From Guideline Grounding to Trustworthy Clinical Decision Support: A Systematic Scoping Review of Patient-Specific LLM Systems

Large language model (LLM) systems can combine clinical guidelines with individual patient information to support patient-specific clinical decisions. Trustworthy guideline-grounded decision support depends on whether recommendations are applicable to the patient, supported by the guideline source, appropriately safeguarded, and clinically evaluated. This systematic scoping review examined how guideline-grounded patient-specific LLM systems represent and integrate guideline and patient information, determine recommendation applicability, support traceability and safeguards, evaluate clinical performance, and provide resources for independent review and reconstruction. Scopus, Web of Science, PubMed, and IEEE Xplore were searched through 1 October 2026. Eligibility required an LLM component, explicit integration of external clinical guideline knowledge, individual patient information, and a clinical decision or appropriateness function. The 160 included publications correspond to 158 analytical records. The corpus is heterogeneous in patient-data provenance, patient representation, guideline representation, model families, integration routes, and implementation patterns. Systems use narrative, structured, longitudinal, conversational, and multimodal patient information, together with guideline text, structured representations, graphs, and formalized logic. OpenAI model families are most frequently represented, followed by Gemini, Llama, DeepSeek, Claude, Qwen, and Mistral. Guideline integration uses guideline content, retrieval, and formalized logic, while implementation patterns include executable logic, prompt-level logic, agent-based, neuro-symbolic, and graph-based designs, and similar-case approaches. Retrieval is used in 85 records, guideline content in 67, and formalized guideline logic in 48. Real patient records or observations are used in 66 records, constructed or simulated cases in 57, patient-derived cases in 15, and mixed-provenance data in 12. Patient-specific applicability factors are described in 145 records (91.8%), whereas explicit applicability logic is confirmed in 30 (19.0%). Recommendation-linked source traces are identified in 53 records (33.5%), while semantic source-support assessment appears in eight (5.1%). External clinical assessment is confirmed in five records (3.2%) and participant-facing use in four (2.5%). Reported resource availability is more common than reviewed study-linked materials, and only four records combine reviewed code or implementation logic, patient or case data, guideline resources, and evaluation assets. Cross-dimensional analyses show that integration routes differ in their patterns of applicability, traceability, and safeguards, while external clinical assessment, participant-facing use, and broad coverage of reviewed study resources occur largely in different subsets of the corpus. These relationships show how system design, patient-data provenance, clinical assessment, and reproducibility intersect across the patient-to-recommendation process. Sensitivity analyses preserve the main pattern of findings. Overall, the least frequently evaluated parts of the patient-to-recommendation pathway concern explicit applicability, semantic source support, safeguard-specific evaluation, external clinical assessment, and participant-facing use. The findings identify a “missing middle” between guideline integration and an evaluated patient-specific recommendation, where patient information must be linked explicitly to the conditions governing that recommendation. Future research should strengthen this patient-to-recommendation connection through explicit applicability logic, source-support assessment, safeguard evaluation, clinical assessment, and study resources supporting reproducibility.

Authors

Institutions

Publication Details

Journal
Machine Learning and Knowledge Extraction
Published
2026-11-12
DOI
https://doi.org/10.3390/make8110376
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

From Guideline Grounding to Trustworthy Clinical Decision Support: A Systematic Scoping Review of Patient-Specific LLM Systems

Vеskа Gаnchеvа
Machine Learning and Knowledge Extraction
Artificial Intelligence in Healthcare and Education
article

From Guideline Grounding to Trustworthy Clinical Decision Support: A Systematic Scoping Review of Patient-Specific LLM Systems

Vеskа Gаnchеvа
article en

Abstract

Large language model (LLM) systems can combine clinical guidelines with individual patient information to support patient-specific clinical decisions. Trustworthy guideline-grounded decision support depends on whether recommendations are applicable to the patient, supported by the guideline source, appropriately safeguarded, and clinically evaluated. This systematic scoping review examined how guideline-grounded patient-specific LLM systems represent and integrate guideline and patient information, determine recommendation applicability, support traceability and safeguards, evaluate clinical performance, and provide resources for independent review and reconstruction. Scopus, Web of Science, PubMed, and IEEE Xplore were searched through 1 October 2026. Eligibility required an LLM component, explicit integration of external clinical guideline knowledge, individual patient information, and a clinical decision or appropriateness function. The 160 included publications correspond to 158 analytical records. The corpus is heterogeneous in patient-data provenance, patient representation, guideline representation, model families, integration routes, and implementation patterns. Systems use narrative, structured, longitudinal, conversational, and multimodal patient information, together with guideline text, structured representations, graphs, and formalized logic. OpenAI model families are most frequently represented, followed by Gemini, Llama, DeepSeek, Claude, Qwen, and Mistral. Guideline integration uses guideline content, retrieval, and formalized logic, while implementation patterns include executable logic, prompt-level logic, agent-based, neuro-symbolic, and graph-based designs, and similar-case approaches. Retrieval is used in 85 records, guideline content in 67, and formalized guideline logic in 48. Real patient records or observations are used in 66 records, constructed or simulated cases in 57, patient-derived cases in 15, and mixed-provenance data in 12. Patient-specific applicability factors are described in 145 records (91.8%), whereas explicit applicability logic is confirmed in 30 (19.0%). Recommendation-linked source traces are identified in 53 records (33.5%), while semantic source-support assessment appears in eight (5.1%). External clinical assessment is confirmed in five records (3.2%) and participant-facing use in four (2.5%). Reported resource availability is more common than reviewed study-linked materials, and only four records combine reviewed code or implementation logic, patient or case data, guideline resources, and evaluation assets. Cross-dimensional analyses show that integration routes differ in their patterns of applicability, traceability, and safeguards, while external clinical assessment, participant-facing use, and broad coverage of reviewed study resources occur largely in different subsets of the corpus. These relationships show how system design, patient-data provenance, clinical assessment, and reproducibility intersect across the patient-to-recommendation process. Sensitivity analyses preserve the main pattern of findings. Overall, the least frequently evaluated parts of the patient-to-recommendation pathway concern explicit applicability, semantic source support, safeguard-specific evaluation, external clinical assessment, and participant-facing use. The findings identify a “missing middle” between guideline integration and an evaluated patient-specific recommendation, where patient information must be linked explicitly to the conditions governing that recommendation. Future research should strengthen this patient-to-recommendation connection through explicit applicability logic, source-support assessment, safeguard evaluation, clinical assessment, and study resources supporting reproducibility.

Machine Learning and Knowledge ExtractionVol. 8(11)
Technical University of Sofia (BG)
Good health and well-being
Artificial Intelligence in Healthcare and Education
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