A pplication of Machine Learning Approaches to D evelop P redictive M odels for Diabetes and Hypertension among Bangladesh Adults

<br>AbstractIntroduction:With rapidurbanization, lifestyle changes, and an aging population, non-communicablediseases (NCDs),includinghypertension and diabetes,pose significant public health challenges inBangladesh andmanyotherlow-and middle-incomecountries. This studyusedmachine learning (ML)approachesto develop predictive models forhypertension and diabetesamong adultsin this country.Methods:BangladeshDemographic andHealthSurvey2022datawere analyzed. This isa nationallyrepresentative cross-sectional survey.Participantre classified as hypertensivewhen theirsystolic bloodpressurewas≥140mmHg, diastolicblood pressure was≥90mmHg, orif they usedantihypertensivemedication.They were classified asdiabetic if theirfasting plasma glucosewas≥7.0mmol/L orthey usedglucose-lowering drugs. Potential predictors included age, gender, education, wealth quintile,overweight/obesity,rural-urbanresidence, and divisionof residence.Descriptiveanalysis was conducted,andsix ML modelswere applied: artificialneuralnetwork (ANN),randomforest,adaptive boosting(AdaBoost),gradientboosting, XGBoost, andsupportvectormachine (SVM). Models’performancewasevaluated via accuracy,area under the curve (AUC), sensitivity, specificity, and F1-score. Featureimportance was assessed to rankrisk factors.Results:The study included13,847 adults, 55% of whom were females.Sensitivity was high across models(up to 0.96 for diabetes and 0.90 for hypertension). However, the overall specificity was low, particularlyfor diabetes (as low as 0.13 in XGBoost).Diabetes and hypertension had prevalence of 16.3% and 20.5%,respectively.The prevalence of both conditionsincreasedwith age, and the highest prevalencewas24.4%for diabetes and 43.3% for hypertensionamongindividuals aged 65 and older. Wealthier and urban residentsexperienced higher rates (diabetes: 24.9% among the richest compared to 9.9% among the poorest;hypertension: 23.3% in urban versus 19.2% in rural areas). Additionally, overweight/obesity was a strongpredictor for both conditions.For diabetes, AdaBoosthadthe highest AUC (0.699) and SVMhadthehighest accuracy (0.836); for hypertension, AdaBoosthad the greatestAUC (0.775) and accuracy (0.799).Hypertension topped diabetes predictors, while overweight/obesitywas the top predictorfor hypertension,followed by age and diabetes. Wealth and gender were moderately influential, with education andgeographic factors less so. Low specificity across models indicated challenges in identifying non-cases.Conclusion:This ML-driven analysisidentifiedthe bidirectional relationship ofhypertensionanddiabetesalong with several other predictors, includingoverweight/obesity,older age, and richer household wealthquintiles.Ourfindings underscore the need for integrated screening and lifestyle interventions targetinghigh-risk groups to mitigatefutureNCD burden.