DiabetesSense

Clinical prediction models live or die by interpretability. A black-box classifier can hit 95% accuracy and still be useless if a clinician can't see why a particular patient was flagged. Diabetes risk already has good baseline accuracy from logistic regression and tree ensembles, so the real research question wasn't 'can we predict?' but 'can we predict and explain in a way clinicians will actually trust?'

Adoption literature on clinical ML is consistent: when clinicians can't trace a prediction back to features they recognise, they reject the tool — even when the tool is more accurate than their own judgement. The interpretability layer isn't optional polish; it's the load-bearing part.

Built on a public diabetes risk dataset with stratified k-fold splits to handle class imbalance. Trained two complementary tree models — Random Forest for variance reduction across heterogeneous feature interactions, Gradient Boosting for sequential refinement on hard examples — and combined them via soft voting. Hyperparameter search via grid search with CV-internal validation.

Wrapped the ensemble in SHAP TreeExplainer, which exploits the additive structure of tree models to compute exact Shapley values rather than approximations. Per-prediction explanations surface the top contributing features as a horizontal bar chart with positive (risk-increasing) and negative (risk-decreasing) contributions colour-coded.

Deployed the model behind a Flask REST API with a React.js frontend that lets clinicians input patient features and get back a risk score plus the feature attribution chart in real time. The chart is the actual product — the score alone wouldn't have been adopted.

Co-authored a Springer book chapter and presented the work at ICSMAI 2024 in Casablanca, Morocco.

~93% classification accuracy on stratified validation, with sub-second per-prediction SHAP explanations served via the REST API. Feature attributions consistently surfaced clinically meaningful drivers (glucose, BMI, age) as the top contributors, which became the basis for clinician trust during pilot review.

Work was peer-reviewed and presented at ICSMAI 2024 in Casablanca, Morocco, with a Springer book chapter publication tied to the conference.

For clinical deployment beyond a paper, the next blockers are calibration and population shift: 93% on a single curated dataset doesn't mean 93% on a different hospital's intake. The model needs Platt-scaled probabilities and a population-shift detector before it's safe at the bedside.

/* TODO: Hammad — add a reflection on the dataset (Pima or other), what you'd want to redo with hindsight, and how you'd partner with a clinical team to validate properly. */