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

BackgroundMachine learning (ML) algorithms may improve the prognosis for serious illnesses such as cancer, identifying patients who may benefit from earlier palliative care (PC) or advance care planning (ACP). We evaluated the impact of various presentation strategies of a hypothetical ML algorithm on clinician prognostic accuracy and decision making.MethodsThis was a randomized clinical vignette survey study among medical oncologists who treat metastatic non-small-cell lung cancer (mNSCLC). Between March and June 2023, clinicians were shown 3 vignettes of patients presenting with mNSCLC. The vignettes varied by prognostic risk, as defined from the Lung Cancer Prognostic Index (LCPI). Clinicians estimated life expectancy in months and made recommendations about PC and ACP. Clinicians were then shown the same vignette with a hypothetical survival estimate from a black-box ML algorithm; clinicians were randomized to receive the ML prediction using absolute and/or reference-dependent prognostic estimates. The primary outcome was prognostic accuracy relative to the LCPI.ResultsAmong 51 clinicians with complete responses, the median years in practice was 7 (interquartile range 3.5-19), 14 (27.5%) were female, 23 (45.1%) practiced in a community oncology setting, and baseline accuracy was 54.9% (95% confidence interval [CI] 47.0-62.8) across all vignettes. ML presentation improved accuracy (mean change relative to baseline 20.9%, 95% CI 13.9-27.9,  < 0.001). ML outputs using an absolute presentation strategy alone (mean change 27.4%, 95% 16.8-38.1,  < 0.001) or with reference dependence (mean change 33.4%, 95% 23.9-42.8,  < 0.001) improved accuracy, but reference dependence alone did not (mean change 2.0% [95% CI -11.1 to 15.0],  = 0.77). ML presentation did not change the rates of recommending ACP nor PC referral (mean change 1.3% and 0.7%, respectively).LimitationsThe singular use case of prognosis in mNSCLC, low initial response rate.ConclusionsML-based assessments may improve prognostic accuracy but not result in changed decision making.ImplicationsML prognostic algorithms prioritizing explainability and absolute prognoses may have greater impact on clinician decision making.Trial Registration: CT.gov: NCT06463977HighlightsWhile machine learning (ML) algorithms may accurately predict mortality, the impact of prognostic ML on clinicians' prognostic accuracy and decision making and optimal presentation strategies for ML outputs are unclear.In this multicenter randomized survey study among vignettes of patients with advanced cancer, prognostic accuracy improved by 20.9% when clinicians reviewed vignettes with a hypothetical ML mortality risk prediction, with absolute risk presentation strategies resulting in greater accuracy gains than reference-dependent presentations alone.However, ML presentation did not change the rates of recommending advance care planning or palliative care referral (1.3% and 0.7%, respectively).ML-based prognostic assessments without explanations improve prognostic accuracy but do not change decisions around palliative care referral or advance care planning.