AI-Based Disease Prediction Model

The problem

Preventive care works when it reaches the right patients early, but identifying who is genuinely at elevated risk means synthesising history, labs and lifestyle factors across a whole panel — which is not something a clinician can do at scale during appointments.

How it works

Gradient-boosted models built with XGBoost and scikit-learn assess early-stage disease risk from patient history, lab results and lifestyle data. PostgreSQL holds the clinical record the features are derived from.

Gradient boosting suited the data: tabular clinical records with meaningful missingness, where a model that handles absent values sensibly matters more than one that needs every field populated. It also keeps feature contributions inspectable, which is what allows a clinician to see why a patient surfaced.

What shaped it

Output is a prioritisation aid, not a diagnosis. The model highlights patients who may warrant preventive attention and leaves the clinical judgement where it belongs — a distinction that shaped how results are surfaced.

Class imbalance is inherent here: most patients do not develop the condition, and a model optimised naively on accuracy learns to say no to everyone. Evaluation focused on catching genuine cases rather than on headline accuracy.

Outcome

Clinicians could prioritise preventive care toward the patients most likely to benefit.