applied ML · complete

Clinical risk prediction

A pressure-ulcer risk model built to be honest about its own numbers: ensemble methods, measured on the metric that survives class imbalance, and audited before any accuracy figure was allowed to stand.

Problem

Clinical risk models are easy to overstate. On imbalanced data — most patients don't develop the condition — a model can score high on raw accuracy while being close to useless at the thing that matters: flagging the patients actually at risk.

Approach

Ensemble methods, audited end to end. The model was scored on AUC-ROC, not accuracy alone, and every stage — from split to metric — was checked for the usual failure modes before a single number was reported.

Findings

Ensemble accuracy 87.62%, AUC-ROC 88.67%. Reported together on purpose: the AUC-ROC is the figure that holds up under class imbalance, and it's the one the audit trusts.

The problem.

Pressure-ulcer risk prediction is a genuinely useful thing to get right — the condition is common, costly, and largely preventable when the at-risk patients are flagged early. It's also a textbook trap for machine learning: the positive class is rare, so a model that predicts "no risk" for everyone can post a high accuracy and be clinically worthless.

That gap — between a number that looks good and a model that helps — is the whole problem. The work was as much about not fooling yourself as it was about the model.

Approach.

The model uses ensemble methods — combining several weaker learners so no single one's blind spot decides the outcome. But the method was the smaller half of the work. The larger half was the audit: checking the split for leakage, checking whether the headline metric survived the class imbalance, and rebuilding the pipeline wherever it didn't.

A model that only looks good on the metric that flatters it hasn't been evaluated — it's been marketed.
— audit notes

Findings.

Ensemble accuracy

87.62%

raw accuracy across the test set

AUC-ROC

88.67%

the figure that holds under class imbalance

Method

ensemble

several learners, no single blind spot

The two numbers are reported side by side deliberately. Accuracy alone would be the misleading one on data this imbalanced; the AUC-ROC is what says the model actually separates the at-risk patients from the rest — and it's the figure the audit signs off on.

The full methodology — dataset handling, the specific audit checks, and the write-up — is available on request.