- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+1)
Read review
Paper summary
This paper proposes AIRC for offline basal-insulin policy learning. It uses trajectories generated by a reference Loop controller, estimates deviations in logged actions and future glucose risk, and uses their relationship to guide the learned policy. The method is evaluated mainly in SimGlucose, with an additional held-out analysis on real Loop observational data.
Critical strengths
- The problem is important and clinically relevant, especially because offline learning avoids unsafe online exploration.
- The method separates adjustment direction from low-glucose risk, which is intuitive.
- The experiments include several offline RL baselines and component ablations, and AIRC performs well in SimGlucose.
Critical weaknesses
- The residual-based response design is not sufficiently justified. In Sections 4.2–4.3, the method does not directly model how insulin affects future glucose. Instead, it relates an action residual to a future-risk residual and uses this relation to determine the policy-improvement direction. I am not convinced that the future-risk residual is a reliable signal for how insulin should be adjusted, since it can also reflect patient dynamics, meal effects, or prediction error. A direct comparison with action-conditioned future glucose prediction or trajectory prediction would help justify this design.
- The high- and low-glucose risk signals contain several hand-designed choices. The specific forms and parameters in Section 4.2, such as the high-glucose softplus shaping, the constants 30 and 10, the cap of 10, the low-glucose sigmoid centered at 90, and the selected 60/120 and 240/360 minute horizons, are not well justified. Since these choices affect the later value estimates, response coefficients, and actor updates, a sensitivity analysis is needed.
- The original Loop controller is missing from the main comparison. The SimGlucose training trajectories are generated by the reference Loop controller, and AIRC learns a policy that is anchored to those logged actions. However, Table 1 compares AIRC only with offline RL baselines and does not report the closed-loop performance of the original Loop controller. Therefore, the current experiments do not show whether AIRC actually improves over the behavior controller that generated the data. A direct Loop-vs-AIRC comparison under the same evaluation setting seems necessary.
Questions for authors
- Why is the residual relation in Sections 4.2–4.3 a better choice than directly modeling future glucose under different actions?
- How sensitive are the results to the hand-designed risk-shaping parameters and prediction horizons?
- How does AIRC compare directly with the original reference Loop controller in closed-loop SimGlucose evaluation?