UC Irvine's CARE-ECG Cuts Medical AI Hallucinations in Half With Causal Reasoning
A new framework grounds ECG language models in causal graphs and counterfactual reasoning, cutting hallucinations to 4% while lifting diagnostic accuracy to 84%.
ProTakeaways
- CARE-ECG routes ECG interpretation through latent biomarkers, a Bayesian causal graph, and a verifier before generating text.
- Hits 0.84 accuracy on Expert-ECG-QA and 0.76 on SCP-mapped PTB-XL under GPT-4.
- Hallucination Rate drops to 0.04-0.06, roughly half of ECG-Chat, Q-HEART, and anyECG-Chat baselines.
- Supports counterfactual queries like whether shortening the QT interval would change MI risk.
- Gains persist under LLaMA 3-8B, suggesting the pipeline, not the backbone, drives improvements.
- No public code release; causal edges reflect observational dependencies, not interventional causality.
A team from UC Irvine, Arizona State, and Cal State Dominguez Hills has proposed CARE-ECG, an ECG-language reasoning framework that tries to fix a persistent problem with medical LLMs: they generate fluent, confident clinical narratives that are not actually tied to what the waveform shows. Rather than mapping raw ECG signals straight into text, CARE-ECG routes every prediction through a discrete set of latent biomarkers, a Bayesian causal graph, and a verifier that checks whether the final explanation matches retrieved evidence.
Why fluent ECG chatbots keep getting it wrong
Recent work has shown that general-purpose LLMs like GPT-4, Gemini, and Claude reach only moderate accuracy on standardized ECG teaching cases, with shared weaknesses in axis determination and bundle branch block interpretation. A separate benchmark from KAIST found something more damning: models possess the medical knowledge to retrieve clinical criteria for a diagnosis, yet exhibit near-zero success rates in maintaining a complete reasoning chain, primarily failing to ground the corresponding findings to the actual visual evidence in the ECG signal.
CARE-ECG targets exactly this gap. Most existing ECG-LLM systems rely on weakly supervised alignment between waveforms and text, combined with query-driven retrieval mechanisms that are agnostic to physiological structure and causal dependencies. The authors show a representative failure where a typical ECG-LLM over-commits to acute STEMI by hallucinating confirmatory patterns such as reciprocal ST depression instead of relying on traceable waveform criteria.