[Paper Review] PPGnet: Deep Network for Device Independent Heart Rate Estimation from Photoplethysmogram

A Shyam, Vignesh Ravichandran+3 more|arXiv (Cornell University)|Mar 21, 2019

Non-Invasive Vital Sign Monitoring8 references4 citations

TL;DR

This paper proposes PPGnet, an end-to-end deep learning model for device-independent heart rate estimation using single-channel wrist PPG signals without requiring patient-specific training or tri-axial accelerometer data. The model achieves a mean absolute error of 3.36 ± 4.1 BPM on the IEEE SPC 2015 dataset and demonstrates feasibility of transfer learning with sparse retraining on a new in-house PPG dataset, enabling generalization across different PPG hardware designs.

ABSTRACT

Photoplethysmogram (PPG) is increasingly used to provide monitoring of the cardiovascular system under ambulatory conditions. Wearable devices like smartwatches use PPG to allow long term unobtrusive monitoring of heart rate in free living conditions. PPG based heart rate measurement is unfortunately highly susceptible to motion artifacts, particularly when measured from the wrist. Traditional machine learning and deep learning approaches rely on tri-axial accelerometer data along with PPG to perform heart rate estimation. The conventional learning based approaches have not addressed the need for device-specific modeling due to differences in hardware design among PPG devices. In this paper, we propose a novel end to end deep learning model to perform heart rate estimation using 8 second length input PPG signal. We evaluate the proposed model on the IEEE SPC 2015 dataset, achieving a mean absolute error of 3.36+-4.1BPM for HR estimation on 12 subjects without requiring patient specific training. We also studied the feasibility of applying transfer learning along with sparse retraining from a comprehensive in house PPG dataset for heart rate estimation across PPG devices with different hardware design.

Motivation & Objective

• To develop a deep learning model for heart rate estimation from wrist PPG signals that does not require patient-specific retraining.
• To eliminate dependency on tri-axial accelerometer data to reduce battery consumption in wearable devices.
• To enable generalization across PPG devices with varying hardware designs through transfer learning.
• To evaluate the model’s robustness under motion artifacts using diverse datasets, including the IEEE SPC 2015 and an in-house ADI dataset.
• To investigate the feasibility of sparse retraining for new PPG devices using a pre-trained model.

Proposed method

• The model uses a hybrid architecture combining 1D convolutional neural network (CNN) layers and long short-term memory (LSTM) layers to extract temporal and spectral features from 8-second single-channel PPG signals.
• The network is trained end-to-end on raw PPG signals without preprocessing, avoiding reliance on external sensor data such as accelerometers.
• Transfer learning is applied by initializing the model with pre-trained weights from the IEEE SPC 2015 dataset and fine-tuning on a sparse subset (15%) of a new in-house PPG dataset (ADI).
• Model convergence is evaluated under four conditions: random initialization on ADI, random initialization on ADI with full data, transfer learning with full ADI data, and transfer learning with 15% ADI data.
• Performance is assessed using mean absolute error (MAE), standard deviation of absolute error (SDAE), and Pearson correlation coefficient (PCC).
• The model is evaluated on both the IEEE SPC 2015 dataset and the in-house ADI dataset to test robustness under varying motion conditions and heart rate ranges (39–155 BPM).

Experimental results

Research questions

• RQ1 Can a deep learning model estimate heart rate from single-channel wrist PPG signals without requiring patient-specific retraining?
• RQ2 Can the model generalize across different PPG hardware designs without retraining from scratch?
• RQ3 Does transfer learning with sparse retraining on a new dataset enable comparable performance to full retraining?
• RQ4 How does the model perform under high-motion conditions, particularly with randomized motion artifacts?
• RQ5 Can the model achieve low error and high correlation without using accelerometer data?

Key findings

• The model achieved a mean absolute error of 3.36 ± 4.1 BPM on the IEEE SPC 2015 dataset, demonstrating state-of-the-art performance without preprocessing or accelerometer data.
• On the IEEE SPC 2015 test set with intense physical movements, the model achieved a mean absolute error of 12.48 ± 14.45 BPM, indicating robustness under motion artifacts.
• When trained from random initialization on the in-house ADI dataset, the model achieved a mean absolute error of 4.10 ± 5.19 BPM with a Pearson correlation coefficient of 0.92.
• Using transfer learning with pre-trained weights from IEEE SPC 2015 and fine-tuning on the full ADI dataset, the model achieved 4.095 ± 5.25 BPM MAE, nearly matching the performance of full training.
• Even with only 15% of the ADI dataset used for retraining, the model achieved 4.11 ± 5.27 BPM MAE, showing that minimal retraining suffices for new hardware devices.
• The model demonstrated adaptability to rapid motion transitions, maintaining accuracy despite high accelerometer energy in the 250–400 window, indicating resilience to motion artifacts.