Table 2.
Fitted Q-evaluation of VentPilot and ablated policies in validation cohorts
| Reward type | AI | Without intensivist feedback | Without high-resolution train data | Without intensivist feedback and high-resolution train data |
|---|---|---|---|---|
| SNUH internal validation cohort | ||||
| Overall | 1.6 (0.8–2.5) | 1.2 (0.5–1.8) | 1.3 (0.5–1.7) | 0.9 (0.2–1.4) |
| VFDs, daysa | 1.5 (0.8–2.9) | 1.0 (0.2–2.1) | 1.1 (0.4–2.1) | 0.8 (0.1–1.8) |
| Preference | 1.5 (0.1–2.8) | -0.2 (− 0.7–0.8) | 1.2 (− 0.4–2.7) | -0.5 (− 1.1–0.7) |
| Mayo Clinic external validation cohort | ||||
| Overall | 1.3 (0.6–2.0) | 0.7 (0.0–1.5) | 0.8 (0.1–1.5) | 0.2 (− 0.7–1.0) |
| VFDs, daysa | 1.2 (0.3–2.2) | 1.0 (0.2–1.9) | 0.6 (− 0.2–1.4) | 0.1 (− 0.7–1.0) |
| Preference | 0.8 (0.5–1.3) | 0.4 (0.0–0.9) | 0.5 (0.1–0.8) | 0.2 (− 0.2–0.5) |
AI artificial intelligence, SNUH Seoul National University Hospital, VFDs ventilator-free days
Values are differences in estimated return of the AI policy minus the return of the clinician policy; positive values indicate higher estimated return than the clinician policy. Overall reward is the weighted sum of a VFDs-based reward and a preference reward generated by a learned preference model trained on intensivist action ratings. Fitted Q-evaluation was performed separately for the overall reward and for each reward component using the corresponding reward definition. “without intensivist feedback” denotes training with the preference reward weight set to 0 (i.e., no preference-based supervision signal). “without high-resolution train data” denotes training without the prospectively collected high-resolution time-series dataset, using only the coarser electronic medical record-derived dataset available in routine documentation
a For the VFDs component, reported differences can be interpreted as estimated differences in VFDs (in days), because the VFD-based return differs from VFDs only by a constant shift