Knowing When to Stop: Healthcare ML Modelling

A healthcare policy study needed to predict hospital length of stay using a 32-country public data panel.

I tested 540 model variations after an initial baseline was already close to the available accuracy ceiling. I investigated what drove the small gains and assessed whether further tuning could produce a meaningful decision improvement.

540 model versions tested

Best improvement was +0.42% over the untuned baseline

Identified imaging capacity per person as the only factor with material explanatory weight

I recommended stopping further model tuning. The remaining lift was too small to justify additional work, and most model accuracy reflected national baselines rather than transferable insight.

Country-level public data can inform policy questions, not decisions about individual hospitals.