Abstract

There is an increased risk of hypoglycemia with insulin stacking, which involves taking a new insulin dose while the previous dose is still working. The works reviewed focus on dose titration or automated delivery of medication, or conversational counseling in recent studies on artificial intelligence (AI), but do not discuss stacking directly. This paper summarizes nine papers on insulin stacking, AI enabled insulin dosing and LLM-enabled clinical support, finding that no study, in its available abstract and/or text, integrates the four factors of continuous glucose monitoring (CGM), insulin history, meals, physical activity, and health-record context, into a stacking risk assessment and communication framework with explicit safety constraints for both the clinician and the patient. We suggest a new five-layer framework that builds a separation between a deterministic clinical core (insulin-on-board estimation, projected-glucose risk tiers, hard safety rules) and a retrieval-grounded generative layer which explains, but will not compute, risk. The framework is theoretical (not implemented) and no experimental or clinical results are provided. Future work is suggested in an evaluation protocol.

Keywords

How to Cite This Article

Rights