Smart Imitator: Learning from Imperfect Clinical Decisions.

Summary

Smart Imitator is an offline RL pipeline that separates clinician actions by quality via adversarial cooperative imitation learning and then learns a reward to derive superior policies. In a sepsis dataset with 19,711 trajectories, SI reduced estimated mortality by 19.6% versus the best baseline and aligned with successful clinical decisions while deviating strategically.

Key Findings

• Adversarial cooperative imitation learning with sample selection stratified clinician policies from optimal to nonoptimal.
• Parameterized reward learning enabled RL to derive policies that outperformed state-of-the-art baselines.
• On sepsis trajectories (n=19,711), SI reduced estimated mortality by 19.6% compared with the best baseline.

Clinical Implications

If prospectively validated, SI could inform bedside decision support to personalize sepsis care and reduce mortality; deployment requires careful safety guards, clinician oversight, and calibration to local practice.

Limitations

• Outcomes are estimated in offline RL without prospective clinical trials; off-policy evaluation may be biased.
• Generalizability to diverse institutions and dynamic clinical workflows remains unproven.

Future Directions

Prospective, randomized clinician-in-the-loop trials; safety-constrained RL; external validation across health systems; and fairness/robustness evaluation.