Set-Prediction-Based J-Peak Detection for Pillow-Based Ballistocardiography

TL;DR

This paper introduces a set-prediction framework for detecting J-peaks in pillow-based ballistocardiography, improving accuracy and efficiency over traditional segmentation methods for sleep heart rate monitoring.

Contribution

It presents a novel set-prediction approach for BCG J-peak detection, eliminating the need for heuristic post-processing and reducing model complexity.

Findings

Outperforms U-Net-based segmentation baseline in detection accuracy

Reduces model parameters and computational complexity

Provides a publicly available BCG-ECG dataset with annotated J-peaks

Abstract

J-peak detection in ballistocardiography (BCG) is a key component of unobtrusive heart rate monitoring during sleep. Most existing approaches formulate this task as a dense time-point segmentation problem and rely on heuristic post-processing to convert continuous responses into discrete peak events, resulting in redundant model structures and sensitivity to parameter settings. In this work, we construct and publicly release a pillow-based BCG--ECG dataset consisting of multi-subject, multi-night natural sleep recordings with manually annotated BCG J-peaks. Based on this dataset, we propose a set-prediction-based J-peak detection framework that directly models peaks as discrete temporal events, eliminating the need for high-resolution segmentation heads and explicit peak suppression. Experimental results show that, under a shared convolutional backbone, the proposed method achieves superior detection performance compared to a U-Net-based segmentation baseline, while substantially reducing model parameters and computational complexity. These results indicate that event-level set prediction provides a concise and efficient modeling paradigm for BCG J-peak detection in sleep monitoring.