Problem
Sepsis requires timely recognition, yet clinical deterioration can be difficult to identify early from high-dimensional patient data. The project investigated whether machine learning could identify patterns associated with sepsis before clinical onset.
Approach
- 1.Explored anonymised hospital data from Mumbai and derived Sequential Organ Failure Assessment scores to represent patient deterioration.
- 2.Performed feature engineering, feature selection, and principal component analysis to identify salient clinical patterns and reduce dimensionality.
- 3.Developed a random-forest model to predict sepsis onset up to six hours before clinical onset and considered its role as decision support rather than a replacement for clinician judgement.
Learnings
- Predictive performance is insufficient without careful communication of uncertainty and fit with clinical workflow.
- Feature selection is both a modelling task and a clinical interpretability task.