Effect of Integrating Machine Learning Mortality Estimates With Behavioral Nudges to Clinicians on Serious Illness Conversations Among Patients With Cancer

Manz, C. R.

doi:10.1001/jamaoncol.2020.4759

Manz ML nudges (SIC)

Effect of Integrating Machine Learning Mortality Estimates With Behavioral Nudges to Clinicians on Serious Illness Conversations Among Patients With Cancer

Patient / Population Intervention / Exposure Comparison Outcome

In 14,607 patients with cancer seen by 78 oncology clinicians, machine-learning mortality predictions delivered as behavioural nudges increased serious illness conversations compared with usual care with weekly cumulative performance emails.

N

14,607patients

Design

Stepped-wedge cluster RCT, 8 clinic groups over 20 weeks, Pennsylvania, USA

Endpoint

Percentage of patient encounters with a documented serious illness conversation

Relevance

1Practice-defining — the trial the guideline rests on.

ResultSICs in 4.6% vs 1.3% of all encounters (adjusted difference 3.3 percentage points, 95% CI 2.3–4.5; P<0.001); among 4,124 high-risk encounters 15.2% vs 3.6% (adjusted difference 11.6 points, 95% CI 8.2–12.5; P<0.001).

Manz CR, et al. JAMA Oncol. 2020;6(12):e204759. 10.1001/jamaoncol.2020.4759

Discussion & critique

A landmark for pairing a validated ML mortality model with behavioural economics: the combination raised the rate of conversations with high-risk patients from 3.6% to 15.2%. Single academic system, 20 weeks, and a process endpoint (documented conversations) rather than goal-concordant care; the 40-week follow-up (Manz ML nudges (long-term)) later showed less end-of-life systemic therapy but no change in hospice or ICU use.