The results support the presentation of the tool as an awareness and triage aid rather than a diagnostic test. The study defined dysglycemia as abnormal blood-glucose regulation, encompassing prediabetes and diabetes-range HbA1c categories, and classified respondents as positive when their HbA1c was at least 5.7%. Model development used Wave 5 of the Indonesia Family Life Survey, a longitudinal socioeconomic and health survey conducted nationally by the RAND Corporation with SurveyMeter in Yogyakarta.

IFLS5 covered 16,204 households and 50,148 respondents in 13 of Indonesia’s 27 provinces, with the survey sample representing approximately 83% of the national population. It was unusual among the Indonesian diabetes-prediction studies reviewed by the authors because it combined broad population coverage with measured HbA1c labels. Earlier studies were described as using 120 to 1,000 records from single facilities or narrow regional populations.

Investigators limited the analysis to respondents aged 20 or older with valid HbA1c measurements, then merged biomarker records with demographic, lifestyle and health information. Complete-case filtering left an analytic sample of 5,052 adults. Of these, 1,753, or 34.7%, met the dysglycemia threshold, while 3,299, or 65.3%, did not.

The analytic sample was 53.6% female and 46.4% male; Javanese participants represented 45.4% and Sundanese participants 12.4%. Compared with the full survey, the female proportion differed by 1.9 percentage points, and the ethnic-group ranking was unchanged, although Javanese representation increased from 41.7% to 45.4%. The sample was older than the full survey because HbA1c measurement was more frequent among older adults, an imbalance consistent with the screening focus but a consideration when generalizing to younger people.

Feature selection followed four linked stages rather than relying solely on statistical association. A literature review established non-invasive candidate predictors and used measured HbA1c as the outcome. A deliberately permissive correlation screen reduced 93 columns to 20 candidate features and the target.

Two Indonesian doctors, one specializing in cardiology and the other in internal medicine, independently reviewed that set. They removed variables considered redundant or uninformative, including sex, education and food and mental-health questionnaire blocks, but reinstated hypertension, family history of diabetes and physical-activity measures despite weak linear correlations. The experts also derived body mass index from weight and height and constructed a Brinkman smoking index from daily cigarette consumption and smoking duration.

SHAP then evaluated the global contribution of the 14 candidate predictors, after which nine features were retained and judged clinically coherent by the same doctors. Five algorithms were compared: logistic regression, random forest, XGBoost, CatBoost and TabNet. The investigators optimized hyperparameters with a grid search and five-fold cross-validation, while testing random oversampling, SMOTE, NearMiss and random undersampling to address the 34.7% prevalence of the positive class.

Random undersampling gave the best scores across the algorithms. XGBoost achieved the strongest overall balance, with a test-set ROC-AUC of 0.72 and precision-recall AUC of 0.60. Logistic regression, a simpler transparent baseline, reached a ROC-AUC of 0.69.

The authors interpret this narrow difference as evidence that the available non-invasive predictors contain limited discriminatory information, rather than showing that greater model complexity would automatically improve screening performance. Subgroup analyses on the held-out test set found nearly identical discrimination for a majority grouping of Javanese and Sundanese participants and a combined minority grouping, with ROC-AUC values of 0.71 and 0.72, respectively. Age-stratified ROC-AUC values were similarly close at 0.69 below age 60 and 0.70 at age 60 or older, but sensitivity and specificity varied markedly.

Among older adults, sensitivity was 0.80 and specificity was 0.32; among younger adults, sensitivity was 0.56 and specificity was 0.79. These results describe performance in the observed test data and do not establish equivalent accuracy for every individual ethnic group, because several small groups were combined to permit subgroup calculations. The model was also trained without survey weights, leaving population-calibrated estimation for future work.

SHAP analysis identified age, BMI and hypertension as the most influential variables in model output. Age had the strongest positive Pearson correlation with the outcome, at 0.24, followed by hypertension at 0.17 and BMI at 0.13. The SHAP dependence analysis showed a nonlinear pattern: age contributed little below 40, began adding to modeled risk thereafter and increased sharply after age 45.

Smoking status, high cholesterol and parental history of diabetes also contributed positively to the model’s output, while higher moderate physical-activity frequency contributed toward lower risk. A domain expert reviewed global and individual explanations and judged them clinically consistent. SHAP was selected over LIME after local LIME explanations varied with random seeds and the definition of a local neighborhood proved ambiguous for tabular data.

For the tree model, TreeExplainer required 0.0075 seconds per explanation compared with 0.4706 seconds for KernelExplainer while producing nearly identical values. The resulting application was built in Kotlin with Jetpack Compose for Android 5.0 and later, supported by a Golang back end. Users enter weight and height, pregnancy history involving delivery of a baby weighing at least 4 kilograms, smoking status and duration, blood-pressure and cholesterol conditions, family history of diabetes-related mortality, and weekly physical activity.

The server runs the model and calculates SHAP contributions before sending the results to the OpenAI API, where GPT-4o generates personalized explanations based on the structured outputs of a medical-explainer prompt. Deterministic code computes each feature’s percentage contribution and passes those values directly to the language model, which is instructed to contextualize rather than independently calculate them. The interface presents the estimated risk as a percentage, places it in four categories from low to very high risk, and uses bar or pie charts to distinguish risk-increasing features in red from risk-decreasing features in green.

Bahasa Indonesia is the application’s primary language. Usability testing offered preliminary evidence that the combined display could be understood by non-specialists. Twelve volunteers—eight men and four women aged 20 to 64—viewed a fixed health profile, completed a five-point questionnaire and then joined semi-structured interviews.

Most had no specialized medical knowledge, and educational backgrounds ranged from elementary and junior high school to university study in several fields. The overall mean comprehension score was 4.31 out of 5, with a standard deviation of 0.88, and was reported as significantly exceeding the study’s target of 4.0. Graphics combined with text received the highest item score, at 4.67, followed by language simplicity at 4.58 and text clarity at 4.50.

Bar and pie charts each scored 3.75 when presented alone. Participants with lower educational attainment initially found the charts difficult but could identify the main message through the red and green cues, while the narrative text helped translate the graphics into personal understanding. Because the pilot included only 12 people from one geographic area and no clinical patients, it cannot establish accessibility across Indonesia’s broader population.

The principal limitation is that the model’s discrimination was modest, its calibration was not formally assessed, and the default operating point could miss many positive cases. At a threshold of 0.50, sensitivity was 0.63, meaning roughly one-third of respondents with dysglycemia in the test set were not flagged. Threshold analysis showed that sensitivity could rise to 0.80 at a specificity of 0.54, and to 0.86 at a specificity of 0.38, but the appropriate balance depends on the deployment context and the consequences of missed cases versus unnecessary confirmatory testing.

The study calculated neither calibration curves nor a Brier score, and the semantic fidelity of the generated descriptions was not formally evaluated. These limitations, together with the small pilot and older-skewed sample, mean the deployed framework has not been validated for diagnosis. The authors identify formal calibration, additional indicators such as waist-to-height ratio, broader usability studies, federated learning and on-device execution as future directions rather than established features or outcomes.

Subject of Research: Population-representative explainable-AI screening for dysglycemia among Indonesian adults

Article Title: A population representative XAI framework for transparent dysglycemia screening in Indonesia

Article References: Dwiyanti, L., Tanadi, R., Willson, B., Radityo, H. A. S., Akbar, S., & Nambo, H. (2026). A population representative XAI framework for transparent dysglycemia screening in Indonesia. Discover Artificial Intelligence, 6(1), Article 1181. https://doi.org/10.1007/s44163-026-02243-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02243-7

Keywords: dysglycemia screening, explainable AI, SHAP, XGBoost, HbA1c, Indonesia, mobile health, IFLS5, language models, population health, population, representative