Why ML for Risk Prediction

Why traditional scoring falls short for NCD risk in Bangladesh.

Conventional risk scores were built on Western population data. Apply them to Bangladeshi adults and you miss the population-specific patterns that actually predict who gets sick.

Bangladesh is changing fast - rapid urbanisation, shifting diets, an aging demographic. Diabetes prevalence now sits at 16.3% and hypertension at 20.5% among adults nationally. These aren't marginal numbers; they represent a compounding non-communicable disease burden that standard screening thresholds weren't designed to catch early.

Machine learning changes the calculus. Instead of applying a fixed formula, ML models learn from the actual data - in this case, the Bangladesh Demographic and Health Survey 2022, a nationally representative sample of 13,847 adults. The models detect interactions between predictors that a clinician or a logistic regression would likely overlook: the way wealth, geography, age, and body weight combine in this specific population.

02

Key Models and Features

Which algorithms led and which predictors drove them.

AdaBoost outperformed the field on both conditions - highest AUC for diabetes (0.699) and for hypertension (0.775, also highest accuracy at 0.799). Six algorithms were tested in total: Artificial Neural Network, Random Forest, AdaBoost, Gradient Boosting, XGBoost, and Support Vector Machine. SVM achieved the highest raw accuracy for diabetes (0.836), though AUC is the more reliable indicator when class balance is uneven.

Condition	Best AUC model	AUC	Best accuracy model	Accuracy
Diabetes	AdaBoost	0.699	SVM	0.836
Hypertension	AdaBoost	0.775	AdaBoost	0.799

Feature importance is where the clinical signal lives. For hypertension prediction, overweight and obesity was the dominant predictor, followed by older age and the presence of diabetes - confirming the bidirectional relationship between the two conditions. For diabetes prediction, hypertension was the single strongest predictor. Wealth quintile and sex were moderately influential across both models; education and geographic division contributed less.

One honest limitation: specificity was low across all models, especially for diabetes. The models are good at flagging true positives - people who do have the condition - but they struggle to rule out people who don't. In practice, that means these tools are best used for population-level screening prioritisation, not individual diagnosis. Pair a high-sensitivity ML screen with clinical follow-up for the flagged group.

You made it to the end.

Sources

Application of machine learning approaches to develop predictive models for diabetes and hypertension among Bangladeshi Adults. - CC0 (public domain dedication)