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

Jacob M. Desman; Zhang Wei Hong; Moein Sabounchi; Ashwin S. Sawant; Jaskirat Gill; Ana C. Costa; Gagan Kumar; Rajeev Sharma; Arpeta Gupta; Paul McCarthy; Veena Nandwani; Doug 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 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

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

Jacob M. Desman
, Zhang Wei Hong
, Moein Sabounchi
, Ashwin S. Sawant
, Jaskirat Gill
, Ana C. Costa
, Gagan Kumar
, Rajeev Sharma
, Arpeta Gupta
, Paul McCarthy
, Veena Nandwani
, Doug 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 Reich, Patricia Kovatch, Roopa Kohli-Seth, Monica Kraft, Pulkit Agrawal, John A. Kellum, Girish N. Nadkarni, Ankit Sakhuja

Research output: Contribution to journal › Article › peer-review

1 Scopus citations

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.

Original language	English (US)
Article number	313
Journal	npj Digital Medicine
Volume	8
Issue number	1
DOIs	https://doi.org/10.1038/s41746-025-01709-9
State	Published - Dec 2025
Externally published	Yes

ASJC Scopus subject areas

Medicine (miscellaneous)
Health Informatics
Computer Science Applications
Health Information Management

Access to Document

10.1038/s41746-025-01709-9

Cite this

Desman, J. M., Hong, Z. W., Sabounchi, M., Sawant, A. S., Gill, J., Costa, A. C., Kumar, G., Sharma, R., Gupta, A., McCarthy, P., Nandwani, V., Powell, D., Carideo, A., Goodwin, D., Ahmed, S., Gidwani, U., Levin, M. A., Varghese, R., Filsoufi, F., ... Sakhuja, A. (2025). A distributional reinforcement learning model for optimal glucose control after cardiac surgery. npj Digital Medicine, 8(1), Article 313. https://doi.org/10.1038/s41746-025-01709-9

Desman, JM, Hong, ZW, Sabounchi, M, Sawant, AS, Gill, J, Costa, AC, Kumar, G, Sharma, R, Gupta, A, McCarthy, P, Nandwani, V, Powell, D, Carideo, A, Goodwin, D, Ahmed, S, Gidwani, U, Levin, MA, Varghese, R, Filsoufi, F, Freeman, R, Shetreat-Klein, A, Charney, AW, Hofer, I, Chan, L, Reich, D, Kovatch, P, Kohli-Seth, R, Kraft, M, Agrawal, P, Kellum, JA, Nadkarni, GN & Sakhuja, A 2025, 'A distributional reinforcement learning model for optimal glucose control after cardiac surgery', npj Digital Medicine, vol. 8, no. 1, 313. https://doi.org/10.1038/s41746-025-01709-9

@article{cc5cbee1cf804e0da2f40b3290b4098c,

title = "A distributional reinforcement learning model for optimal glucose control after cardiac surgery",

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{\textquoteright} 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{\textquoteright}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.",

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

note = "Publisher Copyright: {\textcopyright} The Author(s) 2025.",

year = "2025",

month = dec,

doi = "10.1038/s41746-025-01709-9",

language = "English (US)",

volume = "8",

journal = "npj Digital Medicine",

issn = "2398-6352",

publisher = "Nature Publishing Group",

number = "1",

}

TY - JOUR

T1 - A distributional reinforcement learning model for optimal glucose control after cardiac surgery

AU - Desman, Jacob M.

AU - Hong, Zhang Wei

AU - Sabounchi, Moein

AU - Sawant, Ashwin S.

AU - Gill, Jaskirat

AU - Costa, Ana C.

AU - Kumar, Gagan

AU - Sharma, Rajeev

AU - Gupta, Arpeta

AU - McCarthy, Paul

AU - Nandwani, Veena

AU - Powell, Doug

AU - Carideo, Alexandra

AU - Goodwin, Donnie

AU - Ahmed, Sanam

AU - Gidwani, Umesh

AU - Levin, Matthew A.

AU - Varghese, Robin

AU - Filsoufi, Farzan

AU - Freeman, Robert

AU - Shetreat-Klein, Avniel

AU - Charney, Alexander W.

AU - Hofer, Ira

AU - Chan, Lili

AU - Reich, David

AU - Kovatch, Patricia

AU - Kohli-Seth, Roopa

AU - Kraft, Monica

AU - Agrawal, Pulkit

AU - Kellum, John A.

AU - Nadkarni, Girish N.

AU - Sakhuja, Ankit

PY - 2025/12

Y1 - 2025/12

N2 - 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.

AB - 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.

UR - https://www.scopus.com/pages/publications/105006921885

UR - https://www.scopus.com/pages/publications/105006921885#tab=citedBy

U2 - 10.1038/s41746-025-01709-9

DO - 10.1038/s41746-025-01709-9

M3 - Article

AN - SCOPUS:105006921885

SN - 2398-6352

VL - 8

JO - npj Digital Medicine

JF - npj Digital Medicine

IS - 1

M1 - 313

ER -

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

Abstract

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this