# Efficient and Direct Inference of Heart Rate Variability using Both Signal Processing and Machine Learning

**Authors:** Yuntong Zhang, Jingye Xu, Mimi Xie, Dakai Zhu, Houbing Song, Wei Wang  
**arXiv:** [2303.13637](https://arxiv.org/abs/2303.13637)  
**Date:** 2023-03-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.13637v1/

## Key points

- Heart Rate Variability (HRV) is an important indicator of physical and mental health
- Photoplethysmography (PPG) sensors can be used to infer HRV
- Yuntong Zhang et al. collected a large dataset of PPG signals and HRV ground truth to address limitations of prior studies
- They developed HRV models combining signal processing and machine learning for direct inference of HRV with errors between 3.5% to 25.7%
- Decision Trees (DT) and Multi-level Perceptrons (MLP) are more suitable for small embedded devices with good accuracy
- A new dataset containing three 2-hour-long PPG/ECG traces for one human subject performing various activities was collected
- This study provides an efficient HRV inference methodology using both signal processing and machine learning techniques that can potentially enable the future use of PPG-based HRV monitoring in healthcare on small embedded devices with limited memory size requirements

## Summary

Heart Rate Variability (HRV) is a significant indicator of physical and mental health, measuring the variation of time between consecutive heartbeats. Recent research has shown that photoplethysmography (PPG) sensors can be used to infer HRV. To address the limitations of prior studies relying on signal processing or machine learning (ML), indirect inference of HRV, small datasets, and large ML models which restrict their applications to small embedded devices and potential use in healthcare, Yuntong Zhang et al. collected a large dataset of PPG signals and HRV ground truth. They developed HRV models combining signal processing and ML for direct inference of HRV which achieved errors between 3.5% to 25.7%, outperforming signal-processing-only and ML-only methods. The authors explored different ML models showing that Decision Trees (DT) and Multi-level Perceptrons (MLP) are more suitable for small embedded devices with good accuracy; DT models had an average error of 13.0% with less than 10KB size and inference time less than 10𝜇s while MLP models had an average error of 9.1% with less than 469KB size and inference time less than 1ms. Additionally, they collected a new dataset containing three 2-hour-long PPG/ECG traces for one human subject performing various activities such as work, sleeping, and sitting. This study provides an efficient HRV inference methodology using both signal processing and machine learning techniques that can potentially enable the future use of PPG-based HRV monitoring in healthcare on small embedded devices with limited memory size requirements.