Thalassemia[1] represents a major global health challenge due to its high prevalence and the substantial medical, social and economic burden associated with its management, particularly in low- and middle-income countries (LMICs) such as Sri Lanka [2]. It is a heterogeneous group of genetic disorders characterised by defective synthesis of hemoglobin chains, leading to anemia, ineffective erythropoiesis, and hemolysis [3]. The two principal forms, alpha-thalassemia and beta-thalassemia, vary widely in clinical severity, ranging from asymptomatic microcytosis to life-threatening anemia requiring lifelong transfusion therapy. In Sri Lanka, approximately 3,000 patients [4] depend on regular transfusions, consuming a significant portion of the national healthcare budget [5] [6].

Early screening and diagnosis are critical for preventing thalassemia major births. Current screening methods in Sri Lanka involve assessing red blood cell indices from a full blood count (FBC), peripheral blood smear examination by haematologists, and confirmatory tests such as High-Performance Liquid Chromatography (HPLC). While HPLC and genetic analysis provide very high diagnostic accuracy, these methods are time-consuming, expensive, and dependent on specialised expertise, making widespread implementation challenging. Despite Sri Lanka’s existing prevention program, the birth rate of thalassemia major cases remains inadequately controlled, demonstrating the need for more effective screening strategies. A simple, cost-effective, yet highly sensitive screening approach is therefore needed—one that can efficiently identify individuals at risk and prioritise them for confirmatory testing without overburdening clinicians and laboratory personnel. Such a method should ideally function with basic RBC data and blood smear images, even in rural primary care settings lacking haematologists.

Machine learning has emerged as a powerful tool in medical diagnostics, offering potential improvements in efficiency, accuracy, and cost reduction. Various machine learning approaches have been applied to haematological disorders, but research on thalassemia screening remains limited. Previous studies have either focused on RBC-based classification models or blood image analysis, but not both simultaneously. Integrating these two approaches could significantly enhance diagnostic reliability by combining morphological and haematological features.

However, a critical gap persists: current digital tools do not provide an integrated ML pipeline capable of processing both numerical RBC indices and smear-image features, nor are they designed for practical deployment in LMICs. Many existing systems require high-quality laboratory imaging, powerful computing resources, or specialised environments, making them unsuitable for decentralised or primary-care screening.

In LMICs, the broader deployment of AI tools is challenged by variability in smartphone imaging quality, limited computational capacity, intermittent internet connectivity, and difficulties integrating digital tools into routine health workflows. Addressing these context-specific barriers is essential for achieving real-world impact. A notable innovation of the present study is the development of a machine learning pipeline explicitly designed to operate on average-priced smartphones, without requiring expensive high-tech equipment or internet connectivity. The system features a simple user interface, enabling healthcare workers to quickly and accurately screen individuals, triage a small number for confirmatory testing, and facilitate decentralised testing with minimal infrastructure. Embedding such a tool within national screening pathways has significant potential to enhance prevention efforts, particularly as Sri Lanka prepares to implement premarital thalassemia screening policies.

This study aims to develop a machine-learning–based screening system optimised for high sensitivity and practical implementation, leveraging convolutional neural networks, transfer learning, and complementary ML techniques. A notable innovation of the present study is the development of an integrated pipeline that fuses Red Blood Cell (RBC) indices with CNN-derived blood smear image features into a unified, deployable screening system. The pipeline is explicitly designed to operate on average-priced smartphones, without requiring expensive high-tech equipment or internet connectivity, and features a simple user interface that enables healthcare workers to quickly and accurately screen individuals, triage a small number for confirmatory testing, and facilitate decentralised testing with minimal infrastructure, enabling use even in rural primary care settings without haematologists.

This approach is relevant not only for thalassemia prevention but also for broader digital health implementation, as it demonstrates how AI tools can be adapted for low-resource settings, embedded into existing care pathways, and scaled to support national public-health strategies. By offering a sensitive, efficient, and sustainable solution, the proposed system has significant potential to strengthen thalassemia screening, particularly as Sri Lanka prepares to implement premarital screening policies and expand population-level preventive interventions.

Machine learning techniques have increasingly been applied to automate microscopic image analysis in haematological cytology, addressing critical challenges in blood image processing. The accuracy of these methods depends on multiple factors, including slide preparation, staining quality, and digitisation settings. Optimising algorithm performance often requires standardised image acquisition protocols to ensure consistency and reliability across datasets [7]. One major challenge in blood smear analysis is detecting rare cells in peripheral blood. While analysing common blood cells can be done with small sample sections, the same approach may not be effective for rare cell detection due to limited imaging, potentially leading to false negatives. To mitigate this, studies emphasise the need for analysing larger slide areas, where image acquisition speed becomes a crucial factor [8].

To transition from research to practical application, systematic validation strategies must be established in laboratory environments. Defining comprehensive screening protocols will be essential to maximise the potential of automated recognition methodologies, making them suitable for clinical deployment [9].

Several studies have focused on automated thalassemia detection. Sandanayake et al. [10] developed a system to identify thalassemia using RBC images and machine learning. The system extracted features such as cell shape, size, and colour using edge detection (Canny), Hough Transform, and ellipse fitting, alongside quantitative parameters including RBC count, mean corpuscular volume (MCV), mean corpuscular haemoglobin (MCH), mean corpuscular haemoglobin concentration (MCHC), red blood cell distribution width (RDW), and haemoglobin content. Classification was performed using Artificial Neural Networks (ANN) and Support Vector Machines (SVM), achieving high accuracy (89.4% for females and 97% for males). The approach included a web interface for patient interaction, improving accessibility. However, the system relied solely on RBC morphological features without integrating RBC indices or other haematological parameters, limiting potential diagnostic accuracy. Validation was performed on a relatively small and gender-specific dataset, and modern deep learning methods such as CNNs or transfer learning were not explored.

Khan et al. [11] developed a deep learning approach using Hb electrophoresis images, achieving up to 95.8% accuracy with pre-trained CNNs and confirming the model's focus on relevant haemoglobin bands. Although HPLC remains faster, more accurate, and convenient for confirmation, this study provides insights into transfer learning, pre-processing techniques, and interpretability methods such as Score-CAM, which are valuable for developing image-based screening models like ours.

ThalPred is a machine learning tool designed to discriminate thalassemia trait (TT) from iron deficiency anaemia (IDA) using routine RBC parameters. Differentiating these two causes of hypochromic microcytic anaemia is challenging, especially in Thailand, due to the variability of RBC indices across populations. Using 186 adult samples, multiple models, including decision trees, random forests, SVM, and neural networks, were tested. While traditional indices like KF2 and RI performed moderately well, machine learning models, particularly SVM, achieved higher accuracy and robustness even on independent validation sets. Additionally, interpretable rules extracted from random forests enabled clinicians to screen using simple thresholds of Hb, RBC, RDW, and MCHC. Compared to our model, ThalPred emphasises the value of SVM for imbalanced datasets and highlights the importance of interpretability in clinical tools, reinforcing the benefit of combining robust algorithms with transparent decision rules.

Lin et al. [12] used a Mask R-CNN-based deep learning model to detect and segment thalassemic RBCs from quantitative phase images, achieving 97.8% classification accuracy. Their approach combined instance segmentation with feature-based classifiers to extract single-cell morphology and haemoglobin content, providing detailed interpretability. However, the reliance on specialised digital holographic microscopy limits accessibility and real-time applicability. In contrast, our model uses full blood count parameters and blood smear images captured through a mid-range smartphone via a microscope eyepiece, offering greater accessibility and feasibility for widespread use. Insights from Lin et al. underscore the value of combining morphological information with quantitative features, guiding our approach to integrate image- and parameter-based data for accurate and interpretable thalassemia screening.

Sadiq et al. [13] proposed an ensemble model, SGR-VC, combining SVM, Gradient Boosting Machine (GBM), and Random Forest (RF) to classify carriers and non-carriers using a dataset of 5,066 patients from the Punjab Thalassemia Prevention Program. The model achieved 93% accuracy, precision, recall, and F1-score, outperforming individual classifiers. While earlier works primarily focused on distinguishing β-thalassemia from iron deficiency anaemia or on specific subgroups such as women of fertile age, SGR-VC demonstrates a more generalisable approach across genders and age groups. However, gaps remain in exploring deep learning methods, regression-based models, and datasets from diverse geographic populations, which could further improve predictive power. Insights from this study highlight the effectiveness of ensemble approaches in leveraging complementary strengths of tree-based and probability-based classifiers, and the importance of balanced, normalised datasets for robust carrier detection. Compared to our model, SGR-VC serves as a benchmark showing that combining multiple classifiers can reduce misclassification and improve precision, providing a cost-effective, rapid screening tool suitable for large-scale public health applications.

Purwar et al. [14] proposed a novel approach for thalassemia detection by fusing deep image features from blood smear images with clinical features from RBC indices. Using AlexNet CNN to extract morphological features and ten clinical parameters, such as RBC count, haemoglobin, MCV, and haemoglobin A2, the study applied PCA to reduce redundancy and prevent overfitting in the high-dimensional fused feature set (1011 features) relative to a small sample size (20 patients). Classification using Naive Bayes, Random Forest, and K-NN achieved 99 ± 1% accuracy with 100% sensitivity and specificity. While demonstrating the power of multimodal feature fusion, gaps remain in scaling to larger, more diverse populations and in exploring end-to-end deep learning models without feature engineering. Compared to our model, this study highlights the performance gains achievable when combining clinical and morphological information. Insights suggest that integrating heterogeneous data sources with dimensionality reduction can substantially improve detection accuracy even with limited datasets, providing a practical framework for automated thalassemia screening.

Collectively, these studies demonstrate the promise of machine learning and deep learning for automated thalassemia detection. While prior work often focuses either on blood smear image analysis or RBC-based classification, few approaches integrate both modalities. Combining morphological and haematological features enhances diagnostic reliability and supports scalable, patient-centred screening systems, particularly in resource-limited settings such as Sri Lanka.

The PCA results revealed that MCV and MCH contributed most strongly to the first two principal components, indicating that these parameters alone capture the majority of the variance distinguishing β-thalassemia trait from normal individuals. This observation aligns with clinical knowledge, as microcytosis (low MCV) and hypochromia (low MCH) are hallmark features of thalassemia carriers.

To further illustrate this, all nine algorithms were initially trained using only MCV and MCH as input features. The resulting classification performance is shown in Fig. 4, which displays the separation of positive and negative cases by each algorithm in the test set. The graph demonstrates that even with just these two parameters, the models are able to distinguish BTT-positive from normal cases with reasonable accuracy, highlighting the predictive power of MCV and MCH in preliminary screening. This also informed subsequent model development, where additional parameters were incorporated to further improve classification metrics such as sensitivity, specificity, and overall accuracy.

Table 2 shows the performance metrics of various machine learning models trained on RBC parameters for differentiating BTT-positive (BTT) and normal cases. Overall, the models demonstrated high performance, with Classification Accuracies exceeding 92% across most configurations and AUC values consistently above 95%, reflecting strong discrimination between BTT-positive and normal cases. Sensitivity and specificity were well-balanced, with specificity ranging from approximately 93–94% and sensitivity from 88–94%, which is critical for a screening approach where minimising false negatives is essential. True positive counts (BTT → BTT) ranged from 48 to 50, and true negative counts (Normal → Normal) ranged from 84 to 85, indicating a low rate of misclassification. Models trained on slightly different subsets of RBC parameters, including age, sex, or history of iron supplementation, showed minor variations in metrics, suggesting robustness to feature selection. Among the algorithms, the MLP (100) model achieved the highest AUC (~ 96.88%) with well-balanced sensitivity and specificity, whereas Naive Bayes models exhibited the highest sensitivity (~ 92.45%) while maintaining competitive overall accuracy. Logistic Regression and SVM models also performed strongly and offer the additional advantage of interpretability.

Table 3 summarises the performance of each CNN model on the independent test set, with all metrics reported along with their 95% confidence intervals.

VGG16 achieved the highest classification accuracy (72.1%) and the highest specificity (64.3%) among the models, while maintaining a strong recall (86.67%), indicating a balanced ability to identify both positive and negative samples. Models such as MobileNetV2, ConvNeXtTiny, and InceptionV3 reached perfect recall (100%), but their specificity was extremely low, leading to a high number of false positives and a risk of unnecessary follow-up testing. ResNet50 and DenseNet121 showed moderate performance across all metrics, offering more balanced but overall lower accuracy compared to VGG16.

Overall, the results demonstrate a clear trade-off between recall and specificity among the CNN models. Since the primary objective of this step is screening, avoiding missed positive cases is critical. Therefore, VGG16 was selected for blood smear analysis because it provides high sensitivity (recall 86.67%), ensuring that most potential β-thalassemia carriers are flagged for confirmation testing. Although this increases false positives, the pipeline tolerates this, as all flagged cases will undergo HPLC confirmation. This approach maximises patient safety and supports effective population screening within the dual-modal workflow.

Overall, the results demonstrate a clear trade-off between recall and specificity among the CNN models. Models such as ConvNeXtTiny and MobileNetV2 achieved very high recall, correctly identifying nearly all positive cases, but their extremely low specificity led to many false positives and unnecessary follow-up testing. In contrast, VGG16 provided the best balance, with the highest classification accuracy and specificity while maintaining high sensitivity (86.67%), reducing the risk of missed positive cases.

Since the primary objective of this second step is screening, avoiding false negatives is critical. Therefore, VGG16 was selected as the preferred model for blood smear analysis, as its high sensitivity ensures that most potential β-thalassemia carriers are flagged for confirmation testing. Although this increases false positives, the pipeline is designed to tolerate this, as all flagged cases will undergo HPLC confirmation. This approach maximises patient safety, aligns with clinical screening practice, and supports the overall efficiency of the dual-modal pipeline.

To further interpret the performance of the proposed thalassemia screening pipeline, an error analysis was conducted to examine the characteristics and potential causes of model misclassifications. This analysis aimed to identify patterns among false positives and false negatives across both the RBC-based and image-based models, evaluate the complementary effects of the two-step approach, and reveal factors contributing to residual errors. Understanding these limitations provides valuable insight into model behaviour and guides future improvements in data preprocessing, feature selection, and threshold optimisation.

To further investigate the misclassifications of the Red Blood Cell Indices-based MLP model, scatter plots were generated for multiple pairs of haematological parameters. Each plot highlights true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), allowing visual assessment of the model’s decision boundaries and error patterns. The plots also indicate patient sex and adjust point transparency to emphasise misclassified cases. This multi-pair approach provides an intuitive way to explore the distribution of misclassified samples across different parameter combinations, identify regions where the model struggles, and guide potential improvements in feature selection or thresholding. Figure 5 shows Multi-pair scatter plots showing MLP model predictions on the RBC dataset. Visualisations illustrate the distribution of errors, reveal parameter regions contributing to misclassifications, and provide insights for refining the screening pipeline.

Analysis of the MLP misclassifications revealed that false negatives (FNs) cluster in regions of the MCV vs MCH scatter plot that largely overlap with true negatives (TNs). These samples exhibited relatively higher MCV and MCH values compared to true positives, while their RBC counts were lower than those of typical positive cases but similar to those of negative samples. Biologically, this pattern may reflect borderline or atypical β-thalassemia carriers whose red blood cell indices do not fall within classical microcytic and hypochromic ranges. Such cases are known to occur in carriers with mild phenotypic expression, coexisting iron deficiency, or recent blood transfusions, leading to partial normalisation of haematological parameters. In the present dataset, the observed false negatives did not have recent blood transfusions or iron treatment, suggesting that their borderline haematological profiles may be due to mild genetic variants, heterozygosity with less severe mutation expression, or individual physiological compensation mechanisms that partially normalise RBC indices. Consequently, the model may classify these borderline carriers as negative, illustrating the inherent limitation of using Red Blood Cell indices alone for precise detection in all patient subpopulations.

Conversely, false positives (FPs) were primarily located within regions characteristic of true positives, exhibiting lower MCV and MCH values alongside higher RBC counts. This pattern suggests that while these individuals do not carry the β-thalassemia trait, their haematological parameters resemble those of carriers, possibly due to benign microcytosis, variations in haemoglobin concentration, or other haematological conditions. Such overlap underscores the trade-off between sensitivity and specificity in screening approaches and highlights the necessity of confirmatory testing to accurately distinguish true carriers from individuals with overlapping but non-pathological profiles.

To further interpret the performance of the image-based screening step, misclassifications from the VGG-16 model were examined. False negatives (FNs) and false positives (FPs) were analysed in relation to morphological characteristics observed in the blood smear images, providing insight into patterns that contribute to residual errors. This analysis aimed to identify visual features associated with borderline cases, evaluate limitations in feature extraction by the CNN, and highlight scenarios where morphological ambiguity affects predictive accuracy.

Grad-CAM visualisations were generated for representative true negative (TN), true positive (TP), false positive (FP), and false negative (FN) samples to investigate the image-based model’s attention patterns (Fig. 6). For the TP sample (ID: 63542, Positive Probability: 99.29%), attention was strongly focused on the central regions of RBCs, particularly highlighting morphological features consistent with β-thalassemia. Similarly, the TN sample (ID: 63769, Positive Probability: 1.71%) showed minimal attention on RBCs, with the model correctly ignoring normal cell morphology. In contrast, the FP sample (ID: 63627, Positive Probability: 91.32%) exhibited concentrated attention on RBC clusters, likely misinterpreting normal variations or overlapping cells as thalassemia-like features, resulting in high predicted probability despite the sample being negative. The FN sample (ID: 4303, Positive Probability: 11.72%) displayed diffuse and less distinct attention across RBCs, indicating that subtle or borderline morphological abnormalities were insufficient for confident detection. Collectively, these patterns suggest that misclassifications arise from either over-interpretation of normal variations (FPs) or under-recognition of mild, atypical morphological features (FNs), highlighting inherent limitations of relying solely on image morphology for screening in borderline cases.

To further evaluate the robustness and optimal configuration of the dual-modal screening system, two alternative pipeline architectures were tested by altering the order of Red Blood Cell Indices and image-based analysis. This served as an ablation-style assessment to determine how sequencing the analytical stages influences diagnostic accuracy, sensitivity, and specificity.

In this configuration, the Multi-Layer Perceptron (MLP) model based on full blood count parameters was applied first. Samples classified as negative were subsequently analysed using the VGG-16 CNN model. This approach aimed to minimise the number of smear analyses required while ensuring that no β-thalassemia carriers were missed. The combined pipeline achieved 100% sensitivity and 60.7% specificity, indicating that all positive cases were successfully detected but at the cost of increased false positives. Importantly, this configuration reduced the overall need for smear preparation and image processing by 37.7%, representing a practical advantage in large-scale screening workflows.

The second configuration reversed the order, applying VGG-16–based image analysis first and subsequently re-evaluating negative cases with the MLP model. This sequence also achieved 100% sensitivity, confirming that no positive samples were missed. Although specificity increased slightly to 64.3%, all samples required blood smear preparation and image processing. This introduced significant manual effort, time consumption, and computational workload. Consequently, despite maintaining perfect sensitivity, the operational overhead makes this architecture less practical for large-scale or resource-limited screening programs.

Both pipelines maintained perfect sensitivity, confirming the value of combining haematological and morphological features for comprehensive β-thalassemia screening. However, starting with Red Blood Cell indices analysis followed by image evaluation of negatives provided greater operational efficiency by reducing smear preparation and processing requirements. This indicates that the Red Blood Cell Indices → Image pipeline is more suitable for scalable, resource-conscious screening in clinical and public health settings.

The proposed dual-modal screening pipeline combines Red Blood Cell indices analysis and blood smear image classification to maximise sensitivity while reducing unnecessary laboratory workload. In this study, all positive cases (BTT carriers) were successfully identified, with zero missed cases, demonstrating that the pipeline is highly reliable for screening purposes. By first using the Red Blood Cell indices analysis model to classify cases, only the negative cases from the first step were forwarded to blood smear analysis. Consequently, the number of blood smear preparations was reduced from 152 total samples to 97, effectively saving 37% of laboratory time, reagents, and personnel effort.

This two-step approach ensures that almost all positive cases are flagged for confirmation while minimising unnecessary laboratory procedures, making the pipeline both practical and cost-effective, particularly in settings with limited access to haematologists or specialised laboratory infrastructure. Additionally, by using a high-sensitivity VGG16 model for the second step, the pipeline preserves patient safety without compromising screening efficiency.

This study demonstrates the feasibility and potential utility of a dual-modal machine learning pipeline for β-thalassemia screening in Sri Lanka, combining Red Blood Cell (RBC) indices and peripheral blood smear images. By integrating these two complementary modalities, the system leverages both quantitative haematological features and morphological patterns, providing a sensitive, cost-effective, and scalable approach suitable for deployment in low-resource settings.

Analysis of the RBC dataset revealed clear microcytic and hypochromic patterns among β-thalassemia trait (BTT) carriers, characterised by lower MCV (59.89 ± 4.20 fL) and MCH (19.05 ± 1.37 pg) and elevated RBC counts (5.72 ± 0.65 ×10¹²/L), compared to non-carriers. RDW was also higher among carriers, reflecting increased variability in red cell size. These differences provided a strong discriminatory signal for the initial RBC-based screening step. The MLP machine learning classifier, optimised using these parameters, achieved high specificity, allowing positive cases to be confidently referred for confirmatory HPLC testing while minimising unnecessary smear preparation for individuals screened as negative.

The blood smear image analysis component further enhanced screening sensitivity, effectively detecting subtle morphological changes, including target cells, pencil cells, and other characteristic features of thalassemia. Using the VGG16 CNN architecture with transfer learning, the system reliably identified borderline and atypical carriers that could be missed by RBC indices alone. The integration of structured and image-based data in a sequential pipeline ensured that no true positive cases were overlooked, while maintaining operational efficiency by limiting smear analysis to individuals not flagged in the first step.

This dual-modal approach addresses several challenges faced by national screening programmes in low- and middle-income countries. Firstly, it reduces reliance on specialised laboratory equipment, as smartphone-assisted microscopy provided sufficiently high-quality images. Secondly, the sequential workflow optimises resource utilisation by minimising the number of smears requiring manual examination, thereby saving time, reagents, and labour. Thirdly, given the limited number of haematologists in Sri Lankan hospitals, conducting mass screening is extremely tiring and exhausting for available staff; the model’s portability and compatibility with widely available smartphones enable decentralised screening in rural and primary care settings. This aligns with ongoing efforts to expand thalassaemia prevention programmes in Sri Lanka, including premarital screening initiatives.

By combining RBC indices and smear morphology, the system also improves diagnostic confidence, reducing false positives and negatives compared to single-modality approaches. This is particularly relevant in LMICs, where conventional confirmatory tests like HPLC or genetic analysis may be limited by cost, infrastructure, or accessibility. The proposed pipeline could serve as a triage tool, prioritising individuals for confirmatory testing while reducing the burden on central laboratories.

Despite its promising performance, the study has several limitations. The dataset was derived from a single tertiary referral centre in Kurunegala, limiting geographical and ethnic diversity. While age and sex distributions were balanced, other demographic factors, such as socioeconomic background, ethnicity, and regional prevalence variations, were not captured. Consequently, external validity remains to be established, and multi-centre studies are necessary to confirm generalisability across diverse populations.

The sample size is modest, comprising 152 participants, which may affect the robustness of the findings and the ability to capture less common presentations of β-thalassaemia.

Although smartphone-assisted image acquisition reflects real-world conditions, image quality may vary depending on operator experience, lighting, and device specifications. Data augmentation and CNN training mitigated some of these effects; however, further field validation is necessary to assess robustness under variable community-level conditions.

The current pipeline remains a research prototype and has not yet undergone regulatory review, integration with laboratory information systems, or prospective clinical evaluation.

Future work should focus on expanding the dataset to include multi-centre, nationwide samples and additional demographic variables, enhancing model generalisability and robustness. Prospective studies are also required to evaluate real-world usability, workflow integration, and the impact on screening coverage, particularly in the context of mass thalassaemia prevention programs.

Engagement with policymakers, healthcare administrators, and frontline health workers will be essential to understand operational constraints, gather feedback, and secure support for implementation. The development of a clinician-friendly mobile application with streamlined reporting could facilitate adoption, improve operational efficiency, and enable decentralised screening in rural and primary care settings.

The dual-modal framework could be adapted for other haemoglobinopathies, supporting broader public health screening initiatives and contributing to scalable, sustainable, and equitable approaches to genetic disorder prevention.

While the dual-modal machine learning pipeline demonstrates high sensitivity and practical feasibility, several considerations must be addressed prior to deployment.

Proactively addressing these considerations will enable safe, effective integration of the system into national screening pathways while maintaining patient trust and clinical reliability.

This study demonstrates that a dual-modal machine learning pipeline combining RBC indices and peripheral smear morphology can provide a sensitive, cost-effective, and practical approach for β-thalassemia screening. By enabling early detection and prioritisation of at-risk individuals, the system has potential to strengthen national prevention programs, reduce healthcare burden, and support decentralised screening in low-resource settings. It’s design, optimised for smartphones and minimal infrastructure, illustrates a scalable model for implementing AI-driven diagnostic support in LMICs.