The Impact of Machine Learning Mortality Risk Prediction on Clinician Prognostic Accuracy and Decision Support: A Randomized Vignette Study

This randomized vignette study evaluated how different presentation strategies of a hypothetical ML mortality prediction algorithm influenced oncologists’ accuracy in estimating prognosis for metastatic non-small-cell lung cancer. Clinician prognostic accuracy increased by ~21 percentage points when ML outputs were provided, particularly in absolute risk formats, but this did not alter rates of recommending advance care planning or palliative care referral.

The study underscores that while ML can significantly improve clinician prognostic accuracy, changes in behavior require more than just better predictions; it suggests a need for behavioral intervention designs alongside AI deployment to impact care decisions. These findings have important implications for designing explainable and effective AI decision support tools in oncology.