Predicting knee replacement outcomes.
Up to a quarter of knee replacement patients still report pain a year after surgery. Can we see it coming before the operation?
What it is
Group project at the Eindhoven AI Systems Institute, built on NHS Digital PROMs data (patient-reported outcome measures for hip and knee replacements). The clinical question: which patients will have a poor outcome from total joint replacement, predicted using only information available before surgery?
The problem is heavily imbalanced, so the work centred on the precision-recall trade-off and on decision thresholds a clinician could defend in practice, not just a leaderboard metric.
The work
- Built and compared multiple ML classification models on pre-surgery (T0) patient data
- Defined the outcome variable from PROMs scores and engineered the feature space
- Used precision-recall curves as the primary metric for the imbalanced classes
- Explored decision thresholds and their clinical pros and cons
- Benchmarked providers into performance and quality groups
Concepts
What it taught me
When the unit of analysis is a patient, the decision threshold is not a hyperparameter. It is the whole point.