A distributional reinforcement learning model for optimal glucose control after cardiac surgery

Q: What are the key findings of this study?

The GLUCOSE reinforcement learning algorithm significantly reduced the mean absolute error in insulin dosing compared to clinicians (0.9 vs 1.97 units, p<0.001) in patients after cardiac surgery.

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Hyperglycemia after cardiac surgery (n=6797). Intervention: GLUCOSE reinforcement learning algorithm vs. Clinician-administered insulin. Primary outcome: Mean absolute error (MAE) in insulin dosing (internal testing) (p=<0.001).

Q: What does this research mean for the field?

The GLUCOSE reinforcement learning algorithm significantly reduces the mean absolute error in insulin dosing compared to clinicians in patients after cardiac surgery. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Q: What question did this study set out to answer?

To develop and evaluate a distributional offline reinforcement learning algorithm for optimizing postoperative insulin administration after cardiac surgery.

Jacob Desman; Zhang-Wei Hong; Moein Sabounchi; Ashwin Sawant; Jaskirat Gill; Ana Catarina Costa; Gagan Kumar; Rajeev Sharma; Arpeta Gupta; Paul McCarthy; Veena Nandwani; D. Powell; Alexandra Carideo; Donnie Goodwin; Sanam Ahmed; Umesh Gidwani; Matthew A. Levin; Robin Varghese; Farzan Filsoufi; Robert Freeman; Avniel Shetreat‐Klein; Alexander W. Charney; Ira Hofer; Lili Chan; David J. Reich; Patricia Kovatch; Roopa Kohli‐Seth; Monica Kraft; Pulkit Agrawal; John A. Kellum; Girish N. Nadkarni; Ankit Sakhuja

doi:10.1038/s41746-025-01709-9

Key result

GLUCOSE algorithm linked to ~54% lower insulin dosing error vs clinicians after cardiac surgery.

P<0.001
n=6,797

Figures

Why the study?

Does a distributional reinforcement learning algorithm (GLUCOSE) improve insulin dosing accuracy and estimated glucose control compared to clinicians in patients after cardiac surgery?

Key Points

To develop and evaluate a distributional offline reinforcement learning algorithm for optimizing postoperative insulin administration after cardiac surgery.
Trained the GLUCOSE algorithm on 5,228 patients, tested it on 920 patients, and externally validated it on a distinct cohort of 649 patients.
Assessed performance across multi-phase human validation against clinicians evaluating mean absolute error, safety, effectiveness, and acceptability.
GLUCOSE achieved mean estimated rewards of 0.0 [-0.07, 0.06] in internal testing and -0.63 [-0.74, -0.52] in external validation, outperforming clinician rewards of -1.29 [-1.37, -1.20] and -1.02 [-1.16, -0.89], respectively.
The model achieved significantly lower mean absolute error in insulin dosing compared to clinicians in internal testing (0.9 vs. 1.97 units, p < 0.001) and external validation (1.90 vs. 2.24 units, p = 0.003), while matching or exceeding senior clinicians in safety and effectiveness.

Study Design

Type

Observational (n=6,797)

Multicenter

Yes

Structured PICO

Does a distributional reinforcement learning algorithm (GLUCOSE) improve insulin dosing accuracy and estimated glucose control compared to clinicians in patients after cardiac surgery?

P

Population

6,797 ICU patients after cardiac surgery (6,148 in development dataset [MIMIC-IV], 649 in external validation dataset [eICU-CRD]). In the development dataset: mean age 67.8 ± 11.6 years, 71.1% male.

I

Intervention

GLUCOSE (Glucose Level Understanding and Control Optimized for Safety and Efficacy), a distributional offline reinforcement learning algorithm for optimizing hourly insulin dosing.

C

Comparator

Actual clinician-administered insulin doses and retrospective expert clinician recommendations.

O

Outcome

Mean estimated performance return (reward) and Mean Absolute Error (MAE) in insulin dosing compared to expert clinicians.surrogate

Main Result

Absolute Event Rate: 0.9% vs 1.97%

p-value: p=<0.001

An offline reinforcement learning algorithm (GLUCOSE) can recommend safe, effective, and personalized insulin doses for patients in the ICU after cardiac surgery, outperforming standard clinician dosing.

Limitations

Retrospective study requiring prospective validation
Evaluated only for the first day following cardiac surgery
Does not incorporate explicit dietary data
Excluded patients with ambiguous medication administration records
Dependence on accurate insulin administration records, requiring exclusion of patients with ambiguous data

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Does not yet support changing post-cardiac surgery insulin dosing; hypothesis-generating for RL algorithms and requires prospective RCTs.

Abstract

This study introduces Glucose Level Understanding and Control Optimized for Safety and Efficacy (GLUCOSE), a distributional offline reinforcement learning algorithm for optimizing insulin dosing after cardiac surgery. Trained on 5228 patients, tested on 920, and externally validated on 649, GLUCOSE achieved a mean estimated reward of 0.0 [-0.07, 0.06] in internal testing and -0.63 [-0.74, -0.52] in external validation, outperforming clinician returns of -1.29 [-1.37, -1.20] and -1.02 [-1.16, -0.89]. In multi-phase human validation, GLUCOSE first showed a significantly lower mean absolute error (MAE) in insulin dosing, with 0.9 units MAE versus clinicians' 1.97 units (p < 0.001) in internal testing and 1.90 versus 2.24 units (p = 0.003) in external validation. The second and third phases found GLUCOSE's performance as comparable to or exceeding that of senior clinicians in MAE, safety, effectiveness, and acceptability. These findings suggest GLUCOSE as a robust tool for improving postoperative glucose management.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Desman et al. (2025) conducted an observational in Hyperglycemia after cardiac surgery (n=6,797). GLUCOSE reinforcement learning algorithm vs. Clinician-administered insulin was evaluated on Mean absolute error (MAE) in insulin dosing (internal testing) (p=<0.001). The GLUCOSE reinforcement learning algorithm significantly reduced the mean absolute error in insulin dosing compared to clinicians (0.9 vs 1.97 units, p<0.001) in patients after cardiac surgery.

synapsesocial.com/papers/6a1573a8b2e0231f1582824b https://doi.org/10.1038/s41746-025-01709-9

View Full Paper

Ask AI

Bookmark

Share

A distributional reinforcement learning model for optimal glucose control after cardiac surgery

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

A distributional reinforcement learning model for optimal glucose control after cardiac surgery

Figures

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading