C-GRASP: Clinically-Grounded Reasoning for Affective Signal Processing

TL;DR

This work tackles the challenges of applying LLMs to HRV interpretation by addressing RSA contamination, short-data nonlinear instability, and population-norm bias. It introduces C-GRASP, a guardrailed Retrieval-Augmented Generation pipeline that decomposes HRV interpretation into eight traceable steps and employs a Dual Z-score Priority Hierarchy to prioritize individualized baselines. Evaluated on the DREAMER dataset (414 trials), C-GRASP with high-scale reasoning (e.g., MedGemma3-thinking) achieves 37.3% accuracy in 4-class emotion classification and a Clinical Reasoning Consistency score of 69.6%, with ablations showing the Delta Z-score module as a critical anchor. The approach enables transparent, evidence-based clinical decision support and mitigates spectral and numerical hallucinations, paving the way for safer AI integration in biomedical engineering.

Abstract

Heart rate variability (HRV) is a pivotal noninvasive marker for autonomic monitoring; however, applying Large Language Models (LLMs) to HRV interpretation is hindered by physiological hallucinations. These include respiratory sinus arrhythmia (RSA) contamination, short-data instability in nonlinear metrics, and the neglect of individualized baselines in favor of population norms. We propose C-GRASP (Clinically-Grounded Reasoning for Affective Signal Processing), a guardrailed RAG-enhanced pipeline that decomposes HRV interpretation into eight traceable reasoning steps. Central to C-GRASP is a Z-score Priority Hierarchy that enforces the weighting of individualized baseline shifts over normative statistics. The system effectively mitigates spectral hallucinations through automated RSA-aware guardrails, preventing contamination of frequency-domain indices. Evaluated on 414 trials from the DREAMER dataset, C-GRASP integrated with high-scale reasoning models (e.g., MedGemma3-thinking) achieved superior performance in 4-class emotion classification (37.3% accuracy) and a Clinical Reasoning Consistency (CRC) score of 69.6%. Ablation studies confirm that the individualized Delta Z-score module serves as the critical logical anchor, preventing the "population bias" common in native LLMs. Ultimately, C-GRASP transitions affective computing from black-box classification to transparent, evidence-based clinical decision support, paving the way for safer AI integration in biomedical engineering.

Figures (10)

• Figure 1: The C-GRASP System Architecture. The framework integrates individualized feature normalization, dynamic RAG retrieval, and guardrailed stepwise reasoning to generate clinically traceable reports.
• Figure 2: Task performance metrics (T1--T3) across models.
• Figure 3: Cross-model consistency (C1: State Agreement) between the full C-GRASP system and ablation variants. The figure compares the percentage of state predictions that match the full system baseline across different component configurations (w/o RAG, w/o Guardrails, w/o $\Delta$Z). Higher values indicate greater output stability and consistency.
• Figure 4: Clinical Reasoning Consistency (CRC) scores across models.
• Figure 5: Macro and Weighted F1 across models.