiological studies without predictive modeling; Clinical trials, observational studies, or interventions without forecasting components; Diagnostic AI tools without burden forecasting
Context	Studies conducted in African settings (any of the 54 African countries); Studies using African population data; Multi-country studies including African countries; Studies addressing health system planning, policy, or resource allocation	Studies with no African population or context; Studies focusing exclusively on non-African settings
Study Design	Peer-reviewed journal articles; Conference proceedings and abstracts; Grey literature including technical reports, dissertations, and policy documents; Studies published in English, French, or Portuguese; Studies from database inception to October 2025	Editorials, commentaries, letters, or opinion pieces without original data; Study protocols without results; Systematic reviews or meta-analyses (included for reference mining only)

Table 1. Inclusion and exclusion criteria for study selection

Eligibility criteria defined using the Population, Concept, and Context (PCC) framework for scoping reviews. Studies were included if they examined non-communicable disease burden in African populations, employed artificial intelligence or machine learning predictive modeling approaches, and focused on forecasting or predicting future disease burden. Studies were excluded if they addressed only communicable diseases, employed traditional statistical methods without AI/ML components, or had no African population or context.

2.3 Information Sources and Search Strategy

A comprehensive, systematic search strategy was developed in collaboration with an experienced health sciences librarian. The search was executed across eight electronic databases selected for their coverage of health sciences, computer science, and African-focused literature: PubMed/MEDLINE, Scopus, Web of Science, IEEE Xplore Digital Library, Google Scholar (first 200 results sorted by relevance), African Journals Online (AJOL), CINAHL (Cumulative Index to Nursing and Allied Health Literature), and Embase. The search encompassed all records from database inception through September 25, 2024.

The search strategy employed a comprehensive combination of controlled vocabulary terms (MeSH terms for PubMed, Emtree terms for Embase) and keyword searches addressing three primary concept domains: (1) artificial intelligence and machine learning technologies, (2) non-communicable diseases, and (3) African geographical context. Boolean operators (AND, OR) were utilized to combine search terms within and across concept domains. Search strategies were adapted appropriately for each database while maintaining conceptual consistency.

The core PubMed search strategy is presented in Appendix A. Key search terms included:

AI/ML domain: "Artificial Intelligence" [MeSH], "Machine Learning" [MeSH], "Deep Learning" [MeSH], "Neural Networks, Computer" [MeSH], "algorithms," "predictive model*," "forecasting," "prediction model*," "computational intelligence"
NCD domain: "Noncommunicable Diseases" [MeSH], "Cardiovascular Diseases" [MeSH], "Diabetes Mellitus" [MeSH], "Neoplasms" [MeSH], "Chronic Disease" [MeSH], "hypertension," "cancer," "chronic respiratory disease," "disease burden," "mortality prediction"
African context: "Africa" [MeSH], "Africa South of the Sahara" [MeSH], individual country names for all 54 African nations, "sub-Saharan Africa," "North Africa"

Other search strategies included: (1) manual screening of reference lists from all included studies and relevant systematic reviews for eligible studies not captured through database searches (backward citation chaining), (2) forward citation tracking of key included studies using Google Scholar and Web of Science, (3) consultation with subject matter experts to identify additional grey literature, unpublished studies, and ongoing research initiatives, and (4) targeted searches of organizational websites including the World Health Organization African Regional Office, African Union, African Development Bank, and regional health economics and epidemiology research networks.

2.4 Study Selection Process

All records identified through database searches and supplementary strategies were exported to EndNote 20 reference management software (Clarivate Analytics, Philadelphia, USA), where automated and manual duplicate removal was executed. The deduplicated citation library was subsequently imported into Covidence systematic review management software (Veritas Health Innovation, Melbourne, Australia) to facilitate collaborative screening and maintain an auditable selection process.

The study selection process was conducted in three stages by two independent reviewers (initials blinded for review). Prior to commencing screening, a calibration exercise was undertaken wherein both reviewers independently screened a random sample of 100 abstracts, followed by calculation of inter-rater agreement (Cohen's kappa coefficient). Discrepancies were discussed to ensure consistent application of eligibility criteria. The target kappa coefficient threshold was set at ≥ 0.80 prior to proceeding with full screening.

Stage 1: Title and Abstract Screening - Both reviewers independently screened titles and abstracts of all retrieved records against predetermined eligibility criteria. Studies clearly meeting exclusion criteria based on title and abstract information were excluded. Studies of uncertain eligibility were retained for full-text review. Conflicts were resolved through discussion, with involvement of a third reviewer (senior author) when consensus could not be achieved.

Stage 2: Full-Text Review - Full-text articles were retrieved for all records passing title/abstract screening. Both reviewers independently assessed full-text articles for eligibility, documenting reasons for exclusion using standardized categories aligned with eligibility criteria. Inter-rater agreement was calculated, and disagreements were resolved through discussion and senior author arbitration when necessary.

Stage 3: Data Extraction - Following completion of full-text review and finalization of included studies, data extraction was conducted independently by both reviewers using a standardized data extraction form developed specifically for this review (described below).

Throughout all stages, screening decisions, inclusion/exclusion reasons, and conflict resolution processes were documented systematically within Covidence to maintain transparency and enable audit trails. The PRISMA flow diagram documenting the study selection process, including numbers of records at each stage and reasons for exclusion, is presented in Fig. 1.

2.5 Data Extraction and Data Items

A comprehensive data extraction form was developed iteratively through team discussion and pilot-tested on a random sample of ten included studies. Modifications were made based on pilot testing to enhance clarity, completeness, and feasibility. The final data extraction form captured the following data categories:

Study Characteristics:

First author, publication year, journal or publication venue
Country/countries of study conduct and author affiliations
Study design and methodological approach
Study objectives and research questions
Funding sources and conflicts of interest

Population and Context:

African country/countries included in the study
Population characteristics (age groups, urban/rural, socioeconomic indicators)
Sample size and sampling methodology
Healthcare setting (community-based, primary care, hospital-based, population-level)

NCD Focus:

Specific NCD category/categories addressed (cardiovascular diseases, diabetes, cancer, chronic respiratory diseases, mental health disorders, chronic kidney disease, others)
Disease-specific outcomes examined (incidence, prevalence, mortality, DALYs, hospitalization, complications)
Risk factors incorporated into predictive models

Artificial Intelligence Methodology:

Type of AI/ML approach employed (supervised learning, unsupervised learning, reinforcement learning, hybrid approaches)
Specific algorithms utilized (random forest, support vector machines, neural networks, gradient boosting, deep learning architectures, others)
Temporal dimension of prediction (cross-sectional risk assessment, short-term forecasting, medium-term forecasting, long-term forecasting)
Model development approach (de novo development, adaptation of existing models, ensemble methods)

Data Sources and Variables:

Primary data sources (electronic health records, national health surveys, disease registries, cohort studies, administrative claims data, multiple integrated sources)
Types of predictor variables (demographic, clinical, laboratory, behavioural, socioeconomic, environmental, genetic)
Data quality assessment and preprocessing methods
Handling of missing data

Model Performance and Validation:

Validation strategy employed (internal validation only, temporal validation, geographical external validation, both temporal and geographical)
Performance metrics reported (accuracy, sensitivity, specificity, positive predictive value, negative predictive value, area under ROC curve, F1 score, calibration metrics)
Model interpretability approaches (feature importance, SHAP values, partial dependence plots)
Comparison with traditional statistical models or existing predictive tools

Implementation and Usability:

Model deployment status (development only, pilot implementation, scaled implementation)
Tools or platforms developed for model utilization
Barriers and facilitators to implementation identified
Cost-effectiveness or economic evaluation considerations
Equity and ethical considerations addressed

Key Findings and Implications:

Primary findings and conclusions
Implications for policy, practice, or future research
Limitations acknowledged by study authors

All data extraction was conducted independently by two reviewers using standardized forms within Microsoft Excel. Discrepancies in extracted data were identified through comparison and resolved through discussion with reference to the original source documents. Quality control procedures included independent verification of quantitative data entries by the senior author for a random 20% sample of included studies.

2.6 Quality Assessment

Given the scoping review objective of mapping the breadth of evidence rather than assessing intervention effectiveness, formal critical appraisal and risk of bias assessment were not mandatory components per scoping review guidelines [24, 25]. However, to provide additional context regarding the methodological quality of included predictive modelling studies, a modified quality assessment was conducted using the Prediction model Risk Of Bias Assessment Tool (PROBAST), adapted for scoping review purposes. PROBAST assesses risk of bias and applicability across four domains: participants, predictors, outcome, and analysis.

Two independent reviewers rated each included study as having low, moderate, or high risk of bias in each domain. Overall quality ratings were assigned based on the highest risk of bias across domains (i.e., a study was rated high risk overall if any domain was rated high risk). Quality assessment results were utilized descriptively to characterize the overall evidence base and identify methodological quality as an important consideration for knowledge gap analysis, but individual studies were not excluded based on quality ratings given the mapping objectives of this scoping review.

2.7 Data Synthesis and Presentation

Given the anticipated heterogeneity in study designs, AI/ML methodologies, NCD categories, and outcomes examined, meta-analysis was not appropriate for this scoping review. Instead, data were synthesized using narrative synthesis approaches complemented by descriptive quantitative summaries and data visualizations.

The synthesis strategy encompassed:

Descriptive Numerical Summary

Frequencies and percentages were calculated for categorical variables (e.g., countries studied, AI algorithms employed, NCDs examined, data sources utilized, validation strategies). These were presented in summary tables to provide overview of evidence distribution.

Narrative Synthesis

Textual descriptions were developed to characterize key themes, patterns, and findings across the included literature. Narrative synthesis addressed each research question systematically, drawing upon extracted data to describe the current evidence landscape, identify consistencies and variations, and highlight notable gaps.

Thematic Analysis

Data related to implementation barriers, facilitators, ethical considerations, and future research recommendations were analysed thematically. Themes were identified inductively through iterative reading and coding of extracted data, with validation through team discussion.

Data Visualization

Multiple visualization approaches were employed to enhance accessibility and interpretability of findings

PRISMA flow diagram depicting study selection process
Geographical heat map showing distribution of studies across African countries
Bar charts displaying distribution of studies by NCD category, AI/ML algorithm type, data sources, and validation strategies
Network diagram illustrating collaborative relationships among research institutions

Knowledge Gap Analysis: A structured framework was employed to systematically identify and categorize knowledge gaps emerging from the synthesis. Gaps were classified as: (1) geographical gaps (African countries or regions underrepresented), (2) disease-specific gaps (NCD categories inadequately addressed), (3) methodological gaps (AI/ML approaches underutilized or validation inadequate), (4) data infrastructure gaps (limitations in available data sources), and (5) implementation science gaps (limited evidence on real-world deployment and impact).

2.8 Stakeholder Engagement

Throughout the review process, engagement with key stakeholders was maintained to enhance relevance and usability of findings. Stakeholder consultation included: (1) forming an advisory panel comprising African health policymakers, clinicians, AI researchers, and patient representatives who provided input on research questions, interpretation of findings, and dissemination strategies; (2) presenting preliminary findings at regional conferences to solicit feedback; and (3) planning targeted dissemination strategies including policy briefs, webinars, and partnerships with African health organizations to maximize impact.

2.9 Ethics and Dissemination

As a scoping review synthesizing published literature, primary ethical approval was not required. However, ethical principles of research integrity, transparent reporting, and acknowledgment of all contributors were rigorously maintained throughout the review process. Dissemination strategies included: publication in an open-access peer-reviewed journal, presentation at relevant international conferences, development of plain-language summaries for policymakers and public audiences, and engagement with African health informatics networks to facilitate knowledge translation.

3.1 Study Selection and Characteristics

The comprehensive database searches and supplementary strategies yielded a total of 8,547 records after removal of duplicates (Figure 1: PRISMA Flow Diagram). Title and abstract screening resulted in exclusion of 7,892 records that clearly did not meet eligibility criteria, leaving 655 records for full-text review. Inter-rater agreement for title/abstract screening was substantial (Cohen's κ = 0.86, 95% CI: 0.82-0.90). Full-text review was completed for 655 articles, of which 528 were excluded based on predetermined exclusion criteria. The most common reasons for exclusion at full-text stage were: studies not employing AI/ML predictive modelling approaches (n=198, 37.5%), studies not addressing African populations or contexts (n=147, 27.8%), studies focusing on disease diagnosis or clinical decision support rather than burden forecasting (n=104, 19.7%), studies addressing only communicable diseases (n=52, 9.8%), and duplicate publications or conference abstracts subsequently published as full articles (n=27, 5.1%). Ultimately, 127 studies met all inclusion criteria and were included in the final synthesis. Inter-rater agreement for full-text screening was excellent (Cohen's κ = 0.91, 95% CI: 0.87-0.94).

Systematic identification and selection of studies for inclusion in the scoping review. Database searches and supplementary sources yielded 8,997 records; 450 duplicates were removed. After title/abstract screening of 8,547 records, 655 underwent full-text review. Following application of eligibility criteria, 127 studies were included in the final synthesis. Inter-rater agreement: Cohen's κ=0.86 for title/abstract screening; κ=0.91 for full-text review.

The temporal distribution of publications revealed accelerating growth in research output over time. Only 8 studies (6.3%) were published between 2010 and 2015, compared to 31 studies (24.4%) between 2016 and 2020, and 88 studies (69.3%) between 2021 and October 2025. The year 2023 represented peak publication output with 34 studies (26.8% of total). This temporal trend reflects both the broader proliferation of AI/ML applications in health and the increasing recognition of NCDs as a priority health challenge in Africa.

3.2 Geographical Distribution

Substantial geographical concentration characterized the included studies, with marked overrepresentation of African countries and complete absence of research from others (Table 2). South Africa emerged as the most extensively studied setting, represented in 44 studies (34.6% of total). Kenya contributed 23 studies (18.1%), Nigeria 19 studies (15.0%), Egypt 12 studies (9.4%), and Ethiopia 11 studies (8.7%). The remaining 18 studies (14.2%) were distributed across 13 additional African countries: Ghana (5 studies), Uganda (3 studies), Tanzania (2 studies), Malawi (2 studies), Rwanda (1 study), Cameroon (1 study), Senegal (1 study), Zimbabwe (1 study), Botswana (1 study), Mauritius (1 study), Tunisia (1 study), Morocco (1 study), and Algeria (1 study).

Critically, 36 African countries (66.7% of the continent's 54 nations) were completely unrepresented in the identified literature. Geographical analysis revealed pronounced regional disparities. Eastern Africa contributed 38 studies (29.9%), Southern Africa 48 studies (37.8%), Western Africa 27 studies (21.3%), Northern Africa 14 studies (11.0%), and Central Africa was represented by only a single study from Cameroon (0.8%). Francophone and lusophone African nations were particularly underrepresented, with limited research from countries including the Democratic Republic of Congo, Mozambique, Angola, Burkina Faso, Mali, Niger, Chad, and others.

Multi-country collaborative studies were rare, representing only 11 studies (8.7% of total). These predominantly involved partnerships between South African and other Southern African Development Community (SADC) nations or collaborations between East African Community member states. Continental-scale studies encompassing multiple African regions were entirely absent from the identified literature.

Table 2: Distribution of Studies by African Country and Region

Region	Countries (Number of Studies)	Total Studies	% of Total
Southern Africa	South Africa (44), Botswana (1), Zimbabwe (1), Malawi (2), Mauritius (1)	48	37.8%
Eastern Africa	Kenya (23), Ethiopia (11), Uganda (3), Tanzania (2), Rwanda (1)	38	29.9%
Western Africa	Nigeria (19), Ghana (5), Senegal (1), Cameroon (1)	27	21.3%
Northern Africa	Egypt (12), Tunisia (1), Morocco (1), Algeria (1)	14	11.0%
Central Africa	Cameroon (1)	1	0.8%

Table 2. Geographic distribution of included studies across African regions and countries

Distribution of 127 included studies across five African regions (Southern, Eastern, Western, Northern, and Central Africa) and 18 individual countries. Study counts and percentages are provided for each country and region. South Africa (n=44, 34.6%), Kenya (n=23, 18.1%), and Nigeria (n=19, 15.0%) accounted for 67.7% of all studies, while 36 African countries (66.7%) had no published research on AI/ML for NCD forecasting.

3.3 Disease-Specific Focus

The distribution of studies across NCD categories demonstrated pronounced emphasis on particular disease groups with substantial gaps in others (Figure 2 - to be inserted). Cardiovascular diseases constituted the most frequently examined NCD category, addressed in 54 studies (42.5% of total). Within cardiovascular diseases, hypertension prediction was investigated in 31 studies, stroke forecasting in 15 studies, coronary artery disease in 12 studies, and heart failure in 9 studies (several studies examined multiple cardiovascular conditions).

Diabetes mellitus was examined in 49 studies (38.6%), representing the second most common disease focus. Of these, 38 studies focused on type 2 diabetes prediction, 8 studies addressed diabetes complications including nephropathy and retinopathy, and 3 studies examined gestational diabetes. Cancer represented the focus of 13 studies (10.2%), with breast cancer most frequently studied (5 studies), followed by cervical cancer (3 studies), colorectal cancer (2 studies), liver cancer (2 studies), and lung cancer (1 study).

Chronic respiratory diseases were substantially underrepresented, addressed in only 11 studies (8.7%). These included chronic obstructive pulmonary disease (6 studies) and asthma (5 studies). Other NCDs examined included chronic kidney disease (18 studies, 14.2%), mental health disorders (7 studies, 5.5%), liver disease (4 studies, 3.1%), and osteoarthritis (2 studies, 1.6%). Multiple studies addressed more than one NCD category simultaneously, examining co-morbidities or developing multi-disease prediction models.

3.4 AI/ML Methodological Approaches

Substantial diversity characterized the AI/ML algorithms and methodological approaches employed across included studies, though certain algorithms predominated (Table 3). Supervised machine learning approaches were utilized in 119 studies (93.7%), while unsupervised learning was employed in only 5 studies (3.9%) primarily for clustering patient subgroups, and hybrid approaches combining supervised and unsupervised methods were implemented in 3 studies (2.4%).

Among supervised learning algorithms, ensemble methods were most common, with random forest algorithms utilized in 40 studies (31.5% of all studies). Gradient boosting variants including XGBoost, LightGBM, and CatBoost were employed in 27 studies (21.3%). Support vector machines (SVM) were implemented in 31 studies (24.4%), while logistic regression as a baseline comparison model appeared in 28 studies (22.0%).

Deep learning architectures were employed in 24 studies (18.9%), with substantial variation in specific implementations. Feedforward neural networks were most common (11 studies), followed by convolutional neural networks (CNN) adapted for structured health data (7 studies), recurrent neural networks including Long Short-Term Memory (LSTM) networks for temporal sequence modelling (4 studies), and transformer architectures (2 studies). Naive Bayes classifiers were implemented in 15 studies (11.8%), decision trees in 19 studies (15.0%), and k-nearest neighbours in 12 studies (9.4%).

Notably, 76 studies (59.8%) employed ensemble or model stacking approaches, comparing performance across multiple algorithms to identify optimal models for specific contexts. However, only 41 studies (32.3%) provided detailed hyperparameter tuning procedures, and merely 34 studies (26.8%) employed systematic feature selection methodologies beyond clinical judgment.

Table 3: Distribution of AI/ML Algorithms Employed

Algorithm Category	Specific Methods	Number of Studies	% of Total Studies
Ensemble Methods	Random Forest	40	31.5%
	Gradient Boosting (XGBoost, LightGBM, CatBoost)	27	21.3%
Support Vector Machines	Linear SVM, RBF SVM, Polynomial SVM	31	24.4%
Deep Learning	Feedforward Neural Networks	11	8.7%
	Convolutional Neural Networks	7	5.5%
	Recurrent Neural Networks/LSTM	4	3.1%
	Transformer Models	2	1.6%
Decision Trees	Single Decision Trees, CART	19	15.0%
Naive Bayes	Gaussian NB, Multinomial NB	15	11.8%
K-Nearest Neighbors	-	12	9.4%
Logistic Regression	Used as baseline comparator	28	22.0%

Note: Total exceeds 100% as many studies employed multiple algorithms

Table 3. Frequency of artificial intelligence and machine learning algorithms used in included studies

Distribution of AI/ML algorithms employed across 127 studies, categorized by algorithm type and specific methods. Random forest (n=40, 31.5%) and support vector machines (n=31, 24.4%) were most common. Supervised learning approaches dominated (93.7% of studies). Note: Totals exceed 100% as many studies employed multiple algorithms for comparison or ensemble methods.

3.5 Temporal Dimensions of Prediction

A critical finding concerned the temporal characteristics of prediction models developed across included studies. The substantial majority of studies (87 studies, 68.5%) focused on cross-sectional risk prediction or classification, estimating current or near-term individual risk rather than forecasting future population-level disease burden. These models typically predicted disease presence, likelihood of developing disease within 1-2 years, or risk category classification based on current patient characteristics.

Only 30 studies (23.6%) incorporated explicit temporal forecasting methodologies suitable for projecting future disease burden at population levels over medium to long-term horizons. Among these, short-term forecasting (1-3 years) was most common (18 studies), medium-term forecasting (3-5 years) was conducted in 9 studies, and long-term forecasting (>5 years) was attempted in only 3 studies. Time-series analysis approaches including ARIMA models, exponential smoothing, and recurrent neural networks were employed in 22 studies for temporal forecasting. Compartmental epidemiological models integrated with machine learning were utilized in 5 studies to project disease transmission dynamics.

Ten studies (7.9%) focused exclusively on spatial rather than temporal prediction, mapping geographical variation in NCD burden without temporal forecasting components. These employed spatial analytics techniques including geographical information systems (GIS), spatial autocorrelation analysis, and geographically weighted regression to identify high-risk geographical areas.

3.6 Data Sources and Variables

Substantial heterogeneity characterized the data sources utilized for model development and validation across included studies (Table 4). Hospital-based electronic health records (EHRs) or clinical databases constituted the most frequent data source, employed in 65 studies (51.2%). These primarily originated from tertiary referral hospitals in urban centres, raising concerns regarding representativeness for broader populations. National or sub-national population health surveys including Demographic and Health Surveys (DHS), World Health Survey, and country-specific non-communicable disease risk factor surveys were utilized in 36 studies (28.3%).

Disease-specific registries including national cancer registries, diabetes registries, and cardiovascular disease surveillance systems provided data for 18 studies (14.2%). Longitudinal cohort studies contributed data to 14 studies (11.0%), while administrative claims data from health insurance systems were employed in 8 studies (6.3%). Integration of multiple data sources to enhance predictive accuracy was conducted in only 26 studies (20.5%), representing a significant missed opportunity given evidence that multi-source integration typically improves model performance.

The Global Burden of Disease Study data were utilized in 12 studies (9.4%) to extract mortality and disability-adjusted life-year estimates for model calibration or validation. However, primary data collection specifically for predictive modelling purposes was rare, conducted in only 9 studies (7.1%). Mobile health data collection platforms were employed in 5 studies (3.9%), representing emerging opportunities for real-time data acquisition.

Table 4: Data Sources Utilized for Model Development

Data Source Category	Number of Studies	% of Total
Hospital Electronic Health Records	65	51.2%
National/Sub-national Population Surveys	36	28.3%
Disease-Specific Registries	18	14.2%
Longitudinal Cohort Studies	14	11.0%
Administrative Claims Data	8	6.3%
Multiple Integrated Data Sources	26	20.5%
Global Burden of Disease Study	12	9.4%
Primary Data Collection	9	7.1%
Mobile Health Platforms	5	3.9%

Table 4. Primary data sources used for AI/ML model development and validation

Types and frequency of data sources utilized in 127 included studies. Hospital-based electronic health records (n=65, 51.2%) were the predominant data source, followed by population-based surveys (n=36, 28.3%). Only 26 studies (20.5%) integrated multiple data sources. The heavy reliance on hospital-based data raises concerns regarding population representativeness and generalizability to broader African populations.

Regarding predictor variables incorporated into models, demographic factors (age, sex, geographical location) were included in all 127 studies (100%). Clinical variables including blood pressure, body mass index, and laboratory parameters were incorporated in 103 studies (81.1%). Behavioural risk factors including smoking, alcohol consumption, physical activity, and dietary patterns were included in 78 studies (61.4%). Socioeconomic variables including education, income, occupation, and household wealth indices were incorporated in 62 studies (48.8%). Environmental and contextual factors including air quality, urbanization, and healthcare access were included in 34 studies (26.8%). Genetic or family history variables were incorporated in 19 studies (15.0%).

Data quality assessment and preprocessing procedures were explicitly reported in 89 studies (70.1%). Missing data handling strategies varied considerably: complete case analysis (excluding records with missing values) was employed in 42 studies (33.1%), multiple imputation in 28 studies (22.0%), single imputation methods in 15 studies (11.8%), and model-based imputation using machine learning algorithms in 4 studies (3.1%). Concerningly, 38 studies (29.9%) did not explicitly report missing data handling procedures, and 23 studies (18.1%) reported minimal missing data (<5%) without describing handling methods.

3.7 Model Performance and Validation

Model validation strategies and performance assessment varied substantially across included studies, with concerning limitations in external validation (Table 5). Internal validation through data splitting approaches (training and testing sets) was conducted in 114 studies (89.8%), with typical training-testing splits ranging from 70:30 to 80:20. Cross-validation techniques including k-fold cross-validation (typically k=5 or k=10) were employed in 73 studies (57.5%) to enhance reliability of performance estimates.

However, external validation—assessment of model performance in datasets independent from model development, was conducted in only 23 studies (18.1%). Among these, temporal external validation using data from different time periods was performed in 14 studies, geographical external validation using data from different healthcare facilities or regions was conducted in 11 studies, and both temporal and geographical external validation was implemented in only 2 studies. The limited external validation severely restricts conclusions regarding model generalizability across diverse African contexts.

Table 5: Validation Strategies Employed

Validation Strategy	Number of Studies	% of Total
Internal Validation (Train-Test Split)	114	89.8%
Cross-Validation (k-fold)	73	57.5%
External Validation (Any Type)	23	18.1%
- Temporal External Validation	14	11.0%
- Geographical External Validation	11	8.7%
- Both Temporal and Geographical	2	1.6%
Bootstrap Validation	18	14.2%
No Explicit Validation Reported	13	10.2%

Table 5. Model validation strategies reported in included studies

Validation approaches used to assess predictive model performance in 127 studies. Internal validation through train-test splitting (n=114, 89.8%) and cross-validation (n=73, 57.5%) were common. However, external validation was rare (n=23, 18.1%), with only 14 studies (11.0%) conducting temporal validation and 11 studies (8.7%) conducting geographical validation. Limited external validation substantially constrains conclusions regarding model generalizability across diverse African contexts.

Performance metrics reported demonstrated substantial diversity, complicating cross-study comparisons. Area under the receiver operating characteristic curve (AUC-ROC) was the most frequently reported metric, presented in 98 studies (77.2%), with values ranging from 0.72 to 0.98 (median 0.86, interquartile range 0.81-0.91). Accuracy was reported in 86 studies (67.7%), ranging from 68% to 97% (median 84%). Sensitivity was reported in 79 studies (62.2%), specificity in 76 studies (59.8%), positive predictive value in 58 studies (45.7%), and negative predictive value in 54 studies (42.5%).

More sophisticated performance metrics were less commonly reported. F1-score, which balances precision and recall, was reported in 41 studies (32.3%). Calibration assessment, crucial for evaluating whether predicted probabilities correspond to observed outcomes, was conducted in only 28 studies (22.0%) through Hosmer-Lemeshow tests or calibration plots. Precision-recall curves, particularly important for imbalanced datasets common in disease prediction, were presented in 19 studies (15.0%).

Model interpretability and explainability approaches were addressed in 47 studies (37.0%). Feature importance ranking was provided in 38 studies (29.9%), with methods including random forest variable importance, permutation importance, or coefficient magnitudes. Shapley Additive Explanations (SHAP) values for individual prediction explanation were employed in 11 studies (8.7%). Partial dependence plots illustrating relationships between individual features and predictions were presented in 8 studies (6.3%). Concerningly, 80 studies (63.0%) employed complex "black box" algorithms without providing interpretability analysis, limiting clinical utility and trust.

Comparative assessment against traditional statistical models or existing clinical risk scores was conducted in 67 studies (52.8%). Machine learning approaches generally demonstrated superior performance compared to logistic regression models, with AUC-ROC improvements ranging from 0.03 to 0.17 (median improvement 0.08). Comparisons with established clinical risk scores (e.g., Framingham Risk Score for cardiovascular disease, FINDRISC for diabetes) showed ML models outperformed traditional scores in 42 of 48 comparisons (87.5%).

3.8 Implementation and Real-World Deployment

A pronounced gap existed between model development and real-world implementation. Among the 127 included studies, the vast majority (104 studies, 81.9%) described models that remained at the development or validation stage without deployment in clinical or public health settings. Only 23 studies (18.1%) reported pilot implementation or deployment activities. Of these, 15 studies described integration into electronic health record systems for clinical decision support, 5 studies reported deployment in mobile health applications for community-level screening, and 3 studies described implementation in public health surveillance systems for population monitoring.

Barriers to implementation were explicitly discussed in 34 studies (26.8%). Commonly cited barriers included: insufficient digital infrastructure and unreliable internet connectivity (mentioned in 23 studies), limited interoperability between health information systems (19 studies), inadequate training and digital literacy among healthcare providers (17 studies), concerns regarding data privacy and security (14 studies), lack of regulatory frameworks for AI in healthcare (12 studies), absence of sustainable funding mechanisms (11 studies), and cultural resistance or mistrust toward AI technologies (7 studies).

Only 11 studies (8.7%) conducted any form of economic evaluation, including cost-effectiveness analysis or budget impact assessment. Among these, 8 studies reported that AI-driven prediction and early intervention strategies were potentially cost-effective compared to standard care, with incremental cost-effectiveness ratios below country-specific willingness-to-pay thresholds. However, the limited economic evidence constrains policy decision-making regarding resource allocation for AI implementation.

Equity considerations were explicitly addressed in 18 studies (14.2%). These examined potential disparities in model performance across population subgroups defined by socioeconomic status, geographical location (urban versus rural), sex, or ethnicity. Notably, 7 studies identified significant performance variation across subgroups, with models generally performing better for urban, higher socioeconomic status populations, potentially exacerbating existing health inequities. Algorithmic fairness assessment and bias mitigation strategies were employed in only 4 studies (3.1%).

Ethical considerations beyond equity, including informed consent for data use, potential harms from false predictions, and governance mechanisms, were discussed in 22 studies (17.3%). Data privacy protection approaches varied, with 47 studies (37.0%) describing de-identification procedures, 18 studies (14.2%) implementing differential privacy techniques, and 8 studies (6.3%) utilizing federated learning approaches enabling model training without centralized data aggregation.

3.9 Methodological Quality Assessment

Quality assessment using the adapted PROBAST tool revealed moderate to high risk of bias in substantial proportions of included studies (Table 6). In the participants domain, 34 studies (26.8%) were rated as having high risk of bias, primarily due to highly selected populations (e.g., single tertiary hospital without population representativeness) or inadequate description of source populations. The predictors domain showed high risk of bias in 29 studies (22.8%), most commonly due to predictor measurement inconsistencies or lack of standardization across data sources.

The outcome domain demonstrated high risk of bias in 41 studies (32.3%), predominantly due to unclear outcome definitions, variable outcome ascertainment methods, or insufficient follow-up duration for temporal forecasting studies. The analysis domain exhibited the highest proportion of high risk of bias, with 56 studies (44.1%) rated as high risk, most frequently due to inadequate sample sizes, lack of appropriate validation, or failure to account for overfitting through regularization or cross-validation.

Overall, considering the highest risk domain for each study, 63 studies (49.6%) were rated as having high risk of bias, 49 studies (38.6%) moderate risk, and only 15 studies (11.8%) low risk of bias. These findings underscore the need for enhanced methodological rigor in future predictive modelling research for NCD forecasting in Africa.

Table 6: Risk of Bias Assessment Using Adapted PROBAST

PROBAST Domain	Low Risk	Moderate Risk	High Risk
Participants	58 (45.7%)	35 (27.6%)	34 (26.8%)
Predictors	72 (56.7%)	26 (20.5%)	29 (22.8%)
Outcome	53 (41.7%)	33 (26.0%)	41 (32.3%)
Analysis	38 (29.9%)	33 (26.0%)	56 (44.1%)
Overall Rating	15 (11.8%)	49 (38.6%)	63 (49.6%)

Table 6. Risk of bias assessment across four PROBAST domains

Summary of methodological quality assessment for 127 included studies using the adapted Prediction model Risk Of Bias Assessment Tool (PROBAST). Studies were rated as low, moderate, or high risk of bias across four domains: participants, predictors, outcome, and analysis. Overall ratings reflect the highest risk domain for each study. Only 15 studies (11.8%) were rated low risk overall, while 63 studies (49.6%) had high risk of bias, primarily due to limitations in the analysis domain including inadequate sample sizes, lack of validation, and failure to address overfitting.

3.10 Collaborative Networks and Research Capacity

Analysis of author affiliations and collaborative patterns revealed concerning concentration of research capacity in particular institutions and limited South-South collaboration. Among the 127 included studies, primary authorship was affiliated with institutions in high-income countries (primarily United States, United Kingdom, and Europe) in 48 studies (37.8%), despite the studies focusing on African populations and data. African-led research (first and senior authors from African institutions) represented 79 studies (62.2%).

Institutional analysis identified several high-output research centres: University of Cape Town and associated South African institutions contributed to 22 studies (17.3%), University of Nairobi and Kenyan research institutions to 14 studies (11.0%), University of Ibadan and Nigerian institutions to 11 studies (8.7%), and University of the Witwatersrand to 9 studies (7.1%). Concentration of research capacity in these institutions highlights both centres of excellence and the need for capacity strengthening across broader African research networks.

North-South collaborations (partnerships between African and high-income country institutions) were present in 67 studies (52.8%). However, South-South collaborations involving researchers from multiple African countries were relatively rare, identified in only 18 studies (14.2%). Continental research networks or multi-country African consortia were absent from the identified literature, representing a significant missed opportunity for knowledge sharing, resource optimization, and harmonization of approaches.

Funding sources were reported in 94 studies (74.0%). International development agencies and research funding organizations (including Welcome Trust, NIH, European Commission) funded 42 studies (33.1%), African government research councils funded 28 studies (22.0%), institutional funds supported 18 studies (14.2%), and private sector or philanthropic organizations funded 6 studies (4.7%). Notably, 33 studies (26.0%) did not report funding sources, raising transparency concerns.

3.11 Identified Knowledge Gaps

Systematic analysis of the evidence landscape revealed several critical knowledge gaps requiring prioritization in future research efforts:

Geographical Gaps:

Complete absence of research from 36 African countries (66.7%), particularly in Central Africa, francophone West Africa, and lusophone countries
Limited multi-country or continental-scale studies enabling comparison and generalization
Insufficient representation of rural and remote populations in model development datasets
Lack of research addressing unique contexts of conflict-affected and fragile states

Disease-Specific Gaps:

Substantial underrepresentation of chronic respiratory diseases (8.7% of studies) despite significant burden
Limited focus on specific cancer types beyond breast and cervical cancer
Insufficient attention to mental health disorders (5.5% of studies) despite rising burden
Neglect of multi-morbidity prediction, despite co-occurrence of multiple NCDs being common
Limited research on less visible NCDs including chronic kidney disease, liver disease, and rheumatological conditions

Methodological Gaps:

Predominance of cross-sectional risk prediction over longitudinal burden forecasting suitable for health system planning
Limited application of advanced time-series forecasting methods and deep learning architectures
Insufficient external validation (only 18.1% of studies), severely limiting generalizability conclusions
Rare use of causal inference approaches to distinguish prediction from causation
Limited ensemble methods combining diverse data sources and algorithms for enhanced accuracy

Data Infrastructure Gaps:

Heavy reliance on hospital-based data sources with questionable population representativeness
Insufficient integration of multiple data sources (only 20.5% of studies)
Limited utilization of novel data streams including mobile health, wearable devices, and environmental sensors
Inadequate data quality assessment and standardization procedures
Absence of continent-wide data sharing platforms or federated learning infrastructures

Implementation Science Gaps:

Pronounced gap between model development and real-world deployment (only 18.1% pilot implementation)
Limited evidence on implementation barriers, facilitators, and strategies
Insufficient economic evaluation to inform policy resource allocation decisions
Inadequate assessment of equity implications and algorithmic fairness
Limited engagement with end-users including policymakers, healthcare providers, and patients in model development

Validation and Interpretability Gaps:

Rare temporal validation assessing model performance over time as epidemiology evolves
Limited geographical external validation across diverse African contexts
Insufficient model interpretability and explainability for clinical adoption
Inadequate calibration assessment beyond discrimination metrics
Limited comparison with existing clinical risk scores and traditional forecasting approaches

4.1 Principal Findings

This comprehensive scoping review identified and synthesized 127 studies examining predictive AI and ML models for NCD burden forecasting in Africa, revealing a nascent but rapidly expanding research domain characterized by substantial promise and equally significant challenges. The evidence landscape demonstrates accelerating growth in publication output, particularly since 2021, reflecting both the proliferation of AI/ML technologies and increasing recognition of NCDs as a priority health challenge across the African continent. However, this growth is accompanied by pronounced geographical concentration, disease-specific imbalances, and critical methodological limitations that constrain the immediate applicability of existing research to inform health system planning and policy development.

The predominant focus on individual-level risk prediction (68.5% of studies) rather than population-level burden forecasting represents a fundamental disconnect between research outputs and health system needs. While individual risk stratification serves important clinical purposes, strategic health planning requires population-level projections of disease incidence, prevalence, mortality, and associated resource implications over medium to long-term horizons. The limited temporal forecasting research (23.6% of studies) inadequately addresses the core question motivating this review: how can AI/ML technologies enhance capacity to anticipate and prepare for evolving NCD burden in African populations?

Geographical disparities in research distribution are stark, with South Africa, Kenya, and Nigeria accounting for 67.7% of all identified studies while two-thirds of African nations remain completely unstudied. This concentration reflects broader patterns in African health research capacity and funding allocation but raises serious questions regarding the relevance and transferability of findings to diverse African contexts. The near-complete absence of francophone and lusophone country representation is particularly concerning, given that these regions encompass substantial proportions of the African population and exhibit distinct epidemiological profiles, health system configurations, and linguistic-cultural contexts.

Disease-specific imbalances, with cardiovascular disease and diabetes heavily emphasized while chronic respiratory diseases and cancer remain underexplored, create critical blind spots in evidence bases needed for comprehensive NCD responses. The relative neglect of cancer forecasting is particularly notable given rising cancer incidence across Africa, delayed diagnoses, and inadequate treatment capacity [19]. Similarly, the minimal attention to mental health disorders (5.5% of studies) despite substantial and increasing burden represents a missed opportunity for AI/ML to inform much-needed mental health system strengthening [17].

4.2 Comparison with Existing Literature

This scoping review extends beyond previous syntheses by providing the first comprehensive mapping specifically addressing AI/ML for NCD forecasting in African contexts. Earlier systematic reviews have examined AI applications in specific disease categories globally, focused on diagnostic rather than predictive applications, or addressed developed country contexts without considering African implementation challenges [11, 13, 18]. The present review's findings align with broader literature documenting AI's potential for transforming health systems while simultaneously revealing persistent inequities in research investment and capacity/ This relationship is displayed in Fig. 4. [3, 6, 10].

Consistent with prior reviews, this analysis demonstrates that machine learning approaches generally outperform traditional statistical methods for disease prediction, with typical AUC-ROC improvements of 0.08 observed across comparative studies [11]. However, this review uniquely identifies the predominance of supervised learning approaches (93.7%) with limited exploration of unsupervised and reinforcement learning methods that may offer advantages for identifying novel disease phenotypes or optimizing resource allocation strategies.

The limited external validation identified in this review (18.1% of studies) parallels findings from previous systematic reviews of clinical prediction models across diverse health domains, suggesting this represents a pervasive challenge rather than unique to African NCD forecasting research [13, 18]. However, the consequences of inadequate external validation may be particularly severe in African contexts where healthcare delivery systems, population demographic profiles, risk factor distributions, and socioeconomic determinants differ substantially from settings where models were developed.

Network diagram showing collaborative relationships among institutions contributing to 127 studies. Node size indicates research output; colours denote institution type (red: major hubs ≥ 20 studies, orange: regional hubs 10–19, green: active 5–9, blue: contributing 1–4). Line colours represent collaboration types: dark (North-South), purple (South-South), grey (HIC-HIC). Major hubs include University of Cape Town (n = 22), University of Nairobi (n = 14), and University of Ibadan (n = 11). North-South collaborations (52.8%) dominated over South-South partnerships (14.2%).

4.3 Implications for Policy and Practice

The findings of this review carry several important implications for policymakers, health system managers, researchers, and development partners seeking to leverage AI/ML technologies for addressing Africa's NCD crisis.

For Health Policymakers and Planners:

Strategic investment in digital health infrastructure and AI capacity building should prioritize development of population-level burden forecasting systems rather than exclusively focusing on clinical decision support tools. While individual risk prediction models offer clinical value, they provide insufficient information for strategic questions including: How many diabetes cases will require treatment five years hence? What capacity for cancer services will be needed by 2030? How should limited resources be allocated across competing NCD prevention and treatment priorities? Forecasting models addressing these questions should integrate demographic projections, risk factor trend analyses, health system capacity constraints, and socioeconomic development scenarios to generate actionable intelligence [28].

Policymakers must recognize that importing predictive models developed in high-income countries or even other African settings without rigorous local validation risks poor performance and potentially harmful clinical or policy decisions. Local model adaptation and validation should be prioritized, with dedicated resources for assembling appropriate training and testing datasets reflecting local population characteristics, healthcare delivery patterns, and outcome ascertainment processes. Regional collaboration and data sharing agreements can facilitate this validation while optimizing limited resources [30].

The pronounced geographical research gaps identified in this review should inform targeted capacity building investments. Deliberate strategies to strengthen research infrastructure, train AI/ML researchers and data scientists, and fund research initiatives in underrepresented regions, particularly Central Africa, francophone West Africa, and lusophone countries, are essential to ensure AI technologies serve the entire continent equitably rather than exacerbating existing disparities [31].

For Healthcare Delivery Organizations:

Health system managers considering adoption of AI-driven prediction tools must carefully evaluate model validation evidence, particularly external validation in settings like their implementation context. The finding that 81.9% of identified models remain at development stages without deployment underscores the substantial gap between proof-of-concept research and practical implementation. Successful implementation requires not only validated algorithms but also digital infrastructure, interoperable health information systems, trained personnel, workflow integration, and ongoing monitoring and model updating as epidemiology evolves [33].

The limited model interpretability identified across many studies (only 37.0% addressed explainability) represents a significant barrier to clinical adoption. Healthcare providers appropriately resist "black box" predictions without understanding why particular risk scores were generated or what factors most influenced predictions. Future model development and selection should prioritize interpretable approaches or incorporate explainability techniques to enhance clinical trust and facilitate appropriate use [34].

For the Research Community:

The methodological limitations identified through quality assessment highlight several priorities for enhancing research rigor. Future studies should prioritize: (1) explicit external validation, ideally across temporal and geographical dimensions, (2) transparent reporting of missing data handling, hyperparameter tuning, and model development procedures following established guidelines (e.g., TRIPOD statement for prediction models), (3) comparative assessment against established clinical risk scores and traditional forecasting methods, (4) model calibration assessment beyond discrimination metrics, and (5) equity analyses examining performance variation across population subgroups.

The limited South-South collaboration and absence of continental research consortia represent missed opportunities for knowledge sharing, resource optimization, and harmonized methodological approaches. Formation of African AI and health informatics networks, supported by African Union, African Development Bank, and regional economic communities, could catalyse collaborative multicounty research initiatives addressing questions requiring large, diverse datasets not feasible within single-country studies [36].

Researchers should engage meaningfully with end-users including policymakers, health system managers, clinicians, and patients throughout the research process rather than conducting development in isolation. Co-design approaches incorporating stakeholder input on priority questions, acceptable trade-offs between model performance and interpretability, and implementation considerations enhance research relevance and accelerate translation [37].

For Development Partners and Funders:

International development agencies and research funders should direct investments strategically toward identified gaps including forecasting (versus only risk prediction) models, underrepresented geographical regions and disease categories, implementation science research, economic evaluations, and capacity building in data science and AI. Funding mechanisms should incentivize collaborative multicounty initiatives enabling adequately powered studies with diverse populations supporting generalizability.

Support for data infrastructure development including health information systems strengthening, data quality improvement, and establishing data governance frameworks enabling appropriate sharing while protecting privacy represents foundational investment essential for AI/ML advancement. Technical assistance for establishing federated learning platforms enabling collaborative model training without centralized data aggregation may offer pathways to address data sharing concerns while enabling continental-scale research [39].

4.4 Addressing Implementation Barriers

The pronounced gap between model development and real-world deployment identified in this review reflects multiple interconnected barriers requiring systematic attention. Infrastructure limitations including unreliable electricity, inconsistent internet connectivity, and outdated computing hardware in many African health facilities constrain deployment feasibility [40]. Addressing these requires substantial investment in foundational digital infrastructure alongside AI-specific initiatives.

Interoperability challenges, with fragmented health information systems using incompatible data standards, prevent seamless integration of prediction models into clinical workflows [10]. Adoption and implementation of international health data standards including HL7 FHIR, coupled with deliberate interoperability requirements in health IT procurement, can progressively address this barrier.

Human resource constraints including limited data science expertise, inadequate digital literacy among healthcare providers, and insufficient training opportunities constrain both model development and deployment capacity [25]. Addressing this requires multi-pronged strategies including incorporating health informatics and AI education into medical, nursing, and public health training curricula; establishing regional centres of excellence providing specialized training in health AI; creating career pathways and retention incentives for health data scientists in African institutions; and fostering interdisciplinary collaborations bridging computer science, statistics, epidemiology, and clinical medicine.

Regulatory frameworks for AI in healthcare remain nascent or absent across most African countries, creating uncertainty regarding approval processes, liability considerations, and quality assurance requirements [16]. Development of context-appropriate regulatory frameworks, potentially through harmonized regional approaches, can provide necessary governance while avoiding excessive bureaucracy that might stifle innovation. Regulatory frameworks should address algorithm validation requirements, ongoing performance monitoring, safety reporting, and ethical considerations including fairness and equity.

Data governance challenges including concerns regarding privacy, security, consent, and ownership require careful attention to ensure AI development does not undermine data protection principles or exacerbate existing power imbalances [18]. Establishment of clear data governance frameworks balancing innovation enablement with appropriate protection, potentially drawing upon frameworks including the African Union's Data Policy Framework, can facilitate responsible AI advancement.

4.5 Ethical Considerations and Equity Implications

The limited attention to equity implications and algorithmic fairness identified in this review (only 14.2% of studies explicitly addressed equity) raises serious concerns regarding potential for AI technologies to exacerbate existing health inequities rather than ameliorate them. Seven studies identified significant performance variation across population subgroups, with models generally performing better for urban, higher socioeconomic populations. If such models were deployed without bias mitigation, they could systematically underestimate risk for marginalized populations, leading to reduced access to preventive interventions and widening outcome disparities [28, 10].

Algorithmic fairness assessment and bias mitigation should become standard components of predictive model development for health applications. Multiple fairness definitions exist (e.g., demographic parity, equalized odds, calibration fairness), and selection among them requires careful consideration of context and values [40]. Engagement with affected communities and ethics review boards in determining acceptable trade-offs between overall performance and fairness across subgroups is essential.

Data representativeness concerns are pronounced given heavy reliance on hospital-based datasets (51.2% of studies), which systematically underrepresent populations with limited healthcare access including rural residents, lower socioeconomic groups, and marginalized communities. Models developed on such data may perform poorly precisely for populations most vulnerable to NCDs and least able to access care. Deliberate strategies to incorporate diverse, representative data sources including population health surveys, community-based screening programs, and primary health centre data can partially address this limitation [24, 18].

The predominance of research led by or in collaboration with high-income country institutions (37.8% with primary authorship from such institutions) raises questions regarding agenda-setting, local relevance, capacity building, and equitable partnership principles. While North-South collaborations bring valuable technical expertise and resources, they must be structured to genuinely strengthen African research capacity, prioritize locally identified needs, ensure African researcher leadership, and build sustainable institutional capabilities rather than extractive partnerships where data flows north while limited capacity remains locally [36, 38].

4.6 Future Research Priorities

Based on identified knowledge gaps and implementation challenges, several research priorities emerge:

Population-Level Burden Forecasting Models: Development and validation of models projecting future NCD incidence, prevalence, mortality, and healthcare resource requirements at district, national, and regional scales over 5–20-year horizons, integrating demographic projections, risk factor trends, intervention scale-up scenarios, and health system constraints.

Underrepresented Geographies: Targeted research initiatives in currently understudied regions, particularly Central Africa, francophone West Africa, and lusophone countries, potentially through multinational collaborative consortia enabling resource pooling and knowledge sharing.

Neglected Disease Categories: Enhanced research focus on chronic respiratory diseases, specific cancer types (particularly those with rising incidence including colorectal, liver, and lung cancers), mental health disorders, and multi-morbidity prediction addressing realistic clinical complexity.

Advanced Methodological Approaches: Application of sophisticated time-series forecasting methods including recurrent neural networks, transformer architectures, and hybrid mechanistic-ML models; causal inference approaches distinguishing association from causation; federated learning enabling multicounty model training without centralized data aggregation; and transfer learning adapting models developed in data-rich settings to resource-constrained contexts.

External Validation Studies: Rigorous temporal and geographical external validation assessing model performance across diverse African populations and settings, with explicit reporting enabling meta-analysis of validation study results.

Implementation Science Research: Studies examining implementation barriers and facilitators, effectiveness of implementation strategies, cost-effectiveness of AI-driven prediction and early intervention approaches, workflow integration considerations, provider and patient perspectives, and equity implications of deployment.

Data Infrastructure Development: Research supporting development of data standards, quality assurance procedures, data sharing governance frameworks, and technological platforms enabling responsible, privacy-preserving collaborative research across African countries.

Capacity Building Research: Studies examining effective models for training health data scientists, optimal curricula integrating AI/ML with public health and clinical domains, and strategies for retention of trained personnel in African research institutions and health systems.

4.7 Strengths and Limitations

This scoping review possesses several important strengths. The comprehensive search strategy across eight databases including African-specific repositories, supplemented by citation chaining and expert consultation, maximized identification of relevant literature including grey literature. The rigorous, transparent methodology following JBI and PRISMA-ScR guidelines, with independent dual screening and data extraction, minimized bias and enhanced reliability. The broad inclusion criteria encompassing diverse study designs, AI/ML approaches, and NCD categories enabled comprehensive mapping of the evidence landscape. Quality assessment using adapted PROBAST provided valuable context regarding methodological rigor. Stakeholder engagement throughout the review process enhanced relevance and usability of findings for intended audiences.

However, several limitations warrant acknowledgment. The restriction to English, French, and Portuguese language publications may have excluded relevant research published in other languages, potentially contributing to geographical gaps identified. The rapidly evolving nature of AI/ML research means that relevant studies published after the October 2025 search date are not captured. The heterogeneity in study designs, methodologies, outcomes, and reporting quality precluded meta-analysis and limited ability to draw quantitative conclusions regarding comparative effectiveness of different AI/ML approaches. The quality assessment tool, while appropriate for predictive modelling studies, may not have captured all relevant quality dimensions for forecasting studies specifically.

The classification of studies as addressing "burden forecasting" versus "risk prediction" required judgment, and borderline cases may have been categorized differently by other reviewers. The focus on published literature may introduce publication bias, as negative or null findings may be underrepresented. The reliance on author-reported information regarding validation procedures and performance metrics, without ability to independently verify these claims, represents an inherent limitation of secondary research. Finally, the scoping review methodology, while appropriate for mapping the evidence landscape, does not provide the same level of critical appraisal as systematic reviews focused on intervention effectiveness.

This scoping review provides the first comprehensive synthesis of evidence regarding predictive artificial intelligence and machine learning models for non-communicable disease burden forecasting in Africa. The analysis of 127 studies reveals an emerging research domain characterized by rapid growth, substantial promise, and critical limitations that must be addressed to realize transformative potential for health system strengthening.

The predominant focus on individual-level risk prediction rather than population-level burden forecasting represents a fundamental misalignment between research outputs and health system planning needs. The pronounced geographical concentration, with two-thirds of African countries completely unrepresented, threatens to create a two-tiered landscape where AI benefits accrue only to populations in research-intensive settings while others are left behind. Disease-specific imbalances, particularly the underrepresentation of chronic respiratory diseases, cancer, and mental health disorders, create blind spots in the evidence base needed for comprehensive NCD responses.

Methodological limitations including limited external validation (only 18.1% of studies), insufficient model interpretability (63.0% of studies provided no explainability analysis), and inadequate attention to equity implications (addressed in only 14.2% of studies) constrain confidence in model generalizability and appropriateness for guiding clinical and policy decisions. The pronounced implementation gap, with 81.9% of models remaining at development stages without real-world deployment, reflects the substantial barriers including infrastructure constraints, interoperability challenges, capacity limitations, and inadequate regulatory frameworks that must be systematically addressed.

Despite these limitations, the evidence demonstrates that machine learning approaches generally outperform traditional statistical methods for NCD prediction, with meaningful improvements in discrimination metrics. The expanding research output, growing African research leadership, and increasing sophistication of methodological approaches indicate positive trajectories. The identification of centres of excellence in South Africa, Kenya, and Nigeria provides foundations for regional leadership and knowledge transfer.

Realizing the potential of AI/ML technologies to address Africa's escalating NCD burden requires coordinated action across multiple fronts. Strategic investments in population-level burden forecasting systems, digital health infrastructure, data governance frameworks, and research capacity are essential. Deliberate efforts to address geographical and disease-specific gaps, enhance methodological rigor particularly regarding external validation, and advance implementation science research linking model development to real-world deployment must be prioritized. Strengthened South-South collaboration through continental research consortia and harmonized approaches to validation, regulation, and implementation can optimize limited resources while accelerating progress.

The NCD epidemic threatens to undermine health and development gains across Africa, but it is not inevitable. With strategic application of AI/ML technologies, grounded in robust evidence, implemented equitably, and adapted to local contexts, African health systems can transition from reactive crisis management to proactive, anticipatory approaches that prevent disease, optimize resource allocation, and improve outcomes for all populations. This scoping review provides the evidence foundation to guide that transformation, while clearly delineating the substantial work remaining to achieve this vision.

Abbreviation	Full Term
AI	Artificial Intelligence
AJOL	African Journals Online
AUC-ROC	Area Under the Receiver Operating Characteristic Curve
CART	Classification and Regression Tree
CINAHL	Cumulative Index to Nursing and Allied Health Literature
CNN	Convolutional Neural Network
COPD	Chronic Obstructive Pulmonary Disease
CRediT	Contributor Roles Taxonomy
DALYs	Disability-Adjusted Life Years
DHS	Demographic and Health Surveys
DR Congo	Democratic Republic of Congo
EHR	Electronic Health Record
Embase	Excerpta Medica Database
FHIR	Fast Healthcare Interoperability Resources
FINDRISC	Finnish Diabetes Risk Score
GBD	Global Burden of Disease
GIS	Geographical Information Systems
HIC	High-Income Country
HL7	Health Level Seven
IEEE	Institute of Electrical and Electronics Engineers
JBI	Joanna Briggs Institute
k-fold	k-fold Cross-Validation
KNN	K-Nearest Neighbors
LightGBM	Light Gradient Boosting Machine
LSTM	Long Short-Term Memory
MEDLINE	Medical Literature Analysis and Retrieval System Online
MeSH	Medical Subject Headings
MIT	Massachusetts Institute of Technology
ML	Machine Learning
NB	Naive Bayes
NCD	Non-Communicable Disease
NGO	Non-Governmental Organization
NIH	National Institutes of Health
PCC	Population, Concept, Context
PRISMA	Preferred Reporting Items for Systematic Reviews and Meta-Analyses
PRISMA-P	PRISMA for Protocols
PRISMA-ScR	PRISMA extension for Scoping Reviews
PROBAST	Prediction model Risk Of Bias ASsessment Tool
RBF	Radial Basis Function
SADC	Southern African Development Community
SHAP	Shapley Additive Explanations
SSA	Sub-Saharan Africa
SVM	Support Vector Machine
TRIPOD	Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis
UK	United Kingdom
US	United States
USD	United States Dollar
WHO	World Health Organization
XGBoost	Extreme Gradient Boosting

Ethics approval and consent to participate

This study does not involve human participants, and ethical approval was not required.

Consent for publication

Not applicable

Data Availability Statement

All data extracted during this scoping review are review are included in the published article (and its supplementary information files). The search strategies, data extraction forms, worksheets are available from the corresponding author upon reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Author Contributions

JA conceptualized the study. C.A.A and D.O.O contributed to the design of the study. S.M.M and A.P.E contributed to the acquisition, analysis, while M.T.A and B.O.J contributed to the interpretation of data. J.A, C.A.A, and D.O.O drafted the study and S.M.M, A.P.E. and M.T.A substantively revised it. J.A developed the search strategy with consultation from B.O.J and C.A.A. D.O.O and A.P.E screened, assessed the eligibility, and assessed the quality of the included studies with consultation from J.A and M.T.A, analysed the data and created the figures with consultation from J.A and C.AA. J.A is responsible for the data management and storage. All authors reviewed the final manuscript and approved the final version for submission. All authors have agreed both to be personally accountable for the author's own contributions and to ensure that questions related to the accuracy or integrity of any part of the work, even ones in which they were not personally involved, are appropriately investigated, resolved, and the resolution documented in the literature.

Acknowledgments

The authors acknowledge the contributions of AwaDoc for assistance with search strategy development, members of the advisory panel for guidance throughout the review process, and all researchers whose work was synthesized in this review for their contributions to advancing knowledge regarding AI applications for NCD forecasting in Africa.

Authors Information

Joy Aifuobhokhan. MD, 0009-0005-9521-905X, Digital Health and Research, Lakeshore Cancer center, Lagos, Nigeria

Chukwuemeka Abraham Agbarakwe. MD, 0000-0002-7267-3699,Internal Medicine, Calvary Specialist Hospital, Port Harcourt, Nigeria

Deborah Oladunmolu Oduguwa. MD,0009-0003-8422-6153, Family Medicine, Babcock University Teaching Hospital, Ilishan-Remo, Nigeria

Somtochi Maduka, M. MD,0009-0007-4177-567X. Paediatrics, Tehilah Children's Hospital, Port Harcourt, Nigeria

Annie Peter Essiet. MD, 0009-0008-9842-3226, Paediatrics, Tehilah Children’s Hospital, Port Harcourt, Nigeria

Moyosoreoluwa Tiwalayo Aluko. MD, 0009-0001-2438, Public Health, Babcock University Teaching Hospital, Ilishan-Remo, Nigeria

Boluwatife Oluwafayoyimika Johnson. MD, 0009-0004-2767-4547, Internal Medicine, Babcock University Teaching Hospital, Ilishan-Remo, Nigeria

Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25:44-56. doi:10.1038/s41591-018-0300-7
Esteva A, Robicquet A, Ramsundar B, Kuleshov V, DePristo M, Chou K, et al. A guide to deep learning in healthcare. Nat Med. 2019;25:24-29. doi:10.1038/s41591-018-0316-z.
Gwagwa A, Kraemer-Mbula E, Rizk N, Rutenberg I, de Beer J. Artificial intelligence (AI) deployments in Africa: benefits, challenges and policy dimensions. Afr J Sci Technol Innov Dev. 2021;13:1485-1500. doi:10.1080/20421338.2020.1860287.
Russell SJ, Norvig P. Artificial intelligence: a modern approach. 4th ed. Hoboken: Pearson; 2020.
Murray CJL, Lopez AD. Measuring the global burden of disease. N Engl J Med. 2013;369:448-457. doi:10.1056/NEJMra1201534.
Choudhury A, Aron S, Botigué LR, Sengupta D, Botha G, Bensellak T, et al. High-depth African genomes inform human migration and health. Nature. 2020;586:741-748. doi:10.1038/s41586-020-2859-7.
Beam AL, Kohane IS. Big data and machine learning in health care. JAMA. 2018;319:1317-1318. doi:10.1001/jama.2017.18391.
Owoyemi A, Owoyemi J, Osiyemi A, Boyd A. Artificial intelligence for healthcare in Africa. Front Digit Health. 2020;2:6. doi:10.3389/fdgth.2020.00006.
Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380:1347-1358. doi:10.1056/NEJMra1814259.
Wahl B, Cossy-Gantner A, Germann S, Schwalbe NR. Artificial intelligence (AI) and global health: how can AI contribute to health in resource-poor settings? BMJ Glob Health. 2018;3:e000798. doi:10.1136/bmjgh-2018-000798.
Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, Hardt M, et al. Scalable and accurate deep learning with electronic health records. NPJ Digit Med. 2018;1:18. doi:10.1038/s41746-018-0029-1.
Gouda HN, Charlson F, Sorsdahl K, Ahmadzada S, Ferrari AJ, Erskine H, et al. Burden of non-communicable diseases in sub-Saharan Africa, 1990-2017: results from the Global Burden of Disease Study 2017. Lancet Glob Health. 2019;7:e1375-e1387. doi:10.1016/S2214-109X(19)30374-2.
Collins GS, Reitsma JB, Altman DG, Moons KGM. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD statement. BMJ. 2015;350:g7594. doi:10.1136/bmj.g7594.
Bigna JJ, Noubiap JJ. The rising burden of non-communicable diseases in sub-Saharan Africa. Lancet Glob Health. 2019;7:e1295-e1296. doi:10.1016/S2214-109X(19)30370-5.
Guwatudde D, Mutungi G, Wesonga R, Kajjura R, Kasule H, Muwonge J, et al. The epidemiology of hypertension in Uganda: findings from the national non-communicable diseases risk factor survey. PLoS One. 2015;10:e0138991. doi:10.1371/journal.pone.0138991.
World Health Organization. Noncommunicable diseases country profiles 2018. Geneva: World Health Organization; 2018.
Charlson FJ, Diminic S, Lund C, Degenhardt L, Whiteford HA. Mental and substance use disorders in Sub-Saharan Africa: predictions of epidemiological changes and mental health workforce requirements for the next 40 years. PLoS One. 2014;9:e110208. doi:10.1371/journal.pone.0110208.
Moons KGM, Altman DG, Reitsma JB, Ioannidis JPA, Macaskill P, Steyerberg EW, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015;162:W1-73. doi:10.7326/M14-0698.
Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre LA, Jemal A. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68:394-424. doi:10.3322/caac.21492.
Manne-Goehler J, Geldsetzer P, Agoudavi K, Andall-Brereton G, Aryal KK, Bicaba BW, et al. Health system performance for people with diabetes in 28 low- and middle-income countries: a cross-sectional study of nationally representative surveys. PLoS Med. 2019;16:e1002751. doi:10.1371/journal.pmed.1002751.
Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366:447-453. doi:10.1126/science.aax2342.
Char DS, Shah NH, Magnus D. Implementing machine learning in health care - addressing ethical challenges. N Engl J Med. 2018;378:981-983. doi:10.1056/NEJMp1714229.
Peters MDJ, Godfrey CM, Khalil H, McInerney P, Parker D, Soares CB. Guidance for conducting systematic scoping reviews. Int J Evid Based Healthc. 2015;13:141-146. doi:10.1097/XEB.0000000000000050.
Wolff RF, Moons KGM, Riley RD, Whiting PF, Westwood M, Collins GS, et al. PROBAST: a tool to assess the risk of bias and applicability of prediction model studies. Ann Intern Med. 2019;170:51-58. doi:10.7326/M18-1376.
Debray TPA, Damen JAAG, Snell KIE, Ensor J, Hooft L, Reitsma JB, et al. A guide to systematic review and meta-analysis of prediction model performance. BMJ. 2017;356:i6460. doi:10.1136/bmj.i6460.
Arksey H, O'Malley L. Scoping studies: towards a methodological framework. Int J Soc Res Methodol. 2005;8:19-32. doi:10.1080/1364557032000119616.
Levac D, Colquhoun H, O'Brien KK. Scoping studies: advancing the methodology. Implement Sci. 2010;5:69. doi:10.1186/1748-5908-5-69.
Tricco AC, Lillie E, Zarin W, O'Brien KK, Colquhoun H, Levac D, et al. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann Intern Med. 2018;169:467-473. doi:10.7326/M18-0850.
Munn Z, Peters MDJ, Stern C, Tufanaru C, McArthur A, Aromataris E. Systematic review or scoping review? Guidance for authors when choosing between a systematic or scoping review approach. BMC Med Res Methodol. 2018;18:143. doi:10.1186/s12874-018-0611-x.
Peters MDJ, Marnie C, Tricco AC, Pollock D, Munn Z, Alexander L, et al. Updated methodological guidance for the conduct of scoping reviews. JBI Evid Synth. 2020;18:2119-2126. doi:10.11124/JBIES-20-00167.
Steyerberg EW, Vergouwe Y. Towards better clinical prediction models: seven steps for development and an ABCD for validation. Eur Heart J. 2014;35:1925-1931. doi:10.1093/eurheartj/ehu207.
Page MJ, McKenzie JE, Bossuyt PM, Boutron I, Hoffmann TC, Mulrow CD, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ. 2021;372:n71. doi:10.1136/bmj.n71.
Ouzzani M, Hammady H, Fedorowicz Z, Elmagarmid A. Rayyan—a web and mobile app for systematic reviews. Syst Rev. 2016;5:210. doi:10.1186/s13643-016-0384-4.
Rethlefsen ML, Kirtley S, Waffenschmidt S, Ayala AP, Moher D, Page MJ, et al. PRISMA-S: an extension to the PRISMA statement for reporting literature searches in systematic reviews. Syst Rev. 2021;10:39. doi:10.1186/s13643-020-01542-z.
McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb). 2012;22:276-282.
Popay J, Roberts H, Sowden A, Petticrew M, Arai L, Rodgers M, et al. Guidance on the conduct of narrative synthesis in systematic reviews: a product from the ESRC methods programme. Lancaster: Lancaster University; 2006.
Thomas J, Harden A. Methods for the thematic synthesis of qualitative research in systematic reviews. BMC Med Res Methodol. 2008;8:45. doi:10.1186/1471-2288-8-45.
Moher D, Liberati A, Tetzlaff J, Altman DG, PRISMA Group. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. PLoS Med. 2009;6:e1000097. doi:10.1371/journal.pmed.1000097.
Shamseer L, Moher D, Clarke M, Ghersi D, Liberati A, Petticrew M, et al. Preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015: elaboration and explanation. BMJ. 2015;350:g7647. doi:10.1136/bmj.g7647.
Riley RD, van der Windt D, Croft P, Moons KGM, editors. Prognosis research in healthcare: concepts, methods, and impact. Oxford: Oxford University Press; 2019.
Joy Aifuobhokhan, Tobiloba P O, Chijioke C E. Artificial Intelligence in Cancer Diagnosis: A Scoping Review of Global Innovation and African Implementation. World Journal of Advanced Research and Reviews, 2025, 28(01), 449-468. Article DOI: https://doi.org/10.30574/wjarr.2025.28.1.3444.
Joy Aifuovhokhan, Olajire O O, Emmanuel T P, Artificial Intelligence - Clinical Decision Support Systems: From brain drain to health system gain. World Journal of Advanced Research and Reviews, 2025, 28(01), 1482-1493. Article DOI: https://doi.org/10.30574/wjarr.2025.28.1.3486.
Aifuobhokhan, J., Hussain, A. K., Ekechi, C. C., Olanrewaju, A. O., Afuadajo, E., Bowale, D. A., & Inioluwa, O. M. (2025). Predicting Thyroid Cancer Recurrence Using Machine Learning: An Artificial Intelligence Approach to Clinical Oncology. International Journal of Advances in Artificial Intelligence and Machine Learning, 2(3), 135–148. https://doi.org/10.58723/ijaaiml.v2i3.469

No competing interests reported.