On-device reinforcement learning-based insulin dosing controller for fully automated insulin delivery

Introduction Automated insulin delivery (AID) systems integrate continuous glucose monitoring (CGM) with algorithm-driven insulin dosing to enable closed-loop glycemic management for people with type 1 diabetes (T1D). Recent advances in wearable glucose sensing and insulin delivery technologies have enabled increasingly compact AID systems; however, deploying advanced AI-based dosing algorithms directly on resource-constrained wearable devices remains technically challenging. Some AID architectures also rely on external devices, such as smartphones or dedicated controllers, introducing additional communication dependencies into the closed-loop operation. Methods This study presents an on-device insulin dosing controller based on the soft actor–critic (SAC) algorithm for resource-constrained AID systems. The proposed approach processes CGM data and makes insulin dosing decisions locally, without relying on external computing devices. To enable embedded deployment, a compressed SAC architecture with a lightweight encoder was developed using quantization-aware training and knowledge distillation, together with an online adaptation mechanism for individualized control. The controller was implemented on a resource-constrained microcontroller and evaluated in a hardware-in-the-loop framework using virtual patients from a T1D simulator. Results The proposed controller achieved an average time-in-range of 83% while maintaining real-time on-device execution with millijoule-level energy consumption per control step. Discussion These hardware-in-the-loop and in-silico results demonstrate the technical feasibility of real-time reinforcement learning-based insulin dosing and personalization on a resource-constrained embedded platform, while clinical efficacy and safety require further validation.

Authors

Institutions

Publication Details

Journal
Frontiers in Digital Health
Published
2026-11-13
DOI
https://doi.org/10.3389/fdgth.2026.1967489
Primary Topic
Diabetes Management and Research
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

On-device reinforcement learning-based insulin dosing controller for fully automated insulin delivery

Jina Lee, Sung-Min Park
4 links
Frontiers in Digital Health
Diabetes Management and Research
article

On-device reinforcement learning-based insulin dosing controller for fully automated insulin delivery

Jina Lee, Sung-Min Park
article en

Abstract

Introduction Automated insulin delivery (AID) systems integrate continuous glucose monitoring (CGM) with algorithm-driven insulin dosing to enable closed-loop glycemic management for people with type 1 diabetes (T1D). Recent advances in wearable glucose sensing and insulin delivery technologies have enabled increasingly compact AID systems; however, deploying advanced AI-based dosing algorithms directly on resource-constrained wearable devices remains technically challenging. Some AID architectures also rely on external devices, such as smartphones or dedicated controllers, introducing additional communication dependencies into the closed-loop operation. Methods This study presents an on-device insulin dosing controller based on the soft actor–critic (SAC) algorithm for resource-constrained AID systems. The proposed approach processes CGM data and makes insulin dosing decisions locally, without relying on external computing devices. To enable embedded deployment, a compressed SAC architecture with a lightweight encoder was developed using quantization-aware training and knowledge distillation, together with an online adaptation mechanism for individualized control. The controller was implemented on a resource-constrained microcontroller and evaluated in a hardware-in-the-loop framework using virtual patients from a T1D simulator. Results The proposed controller achieved an average time-in-range of 83% while maintaining real-time on-device execution with millijoule-level energy consumption per control step. Discussion These hardware-in-the-loop and in-silico results demonstrate the technical feasibility of real-time reinforcement learning-based insulin dosing and personalization on a resource-constrained embedded platform, while clinical efficacy and safety require further validation.

Frontiers in Digital HealthVol. 8
Pohang University of Science and Technology (KR)
National Research Foundation of Korea
Good health and well-being
Diabetes Management and Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon