CARE-ECG: Causal Agent-based Reasoning for Explainable and Counterfactual ECG Interpretation

Elahe Khatibi; Ziyu Wang; Ankita Sharma; Krishnendu Chakrabarty; Sanaz Rahimi Moosavi; Farshad Firouzi; Amir Rahmani

cs.LGApr 12, 2026

CARE-ECG: Causal Agent-based Reasoning for Explainable and Counterfactual ECG Interpretation

Authors: Elahe Khatibi, Ziyu Wang, Ankita Sharma, Krishnendu Chakrabarty, Sanaz Rahimi Moosavi, Farshad Firouzi, Amir Rahmani

Organizations: University of California, Irvine; Irvine, California, United States · Arizona State University; Tempe, Arizona, United States · California State University, Dominguez Hills; Carson, California, United States

Abstract

Large language models (LLMs) enable waveform-to-text ECG interpretation and interactive clinical questioning, yet most ECG-LLM systems still rely on weak signal-text alignment and retrieval without explicit physiological or causal structure. This limits grounding, temporal reasoning, and counterfactual "what-if" analysis central to clinical decision-making. We propose CARE-ECG, a causally structured ECG-language reasoning framework that unifies representation learning, diagnosis, and explanation in a single pipeline. CARE-ECG encodes multi-lead ECGs into temporally organized latent biomarkers, performs causal graph inference for probabilistic diagnosis, and supports counterfactual assessment via structural causal models. To improve faithfulness, CARE-ECG grounds language outputs through causal retrieval-augmented generation and a modular agentic pipeline that integrates history, diagnosis, and response with verification. Across multiple ECG benchmarks and expert QA settings, CARE-ECG improves diagnostic accuracy and explanation faithfulness while reducing hallucinations (e.g., 0.84 accuracy on Expert-ECG-QA and 0.76 on SCP-mapped PTB-XL under GPT-4). Overall, CARE-ECG provides traceable reasoning by exposing key latent drivers, causal evidence paths, and how alternative physiological states would change outcomes.

Figures & tables

Figure 1. Motivation. Typical ECG-LLMs may over-diagnose STEMI in mimics by hallucinating confirmatory evidence (e.g., reciprocal ST depression). CARE-ECG mitigates this by reasoning over latent morphological biomarkers (e.g., concave ST morphology) and validating decisions via causal counterfactuals.

Figure 2. The CARE-ECG Framework. The Physiological Stream encodes ECGs into latent biomarkers and performs causal inference to predict

P(y)

and enable counterfactual analysis. The Agentic Stream leverages temporal context and causal drivers to guide retrieval and generation, while a Faithfulness Verifier checks signal-grounded alignment and triggers self-correction when needed.

Figure 3. Agentic reasoning workflow. CARE-ECG integrates temporal context (

\Delta z

) and causal evidence (

\tilde{z}

) before generating a grounded rationale, verified by a faithfulness gate with self-correction and fallback.

Module / Field	Prompted Instruction & Runtime Logic
1. Prompt Construction
Patient Query	Injects: Patient Query: {query}
Causal Factors	Injects: Key Causal Factors (from VAE/Graph): {formatted_causal} . (Logic: Comma-joined list of drivers; returns ’None’ if empty).
Retrieved Evidence	Injects: Retrieved Medical Facts (RAG): followed by enumerated list: [Fact i] {fact_i} .
Gen. Instruction	Static string: “Explain the prediction clearly and medically grounded , and attach citations using fact tags (e.g., [Fact 1] ).”
2. Generation Parameters

Table 1. Response Agent prompt & safety gate. The table summarizes the fact-tagged prompt template, decoding settings, hallucination scoring, and the RAG-only fallback rule used for grounded explanation generation.

Features	PTB-XL	MIMIC-IV ECG	Expert-ECG-QA
Data Characteristics
Content	Standard 12-lead ECGs with SCP-ECG statements.	Large-scale clinical ECG corpus for retrieval testing.	Multimodal QA pairs (Real + Synthetic) covering 12 tasks.
Size	21,837 records (18,885 patients)	$\approx$ 800,000 records ( $\approx$ 160,000 patients)	47,211 QA pairs (Multi-turn supported)
Input $\rightarrow$ GT	Signal $x\rightarrow$ SCP Codes	Signal $x\rightarrow$ Clinical Text	Signal $x$ + Query $q\rightarrow$ Answer
Evaluation Protocol
Core Task	Diagnosis Grounding (Text $\rightarrow$ Code Mapping)	Faithfulness Check (Retrieval Alignment)	Clinical Reasoning (Interactive QA)

Table 2. Evaluation datasets and protocols. PTB-XL evaluates diagnosis grounding via Text

\rightarrow

SCP mapping, MIMIC-IV ECG evaluates retrieval-based faithfulness, and Expert-ECG-QA evaluates interactive clinical reasoning.

Model	GPT-4 Backbone				LLaMA 3–8B Backbone
Model	Acc	Prec	Rec	F1	Acc	Prec	Rec	F1
Dataset: Expert-ECG-QA
CARE-ECG (Ours)	0.84	0.79	0.87	0.83	0.77	0.73	0.80	0.76
ECG-Chat	0.76	0.72	0.80	0.75	0.71	0.68	0.73	0.70
Q-HEART	0.74	0.70	0.77	0.73	0.69	0.66	0.70	0.68
anyECG-Chat	0.71	0.68	0.72	0.70	0.66	0.64	0.67	0.65

Table 3. Answer Quality Metrics on Expert-ECG-QA and PTB-XL. For PTB-XL, free-text outputs were mapped to SCP diagnostic codes using keyword extraction and fuzzy label matching. Bold indicates best performance.

Model	GPT-4 Backbone			LLaMA 3–8B Backbone
Model	CRC	Ground.	Cont.Rel.	CRC	Ground.	Cont.Rel.
Dataset: PTB-XL
CARE-ECG (Ours)	0.91	0.71	0.75	0.86	0.67	0.71
ECG-Chat	0.78	0.62	0.64	0.73	0.56	0.59
Q-HEART	0.75	0.60	0.61	0.71	0.54	0.57
anyECG-Chat	0.72	0.57	0.58	0.69	0.52	0.55

Table 4. Causal and Retrieval Metrics on PTB-XL, MIMIC-IV ECG, and Expert-ECG-QA. Bold denotes the best performance.

Figure 4. Faithfulness–robustness trade-off in the HR–SRS plane across datasets. Each panel shows HR (lower is better) versus SRS (higher is better) for all methods under two backbones. Points closer to the upper-left indicate more evidence-consistent explanations with fewer hallucinations.

Model	GPT-4 Backbone		LLaMA 3–8B Backbone
Model	HR ( $\downarrow$ )	SRS ( $\uparrow$ )	HR ( $\downarrow$ )	SRS ( $\uparrow$ )
Dataset: PTB-XL
CARE-ECG (Ours)	0.06	0.81	0.08	0.77
ECG-Chat	0.12	0.72	0.15	0.68
Q-HEART	0.14	0.70	0.17	0.66
anyECG-Chat	0.15	0.68	0.18	0.64

Table 5. Hallucination and Factuality Metrics. HR : Hallucination Rate (lower

\downarrow

is better); SRS : Summary Relevance Score (higher

\uparrow

is better). Bold indicates best performance.

Variant	Modules					GPT-4 Backbone						LLaMA 3–8B Backbone
Variant	Lat	Graph	RAG	Verif.	CF	CRC	SRS	Gnd.	HR $\downarrow$	Prec	F1	CRC	SRS	Gnd.	HR $\downarrow$	Prec	F1
A0: Latent Baseline	✓					0.73	0.67	0.64	0.13	0.72	0.74	0.68	0.63	0.60	0.16	0.68	0.70
A1: + Causal Graphs	✓	✓				0.83	0.72	0.69	0.10	0.75	0.78	0.78	0.69	0.66	0.13	0.72	0.75
A2: + RAG Injection	✓	✓	✓			0.91	0.76	0.71	0.08	0.76	0.79	0.85	0.72	0.68	0.11	0.74	0.76
A3: + Verifier Fallback	✓	✓	✓	✓		0.96	0.82	0.75	0.05	0.79	0.83	0.90	0.78	0.71	0.08	0.76	0.80
A4: + Counterfactual (Full)	✓	✓	✓	✓	✓	0.97	0.84	0.77	0.04	0.81	0.85	0.91	0.80	0.73	0.07	0.78	0.82

Table 6. Ablation Study on Expert-ECG-QA (GPT-4 vs. LLaMA 3–8B). “Latent” denotes CausalVAE+temporal backbone; “Graph” denotes Bayesian inference over

\tilde{z}

; “RAG” denotes causal-conditioned retrieval; “Verif.” denotes verifier-triggered fallback; “CF” denotes counterfactual minimal edits. Bold indicates best performance.

Panel of Figure 5 — Figure 5. Incremental ablation curves for CARE-ECG. Each panel visualizes how key metrics evolve as modules are progressively added (A0 $\rightarrow$ A4 in Table 6 ). The curves provide a compact view of the monotonic improvements and the stage at which each metric changes most strongly.

Figure 6. Radar-style summary of multi-metric trade-offs. The radar view summarizes the consistent advantage of CARE-ECG across complementary reliability dimensions.

Explore similar work

ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation
Mar 15, 2026Jungwoo Oh, Hyunseung Chung, Junhee Lee +6Clinical Reasoning Multimodal Large Language Models
KAIST · Ajou University School of Medicine · Yonsei University College of Medicine +1
Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations
Aug 10, 2026Hongxiang Gao, He-yang Xu, Yuwen Li +6Electrocardiography Electrocardiogram Classification
State Key Laboratory of Digital Medical Engineering, Southeast University · School of Instrument Science and Engineering, Southeast University
Enhancing Explainable Cardiac Diagnosis with Guide-Grounded Multimodal LLMs
Jul 23, 2026Hai-Nam Duy Vuong, Duy-Anh Bui, Trong-Nghia Nguyen +5Multimodal Grounding Medical Report Generation
Business AI Lab, College of Technology, National Economics University, Vietnam · FPT University, Hanoi, Vietnam · A2I Lab, Phenikaa School of Computing, Phenikaa University, Hanoi, Vietnam +2

Mar 15, 2026J/K move · Enter open · S save

ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation

Jungwoo Oh, Hyunseung Chung, Junhee Lee, Min-Gyu Kim+5

KAIST · Ajou University School of Medicine · Yonsei University College of Medicine+1

Clinical Reasoning Multimodal Large Language Models Multimodal Model Evaluation Electrocardiography

Open paper AskPDF arXiv

While Multimodal Large Language Models (MLLMs) show promising performance in automated electrocardiogram interpretation, it remains unclear whether they genuinely perform actual step-by-step reasoning or just rely on superficial visual cues. To investigate this, we introduce ECG-Reasoning-Benchmark, a novel multi-turn evaluation framework comprising over 6,400 samples to systematically assess step-by-step reasoning across 17 core ECG diagnoses. Our comprehensive evaluation of state-of-the-art models reveals a critical failure in executing multi-step logical deduction. Although models possess the medical knowledge to retrieve clinical criteria for a diagnosis, they exhibit near-zero success rates (6% Completion) in maintaining a complete reasoning chain, primarily failing to ground the corresponding ECG findings to the actual visual evidence in the ECG signal. These results demonstrate that current MLLMs bypass actual visual interpretation, exposing a critical flaw in existing training paradigms and underscoring the necessity for robust, reasoning-centric medical AI. The code and data are available at https://github.com/Jwoo5/ecg-reasoning-benchmark.

Abstract

Explore similar work

ECG-Reasoning-Benchmark: A Benchmark for Evaluating Clinical Reasoning Capabilities in ECG Interpretation

Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations

Enhancing Explainable Cardiac Diagnosis with Guide-Grounded Multimodal LLMs