Scaling trachoma surveillance in endemic areas using machine learning

Abstract

Background

Trachoma remains a leading infectious cause of blindness in endemic regions despite progress towards global elimination. Accurate and scalable diagnosis remains challenging in areas with limited ophthalmic expertise. We developed and evaluated a reproducible transfer learning model using eyelid image-based data.

Methods

We retrospectively analysed anonymised inner eyelid photographs collected from trachoma prevalence surveys conducted in Ethiopia, Tanzania, Australia, Solomon Islands, Colombia, the Gambia, and Guatemala (n = 572 images). Images were categorized as trachoma (n = 251) or not trachoma (n = 321) based on consensus grades from the Tropical Data certified graders. Data were processed and analysed in R (version 4.4.1) using the keras and tensorflow packages interfaced with Python (version 3.10) through reticulate. A ResNet50 convolutional neural network pretrained on ImageNet was fine-tuned for binary classification. The model was trained for up to 30 epochs with early stopping and adaptive learning-rate reduction. Performance on a held-out validation set (n = 62, 12%) was evaluated using accuracy, sensitivity, specificity, and Cohen’s κ.

Results

The ResNet50 model achieved an overall accuracy of 85.4% (95% CI 76.3–92%) on the validation set. The model achieved a sensitivity of 80.9% for detecting trachoma and a specificity of 90.5% for correctly identifying non-trachoma eyes. Agreement between predictions and grader labels was substantial (κ = 0.71, 95% CI 0.58–0.84).

Conclusion

Our findings demonstrate the feasibility of using a reproducible R-based deep learning pipeline for automated trachoma classification in large-scale surveys. These findings demonstrate the feasibility of automated trachoma classification using a transfer-learning framework. Larger datasets will be required before operational deployment in surveillance programmes.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12879-026-13215-8.

Introduction

Trachoma, a neglected tropical bacterial disease caused by Chlamydia trachomatis serotypes A, B, Ba and C, remains the leading infectious cause of blindness worldwide, with approximately 103 million people living in endemic areas at risk globally, and children are disproportionately affected [1–3]. Africa bears the greatest burden, accounting for approximately 68.5% (27.8 million) of all active cases [1, 4]. Transmission occurs through contact with infected ocular or nasal secretions, contaminated fomites, and via mechanical vectors such as the bazaar fly (Musca sorbens), which carries bacteria from infected to uninfected individuals [5]. Ocular infection causes follicular conjunctivitis, often presenting as trachomatous inflammation - follicular (TF). Repeated infections lead to scarring of the eyelid margin, which can turn inwards, causing trachomatous trichiasis (TT). In TT, eyelashes abrade the cornea, ultimately resulting in irreversible blindness [6]. TT is more prevalent in women and the elderly, likely due to child-care roles and higher lifetime infection exposure, respectively [2, 7]. Systematic case detection and surgical management of TT are essential to prevent vision loss in endemic communities [8]. Engaging communities in endemic areas helps address cultural and economic barriers, and inform health policies and funding priorities aimed at enhancing sanitation [1].

The World health Organization (WHO) led Global Elimination of Trachoma (GET2020) initiative missed its 2020 target through the SAFE strategy (Surgery, Antibiotics, Facial cleanliness, and Environmental improvement), with a revised goal for elimination by 2030 [9, 10]. Elimination of trachoma as a public health problem is defined by three criteria: a prevalence of trachomatous trichiasis (TT) unknown to the health system of less than 0·2% among adults aged 15 years or older in each formerly endemic district; a prevalence of trachomatous inflammation - follicular (TF) of less than 5% among children aged 1–9 years, sustained for at least two years without ongoing mass drug administration (oral azithromycin or 1% tetracycline eye ointment); and the presence of a functional system to detect and manage incident TT cases [6]. Despite significant progress in Morocco, The Gambia, and Ghana following early SAFE implementation [11], trachoma remains highly endemic in several African countries. Ethiopia has the highest burden, especially among children aged 1–9 years [1, 3, 12], while Tanzania, Sudan, Niger, Guinea, Uganda, and Kenya have yet to meet elimination thresholds [1, 5, 13–15]. Endemic countries have integrated trachoma elimination into coordinated Neglected Tropical Diseases control programmes, recognizing that economic benefits substantially outweigh implementation costs [11, 16].

Standardised population-based prevalence surveys coordinated by the Tropical Data programme, a WHO-led initiative to classify trachoma and non-trachoma photographs, have provided critical epidemiological data for targeted SAFE implementation [6, 17]. Briefly, 2.5× magnifying loupes are used to examine consenting individuals eyes by trained graders for evidence of trichiasis, and metadata on global positioning system (GPS) coordinates, household-level data on access to water, sanitation, and hygiene are recorded. However, as prevalence declines, larger sample sizes and longitudinal data are required to evaluate risk factors and prevalence, and grading inconsistencies and sampling bias can reduce precision, despite rigorous initial training, even among seasoned graders [1, 4, 18].

Sustained progress toward trachoma elimination is impeded by persistent challenges, including long-term conflict and insecurity limiting fieldwork [6, 19, 20], challenges in surveillance among pastoral nomadic populations traversing endemic regions and international borders [15], overwhelmed ophthalmic clinical officers [4, 7, 21], and prevalence remaining above elimination thresholds despite prolonged mass drug administration [2, 22, 23]. A standardised, reproducible eye image analytical pipeline in R, will help to scale automated trachoma detection reducing the burden of manual expert image review, thereby extending trachoma surveillance [4].

Image-based grading allows remote validation and consistent case assessment at scale, but requires substantial expert time [2, 6, 24]. Machine-learning approaches, particularly convolutional neural networks (CNNs), have shown promise for image-based disease diagnostics [25–28]. However, previous machine-learning studies for trachoma detection were constrained by limited datasets or single-country data sources, limiting their generalizability for surveillance applications [29, 30]. We therefore developed a reproducible R-based transfer learning workflow using diverse, multi-country Tropical Data images to evaluate the potential of automated trachoma classification for surveillance in public health systems in resource-limited settings.

Materials and methods

Study design and data sources

We retrospectively analysed anonymised inner eyelid photographs collated on the Tropical Data database (www.tropicaldata.org) accessed on 26th September 2025, which archives images from standardized trachoma prevalence surveys, conducted between 2006 and 2022. Images were from surveys in Ethiopia, Tanzania, Australia, Solomon Islands, Colombia, the Gambia, and Guatemala (n = 572 images). Images were categorized as trachoma (n = 251) or not trachoma (n = 321) based on consensus grades from the Tropical Data certified graders. A limitation of this dataset was the aggregation of trachomatous trichiasis (TT) and trachomatous inflammation - follicular (TF) images into a single category. This approach was necessitated by inconsistent labelling across the available images, whereby not all trachomatous cases were distinctly classified by clinical stage. Furthermore, only this finite sample size was available for analysis. Although relatively modest, the dataset was analysed using transfer learning with a pretrained architecture, which allows effective training with limited image datasets.

Analysis software

All analyses were conducted in R (version 4.4.1) with the keras and tensorflow packages interfacing Python via the reticulate package. A Python (version 3.10) environment with TensorFlow (version 2.16.2) served as the backend. Additional R packages used included: caret for model evaluation and confusion matrix reporting.

Model development and evaluation

Images were randomly divided into training and validation sets using an 80:20 split, with stratification by class (trachoma vs. not trachoma). The split was seeded to ensure reproducibility.

Images were first standardized to 224 × 224 pixels with three red, green, and blue (RBG) channels, and rescaled to values between 0 and 1 to improve generalizability. The data was augmented as follows: random rotation up to 20°, width and height shifts up to 10%, brightness adjustments between 0.8 and 1.2, random zoom up to 20%, and horizontal flipping. Validation images were rescaled but not augmented.

We implemented a pre-trained transfer learning computer model, ResNet50, to extract features. ResNet50 is a 50-layer deep convolutional neural network (CNN) from the Residual Network (ResNet) pre-trained on ImageNet. ResNet50 was selected because it mitigates the vanishing gradient problem and demonstrates strong performance in medical image classification tasks, performing well with transfer learning even on relatively small datasets. The convolutional base was frozen during initial training to preserve pre-trained weights. The following layers were added to the base model: global average pooling, dropout (rate = 0.4) for regularization, a fully connected dense layer with 256 activated rectified linear units (ReLU) units followed by batch normalization, and a dense sigmoid-activated unit for binary classification.

Models were trained for up to 30 epochs (cycles through the entire dataset), and optimization was performed with the Adam optimizer (learning rate = 1e-5). Two call-back strategies: early stopping based on validation loss (patience = 5, restoring best weights), and learning rate reduction on plateau (factor = 0.5, patience = 2), helped reduce overfitting.

To evaluate the model, predictions were performed on the validation dataset. Performance was evaluated using a confusion matrix and associated metrics (sensitivity, specificity, accuracy, and Cohen’s kappa (κ)) computed with the caret package.

Ethics and data access approval

Permission was granted to access data from the Tropical Data database (www.tropicaldata.org) by the Tropical Database Steering committee. The Tropical Data database program received ethical approval from the London School of Hygiene & Tropical Medicine committee. No new ethical approval was required for this secondary data analysis. These data is anonymized, non-identifiable, and publicly accessible.

Data availability

The R code and trained model weights are available in the supplementary materials.

Results

Overall, the model achieved an accuracy of 85.4% (95% CI 76.3–92%). The model achieved a sensitivity of 80.9% for detecting trachoma and a specificity of 90.5% for correctly identifying non-trachoma eyes. Agreement between predictions and grader labels was substantial (κ = 0.71, 95% CI 0.58–0.84). The model correctly classified 76 of 89 validation images, with balanced performance across both classes (balanced accuracy 85.7%).

Implementation of the model will reduce the number of images requiring expert manual grading while maintaining diagnostic reliability (Fig. 1).

Fig. 1.

Model performance visualization. (A) Schematic of the CNN model used to distinguish images from trachoma and none trachoma cases. (B) Summary of the confusion-matrix metrics from the validation dataset. (C) Performance of the CNN model. (D) Training and validation accuracy over 30 epochs showing good learning. (E) Training and validation loss over 30 epochs showing how error decreases

Our findings demonstrate the feasibility of using a reproducible R-based deep learning pipeline for automated trachoma classification in large-scale surveys. Integrating such tools into trachoma surveillance could support resource-limited national programmes by enabling rapid, consistent image grading and data-driven monitoring of elimination targets.

Discussion

Visual impairment affects over 250 million people globally, including 36 million who are blind, with aetiology varying by setting: age-related conditions such as diabetic retinopathy, macular degeneration, and glaucoma predominate in high-income countries [31, 32], while trachoma remains the leading infectious cause of blindness in low-income settings, where the burden is greatest [1, 4, 32].

Sustained progress toward trachoma elimination is critical [4], but impeded by challenges that limit conventional surveillance: long-term conflict restricting fieldwork [6, 19, 20], difficulties reaching pastoral nomadic communities traversing endemic regions [15], and grading inconsistencies in large-scale eye examination campaigns [4, 7, 21]. Prevalence remains above elimination thresholds even in endemic settings without conflict despite years of mass drug administration interventions [2, 22, 23], underscoring the need for scalable, reproducible analytical approaches that reduce reliance on in-person expert review [4, 7, 21].

Recrudescence of infection after mass drug administration and surgical corrections suggest environmental factors and behavioural change play a strong role in trachoma elimination [5, 18, 23]. Significant scale-up of trainings for ophthalmic surgeons, nurses and officers is crucial to achieve the new 2030 elimination targets [14, 33], however this does not rule out misclassification of trachoma signs especially in large eye examination clinical campaigns [17]. Sustained surveillance that effectively uses robust data for decision making is essential to complement monitoring: trachoma recrudescence, poor surgical intervention outcomes, antibiotic availability, attrition of ophthalmic nurses and trichiasis surgeons, and evaluate the quality of post-elimination care [9].

Advances in bioinformatics are increasingly becoming important to ophthalmic research and care, offering new insights to help mitigate this growing global burden. We demonstrate the feasibility of developing an automated trachoma classification model using a reproducible R-based deep learning pipeline and multi-country survey data. These results are encouraging in the context of existing field grading challenges. Our analysis approach addresses both dataset diversity and accessibility for epidemiologists. It reduces the number of images requiring manual expert review. By using a consistent way to pre-process eye photographs, a pre-trained computer model (ResNet50), and an analysis setup that can be repeated in R, we create a practical approach that could make trachoma detection faster and easier to use.

Although, establishing elimination status in endemic areas is of paramount importance. It is not feasible to diagnose disease or recrudescence status for each member of an at-risk population using in-person community trachoma case finders [18]. More efficient data collation approaches are required to improve the reach by prevalence surveys, especially for hard-to-reach populations. The use of photography not only enables large eye examinations, but provides an opportunity for re-evaluation. The use of machine learning approaches to identify trachoma cases helps to minimise human grading error, while making it feasible to analyse large eye examination datasets [6]. This approach that leverages machine learning to interpret images and establish the prevalence of trachoma can be deployed at a granular level, even among hard-to-reach communities, like those in endemic areas experiencing conflict and among pastoral nomadic communities. Integration of such models into national trachoma surveillance could enhance the precision of prevalence estimates, support monitoring of recrudescence, and improve resource allocation in endemic and post-elimination settings.

Our analysis was limited to binary classification (trachoma versus not trachoma) due to limitations in the available labels. Because TF and TT represent distinct clinical stages with different elimination thresholds, combining them into a single positive class introduces label heterogeneity that may affect model training dynamics, influence class balance, and limit interpretability of the sensitivity estimates. That is, it may not be possible to determine whether the model is specifically detecting features of TF, TT, or both. Separating these stages in future work will be essential for programme decision-making deployment, and for understanding what the model has learned at great detail. Additional evaluation metrics such as receiver operating characteristic (ROC) curves, area under the curve (AUC), and predictive values could provide further insight into model performance, particularly in low-prevalence settings typical of trachoma elimination programmes. These metrics were not evaluated in the present feasibility analysis and can be explored in future studies.

This study demonstrate the feasibility of using a reproducible R-based deep learning pipeline for automated trachoma image classification using transfer learning. Future work using larger datasets will enable seamless integration into surveillance programmes.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(2.1KB, zip)}

Author contributions

BK and CW conceived the study. BK performed analysis. BK and CW wrote the first draft. BK and CW reviewed final draft.

Funding

This study was not supported by any funding source.

Data availability

The R code and trained model weights are available in the supplementary materials.

Declarations

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent to publication

Not applicable.

Competing interests

The authors declare no competing interests.