Independent Assessment of the WHO Skin NTDs App: A Suitable Tool for Leprosy Detection

doi:10.21203/rs.3.rs-8253518/v1

Download PDF

Research Article

Independent Assessment of the WHO Skin NTDs App: A Suitable Tool for Leprosy Detection

https://doi.org/10.21203/rs.3.rs-8253518/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Objective

To independently evaluate the World Health Organization (WHO) Skin Neglected Tropical Diseases (NTDs) app as a clinical decision-support tool for detecting leprosy. The primary objective was to determine whether leprosy appeared within the model's Top-5 diagnostic predictions. A secondary objective was to qualitatively analyse diagnostic error patterns.

Methods

We used a dataset of 439 anonymised clinical images from confirmed leprosy cases (1996–2024), spanning the full Ridley-Jopling spectrum, type 1 and type 2 reactions, and atypical presentations. After excluding 16 images due to processing errors, 423 images were retained: 367 classical leprosy lesions and 56 reactional or atypical leprosy-related presentations. All images were evaluated via the WHO DHIS2 interface. We estimated Top-5 sensitivity (recall) for leprosy and performed a qualitative error analysis focusing on intra-patient inconsistencies and challenging lesion types.

Findings:

The model achieved an overall Top-5 sensitivity (recall) of 84.9%, with higher sensitivity for classical lesions (87.2%) than for reactional or atypical presentations (69.6%). Qualitative review revealed inconsistent predictions for visually similar lesions from the same patient, and misclassifications concentrated among necrotic, inflammatory, and infiltrative lesions.

Conclusion

The WHO Skin NTDs app demonstrates substantial promise as a clinical decision-support and educational tool, especially for classical leprosy. Performance gaps for reactional and atypical forms highlight the need for algorithmic refinement. Enhancing dataset diversity, incorporating lesion segmentation, and integrating patient-level context may improve diagnostic robustness.

Tropical Medicine

Dermatology

Infectious Diseases

Artificial Intelligence and Machine Learning

Leprosy

Hansen’s disease

Artificial intelligence

Deep learning

External analytical validation

WHO Skin NTDs app

Skin NTDs

Diagnostic accuracy.

Leprosy (Hansen’s disease), a neglected tropical disease (NTD), remains a global public health concern. In 2024, 172,717 new leprosy cases were reported from 188 countries and territories, with the highest burden in low- and middle-income regions.¹ Diagnosis is often complex due to the wide clinical spectrum, which ranges from subtle hypochromic macules in paucibacillary forms to diffuse nodular or infiltrative lesions in multibacillary presentations. Reactional states—immune-mediated inflammatory episodes that may occur before, during, or after treatment—add further diagnostic complexity and contribute to misclassification and delayed management.²

Limited specialist availability in endemic regions contributes to substantial diagnostic delays, particularly in remote or resource-constrained settings where frontline providers frequently have limited training in recognising early or atypical manifestations of leprosy.^3,4 These structural gaps are often compounded by persistent stigma, fragmented referral pathways, and operational constraints within primary health services, all of which hinder timely case identification. Consequently, many patients are diagnosed only after the onset of nerve damage and visible deformities, increasing the risk of long-term disability and perpetuating transmission in affected communities.⁵

Advances in artificial intelligence (AI) have created opportunities to support the diagnosis of dermatological conditions and NTDs.^6,7 Techniques such as convolutional neural networks, transfer learning,^8,9 and, more recently, multimodal foundation models integrating large language models with Vision Transformers,¹⁰ have substantially expanded the capacity to analyse skin lesion images at scale. These systems have demonstrated high concordance with dermatologists in experimental settings and, in some narrowly defined diagnostic tasks, have even surpassed human performance.^11–13

The first AI-based approach for leprosy diagnosis demonstrated the feasibility of lesion recognition through image analysis, achieving an accuracy of 91.6%.¹⁴ A subsequent study from Brazil reinforced the potential of AI by combining skin images with demographic and clinical data, reaching 90% accuracy and highlighting the added value of multimodal inputs for more robust classification.¹⁵ A pilot study in Côte d’Ivoire and Ghana extended these findings to a broader spectrum of skin NTDs, applying deep learning to Buruli ulcer, leprosy, mycetoma, scabies, and yaws, reporting over 70% agreement with clinical diagnoses.¹⁶ Despite these encouraging results, most existing studies are constrained by relatively small samples and limited image diversity, underscoring the need for external evaluation in larger, more representative datasets.

In parallel, digital health tools, particularly those incorporating AI, have emerged as a promising strategy to mitigate diagnostic gaps in NTDs. To address case management challenges in frontline settings, the World Health Organization (WHO) developed the Skin NTDs app, a mobile tool that combines offline clinical algorithms with structured learning resources for health workers. The app covers twelve skin NTDs—leprosy, Buruli ulcer, yaws, cutaneous leishmaniasis, post-kala-azar dermal leishmaniasis, onchocerciasis, lymphatic filariasis, mycetoma, chromoblastomycosis, sporotrichosis, scabies, and tungiasis—as well as 24 common skin conditions.¹⁷ In evaluations conducted in Ghana and Kenya, the app achieved a mean quality score of 4.02 (standard deviation 0.47) out of 5, supporting its perceived usefulness as a training and decision-support tool.^18,19

A beta version of the WHO Skin NTDs app incorporates two cascaded AI-based visual classifiers, developed in collaboration with UniversalDoctor and Belle.ai.²⁰ Both models are built on convolutional neural network architectures. The first classifier acts as a broad skin-condition screener, assigning probabilities across 24 common dermatological conditions and indicating when a lesion may warrant further evaluation for a skin-related neglected tropical disease (skin NTD). When a potential NTD is suspected, a second classifier is applied; this NTD-focused model assigns probabilities across the twelve skin NTD categories listed above and returns a ranked list of four suggested NTD diagnoses for each uploaded image. Together, these AI components enable the app to function as both a clinical decision-support tool and a scalable training platform for non-specialist health workers, who can capture and upload anonymised images of skin lesions to obtain a shortlist of possible diagnoses to be integrated with clinical history and examination findings.^7,21

Although the app displays four diagnostic suggestions in routine use, the underlying classifier generates five ranked outputs, which are accessible in the DHIS2 evaluation interface and were used for analytical validation in this study. All analyses in this study refer exclusively to the NTD classifier (second-stage model).

To our knowledge, no previous peer-reviewed study has reported an external analytical validation of the AI-based NTD classifier integrated into the WHO Skin NTDs app. Using an independent dataset of clinically confirmed leprosy cases and their reactional forms, we assessed the classifier’s ability to identify leprosy among the twelve NTD output labels—focusing on Top-5 sensitivity and error patterns—and evaluated its potential utility as a digital health intervention supporting global efforts to control leprosy.

This study aimed to assess the performance of the WHO Skin NTDs app’s AI model in identifying leprosy and its reactional forms. The specific objectives were to assess the model’s sensitivity by verifying whether leprosy appeared among the Top-5 diagnostic predictions generated for each image, and to identify and analyse diagnostic error patterns, with emphasis on reactional states and atypical clinical presentations.

3.1 Study Design

This was an external validation study designed to assess the diagnostic performance of a pre-trained deep learning model currently integrated into the WHO Skin NTDs mobile application. The validation focused on the model’s ability to detect leprosy using an independently curated dataset that had not been used during training or internal validation, as recommended by the WHO.²²

3.2 Data Collection and Dataset

3.2.1 Image Acquisition and Data Security

Clinical images were collected between 1996 and 2024 by leprosy experts. All images originated from patients with confirmed diagnoses of leprosy and represented authentic, real-world clinical scenarios, including a range of lesion types and skin tones.

Strict data protection protocols were applied throughout the study. All images were manually cropped or masked to remove identifiers, anonymised, and securely uploaded to the WHO’s internal SharePoint platform, in accordance with WHO data security standards. Access to images and metadata was restricted to authorised researchers, ensuring data integrity and privacy. A separate test dataset, with no overlap with the training data, was provided to the AI model as images only, without any associated labels or metadata, to ensure unbiased performance evaluation.

The sharing of images adhered to a formal UFES–WHO institutional agreement that regulates authorised use and ensures compliance with applicable ethical and data-protection standards.

3.2.2 External Validation Dataset

The external dataset comprised 439 anonymised clinical images from 208 confirmed leprosy cases. It included a broad spectrum of clinical manifestations, encompassing all Ridley-Jopling classification types and reactional states. Images were linked to structured metadata and corresponding expert-confirmed diagnoses, which were stored separately from the image files. For the present analysis, 16 images with processing errors were excluded, resulting in 423 images divided into two subsets:

Subset 1 (n = 367): Images of classical leprosy lesions, used to assess the model’s ability to detect leprosy among its Top-5 predicted diagnoses;
Subset 2 (n = 56): Images of leprosy reactions (e.g. type 1 reaction, type 2 reaction) and rare or atypical leprosy-related presentations (e.g. Lucio’s phenomenon, erythema multiforme–like lesions), evaluated using the same leprosy-detection endpoint, without subtype-specific analysis due to current model limitations.

Demographic and phenotypic metadata were available for 418 of the 423 images (Subset 1: 362/367; Subset 2: 56/56) and are summarised in Table 1.

Table 1

Demographic and phenotypic characteristics of the external validation subsets*
Variable	Subset 1 (n = 362)	Subset 2 (n = 56)
Sex
Male	56.9% (206)	91.1% (51)
Female	32.6% (118)	8.9% (5)
Missing	10.5% (38)	–
Age group
Adults	87.0% (315)	98.2% (55)
Adolescents	11.0% (40)	–
Children	1.7% (6)	1.8% (1)
Missing	0.3% (1)	–
Fitzpatrick phototype
II	1.1% (4)	7.1% (4)
III	52.5% (190)	64.3% (36)
IV	18.5% (67)	8.9% (5)
V	16.6% (60)	–
VI	11.3% (41)	7.1% (4)
Missing	–	12.5% (7)

*Note: Subset 1 comprises 367 images in total; demographic and phenotypic data were available for 362 images. Subset 2 comprises 56 images, all with complete metadata.

3.2.3 Diagnostic Confirmation

Leprosy diagnosis was established according to WHO guidelines (2018),²³ requiring at least one of the following cardinal signs:

Loss of sensation in a hypopigmented or erythematous skin lesion;

Enlargement of peripheral nerves associated with sensory or motor impairment; or

Detection of acid-fast bacilli in slit-skin smear microscopy.

Paucibacillary cases were classified clinically (≤ 5 lesions and a negative smear), while multibacillary cases were defined by > 5 lesions, a positive smear (bacterial index ≥ 1), or the presence of multiple thickened nerves. When required, diagnosis was supported by additional methods such as histopathology, polymerase chain reaction for M. leprae, anti-phenolic glycolipid-1 serology, or peripheral nerve ultrasound. Leprosy reactions were classified by experienced clinicians according to WHO guidance.²⁴

3.3 Outcome Measures and Evaluation Approach

The primary objective of the study was to assess Top-5 sensitivity (recall), defined as the proportion of confirmed leprosy cases in which leprosy appeared among the model’s five highest-ranked diagnostic predictions within the NTD classifier output. In addition, a Top-N analysis (from Top-1 to Top-5) was conducted to evaluate how often leprosy was prioritised at different ranking levels, thereby providing a more detailed view of the model’s diagnostic hierarchy.

A secondary objective was to conduct a qualitative error analysis to further explore the model’s limitations. This analysis focused on identifying inconsistencies in prediction patterns across images from the same patient and examining misclassified lesion types. Particular attention was given to reactional forms, necrotic and infiltrative lesions, and scenarios in which visually similar images were inconsistently classified by the model.

3.4 Statistical Analysis

All analyses were conducted using R Statistical Software (version 4.2.1; R Foundation for Statistical Computing). Descriptive statistics were reported as absolute frequencies with percentages. Sensitivity was calculated by comparing the algorithm’s Top-5 diagnostic predictions against expert-confirmed diagnoses. Subgroup analyses were performed by subset to identify patterns in model performance.

3.5 Ethical Compliance

This study was approved by the Research Ethics Committee of the Health Sciences Centre, Federal University of Espírito Santo (UFES), under protocol CAAE 83227624.6.0000.5060. The requirement for written informed consent was waived owing to the retrospective nature of the study and the use of fully de-identified data. The research was conducted under a formal collaboration agreement between the WHO Global Neglected Tropical Diseases Programme and UFES, adhering to international ethical standards for research involving human subjects.

4.1 Quantitative Analysis

All performance analyses refer exclusively to the NTD classifier (second-stage model). A total of 439 clinical images from 208 clinically confirmed leprosy cases were initially included in the external dataset; after excluding 16 images due to processing errors, 423 images were analysed, comprising 367 classical leprosy lesions (Subset 1) and 56 reactional or atypical leprosy-related presentations (Subset 2). Overall, leprosy was identified within the Top-5 predictions in 84.9% (359/423) of cases. Sensitivity was 87.2% (320/367) for classical presentations and 69.6% (39/56) for reactional or atypical presentations, as summarised in Table 2.

Table 2

Sensitivity (recall) of the AI model for each subset
Metric	Subset 1	Subset 2	Combined Total
Sensitivity (recall)	87.2%	69.6%	84.9%

The Top-N analysis revealed that leprosy was ranked first in 46.9% of classical cases and 30.4% of reactional or atypical cases, highlighting performance disparities between subsets. Table 3 presents the proportion of cases in which leprosy appeared at each rank position within the model’s five most probable diagnostic outputs.

Table 3

Sensitivity (recall) of the AI model within Top-N algorithm predictions
Dataset	Top-1	Top-2	Top-3	Top-4	Top-5
Subset 1	46.9%	63.5%	73.8%	80.1%	87.2%
Subset 2	30.4%	46.4%	58.9%	62.5%	69.6%

4.2 Case-Level Error Patterns

In multiple cases, images from the same patient produced inconsistent diagnostic outputs. For example, hypopigmented macules from a single patient were classified both with and without leprosy among the Top-5 predictions, depending on the anatomical location. Similarly, a patient with typical facial infiltration and madarosis was misclassified, with leprosy absent from the model’s Top-5 predictions. Nodular and infiltrative lesions—especially those with necrotic features consistent with necrotic erythema nodosum leprosum—were often misclassified, with leprosy not appearing among the Top-5 suggested diagnoses. Hyperpigmented lesions, on the other hand, showed higher rates of correct identification. These inconsistencies suggest difficulties in handling intra-patient variability and in recognising reactional or complex inflammatory patterns.

This is the first article to report an independent external analytical validation of the WHO Skin NTDs app’s AI model for leprosy diagnosis. Using a large and diverse dataset of clinically confirmed cases collected over nearly three decades, we evaluated the model’s ability to recognise both classical leprosy lesions and reactional forms across a broad spectrum of clinical presentations. A key strength of this work lies in the comprehensiveness of the external dataset, which includes images from multiple anatomical sites and spans nearly the full range of Fitzpatrick skin phototypes (all except type I). This diversity enhances the generalisability of the findings and aligns with WHO regulatory expectations for external validation, which requires performance assessment using datasets entirely independent of model development.²²

Overall, the AI model demonstrated high sensitivity, identifying leprosy within the Top-5 predictions in 84.9% of images. As expected, performance was stronger for classical lesions (87.2%) than for reactional or atypical presentations (69.6%), which are inherently more difficult to diagnose due to their variable and often inflammatory nature. This discrepancy underscores the need for targeted refinement of the model to better capture immune-mediated or atypical forms. Subset characteristics may also have contributed: Subset 2 consisted predominantly of adults with intermediate skin phototypes and displayed a more homogeneous phenotypic profile, whereas Subset 1 included a broader distribution of ages and pigmentation patterns. The Top-N analysis reinforced this performance gap, with leprosy ranking first in nearly half of classical lesions but in only one-third of reactional or atypical lesions.

A recent abstract by Madhumita (2025)²⁵ presented a preliminary evaluation of the WHO Skin NTDs app’s AI algorithm using 300 leprosy cases, reporting higher sensitivity for polar forms and reduced accuracy in borderline and reactional states. Our findings extend this early evidence by providing a more robust external validation with a larger dataset and by incorporating a detailed qualitative assessment of error patterns.

Beyond aggregate performance metrics, qualitative analysis revealed inconsistencies in predictions for visually similar lesions from the same patient. These inconsistencies may stem from limited intra-patient variability within the training data or from challenges in capturing spatial context. Necrotic, erythematous, and infiltrative lesions were particularly prone to misclassification, reinforcing known limitations of convolutional models in recognising complex inflammatory patterns.²⁶ To address these issues, future development should prioritise dataset enrichment with underrepresented reactional forms and rare phenotypes, which remain poorly represented in most existing AI pipelines for leprosy and other skin NTDs.^15,16 Complementary approaches such as semantic segmentation or other pixel-level techniques may help the model focus on lesion morphology rather than background or surrounding skin, an approach shown to improve performance in dermatological image analysis, particularly for erythematous, necrotic, and infiltrative lesions.^9,27 Incorporating patient-level context—through multi-image analysis, longitudinal image sets, or whole-body frameworks—could also reduce inconsistencies across images from the same case.^10,28 In addition, ensemble learning architectures may enhance stability and generalisability of diagnostic outputs across heterogeneous clinical presentations.^12,29 These refinements would better align the model’s capabilities with the complex spectrum of leprosy manifestations observed in clinical practice.

The practical relevance of this app is considerable, particularly in primary-care or low-resource settings where access to dermatologists or leprosy specialists is scarce. In such contexts, AI-generated ranked differentials—even when the correct diagnosis is not first—can facilitate case suspicion and prompt referral.^6,7 This potential is underscored by comparisons with the diagnostic accuracy of primary-care clinicians for common skin diseases, which ranges from 19% to 36%.^21,30 Considering that leprosy is far less common and often unfamiliar to general practitioners, a tool achieving 87.2% sensitivity for classical presentations represents a meaningful adjunct to clinical care.

Leprosy remains a significant public health challenge in many endemic regions.¹ Early detection is essential to prevent disability and interrupt transmission. Digital tools such as the WHO Skin NTDs app may help reduce diagnostic disparities, especially where specialists are limited. By combining image-based analysis with an accessible, offline-capable mobile interface, the app offers both decision-support and educational value for non-specialist providers.^2,17 However, despite growing evidence that deep learning systems can approach or exceed clinician performance in specific tasks, clinical adoption remains limited by the scarcity of robust external evaluations. Inconsistent reporting practices and poor transparency further restrict trust and interpretability, highlighting the need for standardised evaluation frameworks tailored to AI in medical imaging.^12,22,31

This study has several limitations. First, the dataset included only confirmed leprosy cases, preventing the calculation of specificity or false-positive rates. Second, this was not a prospective implementation study; thus, real-world workflow integration, user behaviour and clinical impact were not assessed. Third, we did not include direct comparisons with human raters, which are essential for contextualising model performance and informing clinical deployment strategies.^12,22 Fourth, although the AI inference mechanism remains the same across platforms, testing was performed on the DHIS2 interface rather than via the mobile app, and field conditions may introduce additional variability. Finally, most images depicted isolated lesions without accompanying metadata, limiting the ability to emulate real-world diagnostic reasoning.^15,32 Previous work suggests that combining images with structured clinical information substantially improves performance,¹² and future iterations could benefit from such multimodal input.¹⁰

In conclusion, the WHO Skin NTDs app demonstrates considerable promise as a clinical decision-support tool for leprosy, especially in endemic and resource-limited settings. This study confirms its good performance in detecting classical leprosy lesions while identifying persistent gaps in the recognition of reactional and atypical forms. To improve its clinical utility and reliability, we recommend: (i) developing external evaluation datasets that include a broad range of non-leprosy and look-alike conditions to enable robust estimation of specificity and false-positive rates; (ii) increasing the representation of reactional and rare phenotypes in future evaluations and model refinement; and (iii) conducting clinician comparison studies and prospective field evaluations. With continued refinement and responsible integration, AI-powered digital tools such as the WHO Skin NTDs app can strengthen early detection, reduce disability, and accelerate progress towards global elimination targets.

Weekly Epidemiological Record. Global leprosy (Hansen disease) update, 2024: beyond zero cases – what elimination of leprosy really means. Wkly Epidemiol Rec (2025) ;100(37):365–84. Available from: https://www.who.int/publications/i/item/who-wer10037-365-384 (accessed Nov 21, 2025)
Grijsen ML, Nguyen TH, Pinheiro RO et al (2024) Leprosy. Nat Rev Dis Primers 10(1):90. 10.1038/s41572-024-00575-1
Yotsu RR, Fuller LC, Murdoch ME et al (2023) A global call for action to tackle skin-related neglected tropical diseases (skin NTDs) through integration: An ambitious step change. PLoS Negl Trop Dis 17(6):e0011357. 10.1371/journal.pntd.0011357
Deps PD, Kahawita I, Yotsu R, Lambert SM (2025) Leprosy research updates: Shaping the future of global health. Indian J Med Res 161(2):120–124. 10.25259/IJMR_195_2025
Deps PD, Guedes BV, Bucker Filho J, Andreatta MK, Marcari RS, Rodrigues LC (2006) Delay in the diagnosis of leprosy in the Metropolitan Region of Vitória, Brazil. Lepr Rev 77(1):41–47
Fernandes JRN, Teles AS, Fernandes TRS et al (2023) Artificial Intelligence on Diagnostic Aid of Leprosy: A Systematic Literature Review. J Clin Med 13(1):180. 10.3390/jcm13010180
Deps PD, Yotsu R, Furriel BCRS, de Oliveira BD, de Lima SL, Loureiro RM (2024) The potential role of artificial intelligence in the clinical management of Hansen's disease (leprosy). Front Med (Lausanne) 11:1338598. 10.3389/fmed.2024.1338598
Esteva A, Kuprel B, Novoa RA et al (2017) Dermatologist-level classification of skin cancer with deep neural networks. Nature 542(7639):115–118. 10.1038/nature21056
Shin HC, Roth HR, Gao M et al (2016) Deep Convolutional Neural Networks for Computer-Aided Detection: CNN Architectures, Dataset Characteristics and Transfer Learning. IEEE Trans Med Imaging 35(5):1285–1298. 10.1109/TMI.2016.2528162
Panagoulias DP, Tsoureli-Nikita E, Virvou M, Tsihrintzis GA (2025) Dermacen analytica: A novel methodology integrating multi-modal large language models with machine learning in dermatology. Int J Med Inf 199:105898. 10.1016/j.ijmedinf.2025.105898
Liu Z, Zhang Y, Wang K, Xie F, Liu J (2025) Early diagnosis model of mycosis fungoides and five inflammatory skin diseases based on a multimodal data-based convolutional neural network. Br J Dermatol 193(5):968–977. 10.1093/bjd/ljaf212
Liu X, Faes L, Kale AU et al (2019) A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis. Lancet Digit Health 1(6):e271–e297. 10.1016/S2589-7500(19)30123-2
Winkler JK, Blum A, Kommoss K et al (2023) Assessment of Diagnostic Performance of Dermatologists Cooperating With a Convolutional Neural Network in a Prospective Clinical Study: Human With Machine. JAMA Dermatol 159(6):621–627. 10.1001/jamadermatol.2023.0905
Baweja HS, Parhar T. Leprosy lesion recognition using convolutional neural networks. International Conference on Machine, Learning, Cybernetics (2016) (ICMLC), Jeju, Korea (South), 2016, pp. 141–145. 10.1109/ICMLC.2016.7860891
Barbieri RR, Xu Y, Setian L et al (2022) Reimagining leprosy elimination with AI analysis of a combination of skin lesion images with demographic and clinical data. Lancet Reg Health Am 9:100192. 10.1016/j.lana.2022.100192
Yotsu RR, Ding Z, Hamm J, Blanton RE (2023) Deep learning for AI-based diagnosis of skin-related neglected tropical diseases: A pilot study. PLoS Negl Trop Dis 17(8):e0011230. 10.1371/journal.pntd.0011230
Sall D, Jockers D, Dioussé P et al (2025) Potential of the World Health Organization's Skin NTDs App to Support and Improve the Detection of Skin-Related Neglected Tropical Diseases: Protocol for a Performance Evaluation and Feasibility Study in Senegal. JMIR Res Protoc 14:e69420. 10.2196/69420
Cano M, Ruiz-Postigo JA, Macharia P et al (2024) Evaluating the World Health Organization's SkinNTDs App as a Training Tool for Skin Neglected Tropical Diseases in Ghana and Kenya: Cross-Sectional Study. J Med Internet Res 26:e51628. 10.2196/51628
Frej A, Cano M, Ruiz-Postigo JA et al (2022) Assessing the Quality of the World Health Organization's Skin NTDs App as a Training Tool in Ghana and Kenya: Protocol for a Cross-sectional Study. JMIR Res Protoc 11(12):e39393. 10.2196/39393
World Health Organization. Skin-related neglected tropical diseases: WHO launches new version of the WHO Skin NTDs mobile application [Internet] (2023) Available from: https://www.who.int/news/item/09-10-2023-skin-related-neglected-tropical-diseases--who-launches-new-version-of-the-who-skin-ntds-mobile-application (accessed Nov 21, 2025)
Dulmage B, Tegtmeyer K, Zhang MZ, Colavincenzo M, Xu SA, Point-of-Care (2021) Real-Time Artificial Intelligence System to Support Clinician Diagnosis of a Wide Range of Skin Diseases. J Invest Dermatol 141(5):1230–1235. 10.1016/j.jid.2020.08.027
World Health Organization. Regulatory considerations on artificial intelligence for health [Internet]. Geneva: WHO (2023) Available from: https://www.who.int/publications/i/item/9789240078871 (accessed Nov 21, 2025)
World Health Organization. Guidelines for the diagnosis, treatment and prevention of leprosy [Internet] (2018) Available from: https://www.who.int/publications/i/item/9789290226383 (accessed Nov 21, 2025)
World Health Organization. Leprosy/Hansen disease: management of reactions and prevention of disabilities [Internet] (2020) Available from: https://www.who.int/publications/i/item/9789290227595 (accessed Nov 21, 2025)
Madhumita M (2025) Retrospective evaluation of the WHO Skin NTD App for diagnosing leprosy. In: Book of Selected Abstracts – 2nd WHO Global Meeting on Skin NTDs; Mar 25–28; Geneva, Switzerland. World Health Organization; 2025. pp. 258–260
Schielein MC, Christl J, Sitaru S et al (2023) Outlier detection in dermatology: Performance of different convolutional neural networks for binary classification of inflammatory skin diseases. J Eur Acad Dermatol Venereol 37(5):1071–1079. 10.1111/jdv.18853
Kshatri SS, Singh D (2023) Convolutional neural network in medical image analysis: a review. Arch Comput Methods Eng 30(5):2793–2810. 10.1007/s11831-023-09898-w
Primiero CA, Betz-Stablein B, Ascott N et al (2024) A protocol for annotation of total body photography for machine learning to analyze skin phenotype and lesion classification. Front Med (Lausanne) 11:1380984. 10.3389/fmed.2024.1380984
Liu Y, Jain A, Eng C et al (2020) A deep learning system for differential diagnosis of skin diseases. Nat Med 26(6):900–908. 10.1038/s41591-020-0842-3
Groh M, Badri O, Daneshjou R et al (2024) Deep learning-aided decision support for diagnosis of skin disease across skin tones. Nat Med 30(2):573–583. 10.1038/s41591-023-02728-3
Tejani AS, Klontzas ME, Gatti AA et al (2024) Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. Radiol Artif Intell 6(4):e240300. 10.1148/ryai.240300
Mendonça Ramos Simões M, Rocha Lima F, Barbosa Lugão H et al (2025) Development and validation of a machine learning approach for screening new leprosy cases based on the leprosy suspicion questionnaire. Sci Rep 15(1):6912. 10.1038/s41598-025-91462-6

The authors declare potential competing interests as follows: Dr Ruiz-Postigo is affiliated with the World Health Organization (WHO); this affiliation did not influence the study design, data analysis, or interpretation of the results. The other authors declare no conflicts of interest.

Download PDF

Version 1

posted

You are reading this latest preprint version

Independent Assessment of the WHO Skin NTDs App: A Suitable Tool for Leprosy Detection

Status:

Version 1

Abstract

Objective

Methods

Findings:

Conclusion

1. Introduction

2. Objectives

3. Methodology

3.1 Study Design

3.2 Data Collection and Dataset

3.2.1 Image Acquisition and Data Security

3.2.2 External Validation Dataset

3.2.3 Diagnostic Confirmation

3.3 Outcome Measures and Evaluation Approach

3.4 Statistical Analysis

3.5 Ethical Compliance

4. Results

4.1 Quantitative Analysis

4.2 Case-Level Error Patterns

5. Discussion

References

Additional Declarations

Status:

Version 1