Interpretation of medical machine learning models should be based on clinical impact

Samenvatting

Assessing the clinical performance of a machine learning (ML) model is crucial: for example, is our model successful at predicting which patients should undergo a prostate biopsy or not? A naïve statistical metric in a clinical context is accuracy, which assumes that every error we make is equally impactful. This is obviously not the case: an unnecessary biopsy is much less impactful than missing a very aggressive type of cancer leading to the death of a patient. Accuracy thus ignores the medical benefit or cost of a particular clinical decision we make based on the ML model. In our demonstration we want to explore a clinically oriented way to interpret statistical models, Decision Curve Analysis, for a concrete case involving which patients should receive a prostate biopsy or not.