Scholarlink

Photoplethysmography (PPG) signals are increasingly being integrated into wearable devices for health monitoring due to their continuous, non-invasive, and low-cost nature. Beyond physiological assessment, PPG has emerged as a promising biometric modality for user identification, offering potential for personalized healthcare services and enhanced security in wearable environments. This paper presents a survey of PPG-based biometric recognition methods, categorized into traditional machine learning approaches and deep learning approaches. Ablation experiments conducted on two benchmark datasets, PPG-DaLiA and ScientISST MOVE, demonstrate that deep feature representations yield superior performance compared to handcrafted features. Building upon these findings, we propose four LSTM-based hybrid deep learning architectures—CNN–BiLSTM, CNN–LSTM, CNN–BiLSTM with Attention, and Multi-branch CNN–BiLSTM—that effectively combine local feature extraction and temporal dependency modeling. Experimental results show that the proposed models achieve 1–4% higher accuracy than baseline deep learning and hybrid methods, with CNN–BiLSTM with Attention consistently achieving the best performance. These results highlight the potential of deep feature embeddings and hybrid sequence models for robust PPG-based biometric identification in wearable applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Efficient Deep Feature Embedding and Hybrid CNN–LSTM Models for Wearable PPG Biometric Recognition

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Truc T. T. Trinh,
  • Dinh Thuan Nguyen

摘要

Photoplethysmography (PPG) signals are increasingly being integrated into wearable devices for health monitoring due to their continuous, non-invasive, and low-cost nature. Beyond physiological assessment, PPG has emerged as a promising biometric modality for user identification, offering potential for personalized healthcare services and enhanced security in wearable environments. This paper presents a survey of PPG-based biometric recognition methods, categorized into traditional machine learning approaches and deep learning approaches. Ablation experiments conducted on two benchmark datasets, PPG-DaLiA and ScientISST MOVE, demonstrate that deep feature representations yield superior performance compared to handcrafted features. Building upon these findings, we propose four LSTM-based hybrid deep learning architectures—CNN–BiLSTM, CNN–LSTM, CNN–BiLSTM with Attention, and Multi-branch CNN–BiLSTM—that effectively combine local feature extraction and temporal dependency modeling. Experimental results show that the proposed models achieve 1–4% higher accuracy than baseline deep learning and hybrid methods, with CNN–BiLSTM with Attention consistently achieving the best performance. These results highlight the potential of deep feature embeddings and hybrid sequence models for robust PPG-based biometric identification in wearable applications.