Peak-R1: Instruction-Tuned Large Language Models for Robust J-Peak Detection in Cardiomechanical Signals

Cardiomechanical signals, encompassing ballistocardiography (BCG) and the bodyseismogram (BSG), represent a promising modality for unobtrusive and continuous assessment of cardiovascular health. The J-peak, a key fiducial point within the cardiomechanical signal, serves as a robust surrogate for cardiac timing, underpinning heart rate (HR) estimation and hemodynamic modeling. However, precise J-peak localization is frequently confounded by annotation ambiguities, inter-subject signal variability, and motion artifacts. We introduce \textbf{Peak-R1}, a novel framework that leverages an instruction-tuned Large Language Model (LLM) for robust J-peak detection. Central to our approach is a peak-extraction front end that transforms raw BCG segments into compact peak sequences. This peak-centric representation reduces noise and introduces a principled inductive bias, guiding the LLM to focus on physiologically meaningful events and thereby improving its reasoning over time-series data. Peak-R1 is trained via a two-stage strategy: (i) supervised fine-tuning (SFT) to establish stable output formatting and baseline signal interpretation, followed by (ii) reinforcement learning (RL) with Group Relative Policy Optimization (GRPO). The RL stage employs a multi-objective reward function to jointly optimize for output validity, HR consistency, absolute localization accuracy, and detection completeness. The framework achieves an F1 score of 0.930 and HR mean absolute error (MAE) of 0.399 BPM on the Kansas dataset, while maintaining robust performance (F1: 0.770, HR MAE: 7.002 BPM) on the more challenging hospital-BSG dataset. Our ablation studies confirm the necessity of the peak-extraction front end and reveal that RL tuning is critical for improving detection accuracy.