AI Clin Med

AI in Clinical Medicine, ISSN 2819-7437 online, Open Access

Article copyright, the authors; Journal compilation copyright, AI Clin Med and Elmer Press Inc

Journal website https://aicm.elmerpub.com

Review

Volume 2, August 2026, e30

Large Language Models for Diabetes Care Planning, Patient Education and Patient Safety

↓ **Table 1.** Potential Clinician-Facing LLM-CDS Use Cases In Diabetes Care
Clinical domain	Potential LLM contribution	Required clinical inputs	Safeguard	Supporting references
CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; GLP-1 RA: glucagon-like peptide-1 receptor agonist; HbA1c: glycated hemoglobin; LLM: large language model.
Medication optimization	Draft options for treatment intensification, de-intensification, renal-dose review, hypoglycemia-risk review, and cardiorenal risk reduction	Diagnosis, HbA1c, eGFR, albuminuria, atherosclerotic cardiovascular disease status, heart failure status, hypoglycemia history, current therapy, treatment goals	Link each suggestion to current guideline or formulary source; require clinician approval	[5, 6, 9, 10, 17]
GLP-1 RA perioperative planning	Summarize aspiration-risk considerations, glycemic consequences of withholding therapy, and perioperative trade-offs	Procedure type, fasting plan, gastrointestinal symptoms, diabetes control, anesthetic plan, indication for GLP-1 RA	Require anesthetist or endocrinologist review; avoid automatic continuation or withholding advice	[1, 6, 17]
CGM narrative interpretation	Convert glucose metrics and patterns into clinician-readable summaries, including time in range, nocturnal hypoglycemia, variability, and possible behavioral triggers	CGM metrics, insulin timing, meals, activity, symptoms, medication changes	Verify LLM-generated summaries against raw CGM data and device reports	[5, 7, 8]
Visit preparation	Generate structured pre-consult summaries from electronic health record data, recent laboratory results, medication lists, and complication screening status	Diagnoses, medications, laboratory trends, screening results, prior notes, admissions	Clinician confirms factual accuracy before use in documentation or decision-making	[7–14, 17, 18]
Patient education	Draft tailored explanations on medicines, monitoring, diet, weight, physical activity, and culturally relevant nutritional approaches	Literacy level, language, dietary pattern, cultural context, treatment plan, comorbidities	Use approved patient-education material; clinician review before release	[2–4, 7, 8, 18, 19]
Multidisciplinary care coordination	Draft referral letters, shared care plans, follow-up checklists, and team communication summaries	Complication status, care goals, pending investigations, team roles, patient preferences	Maintain audit trail; clarify responsible clinician and intended recipient	[17, 20, 21]

↓ Table 1. Potential Clinician-Facing LLM-CDS Use Cases In Diabetes Care

Clinical domain

Potential LLM contribution

Required clinical inputs

Safeguard

Supporting references

CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; GLP-1 RA: glucagon-like peptide-1 receptor agonist; HbA1c: glycated hemoglobin; LLM: large language model.

Medication optimization

Draft options for treatment intensification, de-intensification, renal-dose review, hypoglycemia-risk review, and cardiorenal risk reduction

Diagnosis, HbA1c, eGFR, albuminuria, atherosclerotic cardiovascular disease status, heart failure status, hypoglycemia history, current therapy, treatment goals

Link each suggestion to current guideline or formulary source; require clinician approval

[5, 6, 9, 10, 17]

GLP-1 RA perioperative planning

Summarize aspiration-risk considerations, glycemic consequences of withholding therapy, and perioperative trade-offs

Procedure type, fasting plan, gastrointestinal symptoms, diabetes control, anesthetic plan, indication for GLP-1 RA

Require anesthetist or endocrinologist review; avoid automatic continuation or withholding advice

[1, 6, 17]

CGM narrative interpretation

Convert glucose metrics and patterns into clinician-readable summaries, including time in range, nocturnal hypoglycemia, variability, and possible behavioral triggers

CGM metrics, insulin timing, meals, activity, symptoms, medication changes

Verify LLM-generated summaries against raw CGM data and device reports

[5, 7, 8]

Visit preparation

Generate structured pre-consult summaries from electronic health record data, recent laboratory results, medication lists, and complication screening status

Diagnoses, medications, laboratory trends, screening results, prior notes, admissions

Clinician confirms factual accuracy before use in documentation or decision-making

[7–14, 17, 18]

Patient education

Draft tailored explanations on medicines, monitoring, diet, weight, physical activity, and culturally relevant nutritional approaches

Literacy level, language, dietary pattern, cultural context, treatment plan, comorbidities

Use approved patient-education material; clinician review before release

[2–4, 7, 8, 18, 19]

Multidisciplinary care coordination

Draft referral letters, shared care plans, follow-up checklists, and team communication summaries

Complication status, care goals, pending investigations, team roles, patient preferences

Maintain audit trail; clarify responsible clinician and intended recipient

[17, 20, 21]

↓ **Table 2.** Safety Architecture for LLM-Enabled Diabetes CDS
Layer	Function	Practical design feature	Example in diabetes care	Supporting references
CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; LLM: large language model.
Data governance	Protect privacy, confidentiality, and data integrity	Role-based access, audit logs, data minimization, de-identification where appropriate	Restrict CGM and electronic health record access to authorized members of the care team	[18, 19]
Retrieval layer	Ground outputs in approved clinical evidence	Retrieval-augmented generation using local guidelines, formularies, institutional protocols, and current standards	Retrieve current diabetes pharmacotherapy, CGM, renal, and cardiovascular guidance before drafting advice	[5, 6, 14, 17, 18]
Clinical logic layer	Constrain model outputs using safety rules	Rule checks for eGFR, pregnancy, hypoglycemia history, allergies, sodium-glucose cotransporter-2 inhibitor sick-day rules, and perioperative status	Flag renal dosing concerns, hypoglycemia risk, or perioperative medication issues	[1, 5, 6, 17]
Explanation layer	Make the basis of outputs reviewable	Source-linked rationale, missing-data list, uncertainty statement, and distinction between retrieved facts and generated suggestions	State: “albuminuria unavailable; renal-risk recommendation incomplete”	[14, 17, 18]
Human review layer	Preserve professional judgment and accountability	Mandatory clinician approval before patient-facing advice or treatment changes	Clinician edits and approves medication or education plan before release	[17–19, 21]
Monitoring layer	Detect failures after deployment	Error reporting, performance audit, subgroup analysis, model-version tracking, and drift assessment	Monitor advice quality across age, language, renal function, ethnicity, disability, and technology access	[18–21, 25, 26]

↓ Table 2. Safety Architecture for LLM-Enabled Diabetes CDS

Layer

Function

Practical design feature

Example in diabetes care

Supporting references

CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; LLM: large language model.

Data governance

Protect privacy, confidentiality, and data integrity

Role-based access, audit logs, data minimization, de-identification where appropriate

Restrict CGM and electronic health record access to authorized members of the care team

[18, 19]

Retrieval layer

Ground outputs in approved clinical evidence

Retrieval-augmented generation using local guidelines, formularies, institutional protocols, and current standards

Retrieve current diabetes pharmacotherapy, CGM, renal, and cardiovascular guidance before drafting advice

[5, 6, 14, 17, 18]

Clinical logic layer

Constrain model outputs using safety rules

Rule checks for eGFR, pregnancy, hypoglycemia history, allergies, sodium-glucose cotransporter-2 inhibitor sick-day rules, and perioperative status

Flag renal dosing concerns, hypoglycemia risk, or perioperative medication issues

[1, 5, 6, 17]

Explanation layer

Make the basis of outputs reviewable

Source-linked rationale, missing-data list, uncertainty statement, and distinction between retrieved facts and generated suggestions

State: “albuminuria unavailable; renal-risk recommendation incomplete”

[14, 17, 18]

Human review layer

Preserve professional judgment and accountability

Mandatory clinician approval before patient-facing advice or treatment changes

Clinician edits and approves medication or education plan before release

[17–19, 21]

Monitoring layer

Detect failures after deployment

Error reporting, performance audit, subgroup analysis, model-version tracking, and drift assessment

Monitor advice quality across age, language, renal function, ethnicity, disability, and technology access

[18–21, 25, 26]

↓ **Table 3.** Suggested Evaluation Domains for LLM-CDS in Diabetes Care
Evaluation domain	Example metric	Why it matters	Supporting references
ADA: American Diabetes Association; CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; HbA1c: glycated hemoglobin; LLM: large language model.
Factual accuracy	Percentage of statements supported by cited guideline, formulary, or patient record	Reduces hallucinated, unsupported, or outdated advice	[14, 17, 18]
Guideline concordance	Agreement with ADA or local diabetes standards across predefined scenarios	Tests whether outputs align with accepted diabetes care	[5, 6, 17]
Missing-data awareness	Rate of appropriate requests for absent HbA1c, eGFR, albuminuria, pregnancy status, hypoglycemia history, or CGM data	Prevents premature recommendations when critical information is absent	[5, 6, 15, 16]
Safety	Serious error rate, near-miss rate, and rate of unsafe treatment suggestions in simulation or live use	Captures clinically meaningful harm beyond answer accuracy	[17–21]
Calibration	Agreement between expressed certainty and correctness	Reduces overconfident incorrect outputs	[14–16]
Equity	Performance across language, age, sex, ethnicity, socioeconomic status, disability, rurality, renal disease, and technology access	Prevents amplification of diabetes-related health disparities	[18, 19, 21, 26, 27]
Workflow impact	Consultation time, documentation burden, alert fatigue, clinician acceptance, and override rate	Determines whether the tool improves or disrupts clinical work	[20, 21, 25, 26]
Patient outcomes	HbA1c, time in range, severe hypoglycemia, admissions, medication adherence, patient understanding, and treatment satisfaction	Links LLM-CDS to clinically meaningful benefit	[5, 6, 20, 21]

↓ Table 3. Suggested Evaluation Domains for LLM-CDS in Diabetes Care

Evaluation domain

Example metric

Why it matters

Supporting references

ADA: American Diabetes Association; CDS: clinical decision support; CGM: continuous glucose monitoring; eGFR: estimated glomerular filtration rate; HbA1c: glycated hemoglobin; LLM: large language model.

Factual accuracy

Percentage of statements supported by cited guideline, formulary, or patient record

Reduces hallucinated, unsupported, or outdated advice

[14, 17, 18]

Guideline concordance

Agreement with ADA or local diabetes standards across predefined scenarios

Tests whether outputs align with accepted diabetes care

[5, 6, 17]

Missing-data awareness

Rate of appropriate requests for absent HbA1c, eGFR, albuminuria, pregnancy status, hypoglycemia history, or CGM data

Prevents premature recommendations when critical information is absent

[5, 6, 15, 16]

Safety

Serious error rate, near-miss rate, and rate of unsafe treatment suggestions in simulation or live use

Captures clinically meaningful harm beyond answer accuracy

[17–21]

Calibration

Agreement between expressed certainty and correctness

Reduces overconfident incorrect outputs

[14–16]

Equity

Performance across language, age, sex, ethnicity, socioeconomic status, disability, rurality, renal disease, and technology access

Prevents amplification of diabetes-related health disparities

[18, 19, 21, 26, 27]

Workflow impact

Consultation time, documentation burden, alert fatigue, clinician acceptance, and override rate

Determines whether the tool improves or disrupts clinical work

[20, 21, 25, 26]

Patient outcomes

HbA1c, time in range, severe hypoglycemia, admissions, medication adherence, patient understanding, and treatment satisfaction

Links LLM-CDS to clinically meaningful benefit

[5, 6, 20, 21]