---
title: 'CliCARE: Guideline-Grounded Clinical AI'
url: https://www.emergentmind.com/topics/clicare
type: topic
---

# CliCARE: Guideline-Grounded Clinical AI

CliCARE (Grounding Large Language Models in Clinical Guidelines for Decision Support over Longitudinal Cancer Electronic Health Records) is a framework designed to enhance clinical decision support by grounding large language models (LLMs) in formal clinical practice guidelines through the structured modeling of longitudinal oncology electronic health records (EHRs). CliCARE addresses fundamental challenges in EHR-based AI—namely, processing extensive multilingual records, mitigating clinical hallucination, and enabling robust, reliable evaluation—by transforming unstructured patient histories into temporally structured graphs and aligning these with prescriptive, process-oriented guideline knowledge. This approach enables the generation of evidence-grounded clinical summaries and actionable recommendations with substantially improved clinical validity and robustness [2507.22533].

## 1. System Architecture

CliCARE operates through three principal stages, culminating in prompted clinical decision support:

1. **EHR-to-TKG Transformation**: For each patient $p$, a sequence of EHR documents $D_p = (d_{\tau_1}, \ldots, d_{\tau_n})$ is divided into historical notes and a most-recent note (History of Present Illness). Compression and extractive summarization are performed using a Longformer, yielding a condensed summary $S_p^{\mathrm{hist}}$ and extracted event list $E_p = \mathrm{ExtractEvents}(S_p^{\mathrm{hist}} \oplus d_{\tau_n})$.

2. **Temporal Knowledge Graph (TKG) Instantiation**: Patient events are grounded to a biomedical knowledge graph $G_B$ via entity linking $\phi$, creating a patient-specific TKG $G_t = (\mathcal{E}_t, \mathcal{R}_t, T)$, where entities $e = (e_B, \tau, A)$ encode biomedical concept, timestamp, and attributes. Edge set $\mathcal{R}_t$ encodes intra- and inter-encounter relations (e.g., "follows", "co-occurs").

3. **Guideline Knowledge Graph Construction and Alignment**: The system constructs a guideline graph $G_g = (E_g, R_g)$ from process-oriented clinical practice guidelines (e.g., NCCN) and enumerates all normative decision paths $Pa_k = \langle s_1, \dots, s_l \rangle$. Patient trajectories $Tr_p = \langle e_1, \dots, e_m \rangle$ from the TKG are semantically aligned to these paths using BERT-based scoring, LLM-based reranking, and bootstrapped iterative expansion.

4. **Generation**: An LLM, either specialist fine-tuned or generalist zero-shot, is prompted with the compressed event list $E_p$, the aligned guideline path $Pa^*$, and explicit instructions to produce both a retrospective clinical summary $(T_{CS})$ and a next-step recommendation $(T_{CR})$.

## 2. Temporal Knowledge Graph Construction

The TKG formalism serves to transform longitudinal, multilingual, and unstructured EHRs into a structured sequence of temporally indexed clinical events. For each patient:

- **Event Extraction**: Events $e_i = (e_B^i, \tau_i, A_i)$ encode links to a global biomedical ontology and are temporally ordered.
- **Graph Building**: The patient-specific TKG $G_t = (\mathcal{E}_t, \mathcal{R}_t, T)$ is assembled, where $\mathcal{E}_t = \{e_1, \ldots, e_m\}$, $\mathcal{R}_t \subseteq \mathcal{E}_t \times \mathcal{R}_B \times \mathcal{E}_t$, and $T$ is the set of discrete timestamps. The patient trajectory $Tr_p = \langle e_1, \ldots, e_m \rangle$ encodes longitudinal clinical evolution with enriched knowledge-graph semantics.

This structure enables long-range temporal dependency modeling and facilitates downstream mapping to process-oriented guideline workflows.

## 3. Normative Guideline Graph Modeling

The guideline knowledge graph $G_g$ is derived from authoritative clinical guidelines and encodes recommended workflows as directed acyclic graphs:

- **Node Set**: $E_g$ consists of abstract guideline concepts (e.g., Cancer Stage, Treatment Line).
- **Edge Set**: $R_g$ corresponds to logical and recommendation-driven relations.
- **Path Enumeration**: All valid decision-making sequences $Pa_k = \langle s_1, \ldots, s_l \rangle$ are enumerated, representing complete prescriptive workflows from diagnosis to management.

This modeling enables direct alignment between real-world patient histories and formal normative pathways.

## 4. Patient-Guideline Alignment Mechanism

Alignment proceeds via a three-step process:

1. **Semantic Scoring**: Each patient trajectory $Tr_p$ is matched with each guideline path $Pa_k$ via a BERT-based similarity metric:

   $$
   \mathrm{Score}(Tr_p, Pa_k) = \sum_{j=1}^l \max_{e_i \in Tr_p} \cos\_sim(f_{\mathrm{BERT}}(\mathrm{desc}(s_j)), f_{\mathrm{BERT}}(\mathrm{desc}(e_i)))
   $$

   The optimal path $Pa^*$ is selected by maximizing this score.

2. **LLM-Based Reranking**: Candidate path-and-score pairs are reranked zero-shot by an LLM, yielding an ordered shortlist of aligned workflows.

3. **Alignment Expansion**: Unaligned events are matched to guideline nodes by maximizing aggregate similarity with already aligned event-stage pairs, using bootstrapped expansion (again leveraging BERT representations).

This hierarchical method ensures principled, semantically robust coverage of patient timelines and prescriptive standards.

## 5. Generation of Clinical Summaries and Recommendations

The final decision support outputs are produced by prompting an LLM with a five-part template ("CliPAGE"):

1. Compressed event list (longitudinal EHR, from TKG construction)
2. Most recent EHR note
3. Grounding evidence (the aligned guideline path $Pa^*$)
4. Request for a retrospective Clinical Summary $(T_{CS})$
5. Request for a prospective Clinical Recommendation $(T_{CR})$

This prompt induces the model to weight aligned nodes more heavily, promoting guideline-conformant reasoning. Grounding is achieved purely via in-context exemplars—no additional trainable attention mechanisms are introduced. Outputs consist of two structured sections: summary and recommendation.

## 6. Evaluation Protocol and Benchmarks

CliCARE is evaluated on two datasets:

- **CancerEHR ($D_{CEHR}$)**: 2,000 Chinese-language cases, up to 21,000 tokens per record.
- **MIMIC-Cancer ($D_{MC}$)**: 2,000 English-language cases sampled from MIMIC-IV.

Baselines include standard RAG pipelines with open-source LLMs (Mistral-7B, Mistral-Instruct, Qwen-3-8B, BioMistral-7B), long-context strategies (BriefContext), and KG-enhanced architectures (MedRAG, KG2RAG, GNN-RAG).

Metrics are LLM-as-Judge scores (composite of GPT-4.1, Claude 4.0 Sonnet, Gemini 2.5 Pro) rated 1–5 on Factual Accuracy, Completeness, Clinical Soundness, and Actionability for both $T_{CS}$ and $T_{CR}$. Correlation with oncologist assessments is measured by Spearman’s $\rho$.

| Method         | $T_{CS}$ (Qwen / Gemini) | $T_{CR}$ (Qwen / Gemini) |
|----------------|-------------------------|--------------------------|
| RAG ($D_{CEHR}$)   | 1.485 / 2.735           | 1.527 / 2.818             |
| CliCARE           | 3.173 / 4.976           | 3.215 / 4.965             |
| Ablation: -TKG    | 1.485                   | 1.527                     |
| Ablation: -LLM RR | 2.857                   | 2.866                     |
| Ablation: -Align  | 3.012                   | 3.035                     |

On $D_{MC}$, CliCARE-based models achieve +0.070 to +0.835 improvement across systems.

Length robustness findings indicate Qwen-3-8B performs best on short records, degrading at longer lengths, whereas Gemini 2.5 Pro maintains or improves performance at longer sequence lengths.

## 7. Impact and Significance

CliCARE’s integration of patient-specific temporal graphs and prescriptive guideline alignment substantially improves clinical validity, completeness, and actionability in both LLM-generated summaries and recommendations for oncology. Human-LLM scoring correlation (Spearman’s $\rho \approx 0.70$) confirms that the automated metric tracks domain expert judgment. Key ablation studies demonstrate the indispensability of TKG compression, LLM-based reranking, and alignment expansion to system performance. A plausible implication is that formalizing the explicit mapping from clinical observations to prescriptive pathways can simultaneously address hallucination and evaluation bottlenecks inherent in prior LLM-based EHR support systems.

CliCARE sets a benchmark for robust, guideline-grounded, multilingual clinical AI on longitudinal data and provides a reproducible protocol for rigorous system evaluation in oncology [2507.22533].

Source: https://www.emergentmind.com/topics/clicare