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

Area of Science:

• Oncology
• Medical Informatics
• Health Services Research

Background:

• Machine learning (ML) algorithms show potential in improving prognostic accuracy for serious illnesses like cancer.
• Identifying patients who could benefit from earlier palliative care (PC) or advance care planning (ACP) is crucial.
• The impact of ML presentation strategies on clinician decision-making requires further investigation.

Purpose of the Study:

• To evaluate how different presentation strategies of a hypothetical ML algorithm affect clinician prognostic accuracy.
• To assess the impact of ML predictions on clinicians' decisions regarding palliative care and advance care planning.
• To determine the optimal presentation strategy for ML prognostic estimates.

Main Methods:

• A randomized clinical vignette survey study was conducted among medical oncologists treating metastatic non-small-cell lung cancer (mNSCLC).
• Clinicians reviewed patient vignettes with varying prognostic risks and estimated life expectancy, recommending PC and ACP.
• Clinicians were then shown the same vignettes with hypothetical ML survival estimates, randomized by absolute and/or reference-dependent presentations.

Main Results:

• ML presentation significantly improved prognostic accuracy by 20.9% (P < 0.001).
• Absolute risk presentation strategies, with or without reference dependence, yielded greater accuracy gains.
• ML presentation did not significantly alter the rates of recommending ACP (1.3% change) or PC referral (0.7% change).

Conclusions:

• ML-based prognostic assessments can enhance clinician prognostic accuracy.
• Current ML presentation strategies do not appear to change clinical decision-making regarding PC or ACP referrals.
• Future ML algorithms should prioritize explainability and absolute prognoses for potentially greater impact on clinical decisions.