Wearable PPG-Based Atrial Fibrillation Detection Using a Hybrid CNN-LSTM Model: Performance Evaluation Across Signal Quality and Noise Conditions

Author(s): Priyanka Joshi, Rohit Bansal, Megha Tiwari

Affiliation: Department of Electronics and Communication Engineering, Shri Vaishnav Institute of Technology and Science, Indore, Madhya Pradesh, India

Page No: 14-20

Volume, Issue & Publishing Year: Volume 3 Issue 3,Aug-2026

Journal: International Journal of Advanced Engineering Application (IJAEA)

ISSN NO: 3048-6807

Abstract:
Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia and a leading independent risk factor for ischaemic stroke, yet its frequently paroxysmal and asymptomatic presentation makes opportunistic, continuous screening through low-cost wearable photoplethysmography (PPG) sensors an attractive complement to intermittent clinical electrocardiography. The principal barrier to reliable wearable-based AF detection is motion-artefact and ambient-noise corruption of the PPG signal, which degrades the rhythm-irregularity features that distinguish AF from normal sinus rhythm and necessitates explicit signal-quality assessment ahead of classification. This study develops and evaluates a hybrid convolutional neural network-long short-term memory (CNN-LSTM) architecture for AF detection from single-channel wrist-worn PPG, incorporating a signal quality index (SQI)-based segment rejection stage trained and validated on a dataset of 41,600 thirty-second PPG segments (8,360 unique recording sessions) collected from 214 subjects, of which 3,680 segments were retained as AF-positive following cardiologist-verified ECG-synchronised labelling. The proposed CNN-LSTM model is benchmarked against a Random Forest baseline using handcrafted heart-rate-variability features, a standalone CNN, and a standalone LSTM. The CNN-LSTM model achieves the highest test-set performance with an accuracy of 94.5%, sensitivity of 91.5%, specificity of 95.7%, F1-score of 90.3%, and area under the receiver operating characteristic curve (AUC) of 0.978, outperforming the Random Forest baseline (AUC 0.912) by a statistically significant margin. Stratified performance analysis across signal-to-noise ratio (SNR) bins reveals that sensitivity falls from 97.8% at SNR greater than 15 dB to 58.2% at SNR below 0 dB, underscoring the necessity of the SQI-based rejection stage, which excludes 480 of 4,160 (11.5%) candidate segments per session-equivalent batch as unsuitable for reliable classification. Power spectral density and RR-interval Poincaré analysis confirm that the model's discriminative capacity derives from the characteristic beat-to-beat irregularity and dominant-frequency dispersion that distinguish AF from sinus rhythm. The results support the feasibility of CNN-LSTM-based PPG screening as a pre-clinical triage tool for opportunistic AF detection in continuous wearable monitoring contexts, contingent on robust signal-quality gating to maintain diagnostic reliability under real-world motion and noise conditions.

Keywords: atrial fibrillation, photoplethysmography, wearable sensors, CNN-LSTM, deep learning, signal quality index, heart rate variability, arrhythmia classification, ROC analysis

Reference:

• [1] Aliamiri, A., & Shen, Y. (2018). Deep learning based atrial fibrillation detection using wearable photoplethysmography sensor. IEEE EMBS International Conference on Biomedical & Health Informatics, 442-445.
• [2] Bashar, S. K., Han, D., Hajeb-Mohammadalipour, S., Ding, E., Whitcomb, C., McManus, D. D., & Chon, K. H. (2019). Atrial fibrillation detection from wrist photoplethysmography signals using smartwatches. Scientific Reports, 9, 15054.
• [3] Chiang, H. T., Hsieh, Y. Y., Fu, S. W., Hung, K. H., Tsao, Y., & Chien, S. Y. (2019). Noise reduction in ECG signals using fully convolutional denoising autoencoders. IEEE Access, 7, 60806-60813.
• [4] Eerikainen, L. M., Bonomi, A. G., Schipper, F., Dekker, L. R. C., Vullings, R., de Morree, H. M., & Aarts, R. M. (2018). Comparison between electrocardiogram- and photoplethysmogram-derived features for atrial fibrillation detection in free-living conditions. Physiological Measurement, 39(8), 084001.
• [5] Elgendi, M. (2012). Optimal signal quality index for photoplethysmogram signals. Bioengineering, 3(4), 25-31.
• [6] Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
• [7] Hwang, H. M., Lee, J. Y., & Joo, S. (2020). A deep learning approach for detecting atrial fibrillation using wrist-worn photoplethysmography. Sensors, 20(15), 4274.
• [8] Lubitz, S. A., Faranesh, A. Z., Selvaggi, C., Atlas, S. J., McManus, D. D., Singer, D. E., & Lee, S. (2022). Detection of atrial fibrillation in a large population using wearable devices. Circulation, 146(19), 1415-1424.
• [9] Perez, M. V., Mahaffey, K. W., Hedlin, H., Rumsfeld, J. S., Garcia, A., Ferris, T., & Turakhia, M. P. (2019). Large-scale assessment of a smartwatch to identify atrial fibrillation. New England Journal of Medicine, 381(20), 1909-1917.
• [10] Poh, M. Z., Poh, Y. C., Chan, P. H., Wong, C. K., Pun, L., Leung, W. W. C., & Siu, C. W. (2018). Diagnostic assessment of a deep learning system for detecting atrial fibrillation in pulse waveforms. Heart, 104(23), 1921-1928.
• [11] Selvaraj, N., Jaryal, A., Santhosh, J., Deepak, K. K., & Anand, S. (2008). Assessment of heart rate variability derived from finger-tip photoplethysmography. Journal of Medical Engineering & Technology, 32(6), 479-484.
• [12] Shen, Y., Voisin, M., Aliamiri, A., Avati, A., Hannun, A., & Ng, A. (2019). Ambulatory atrial fibrillation monitoring using wearable photoplethysmography with deep learning. Proceedings of the 25th ACM SIGKDD International Conference, 1909-1916.
• [13] Sweeting, M. J., et al. (2021). Pulse waveform analysis for atrial fibrillation screening using machine learning: a systematic review. International Journal of Cardiology, 326, 152-159.
• [14] Tison, G. H., Sanchez, J. M., Ballinger, B., Singh, A., Olgin, J. E., Pletcher, M. J., & Marcus, G. M. (2018). Passive detection of atrial fibrillation using a commercially available smartwatch. JAMA Cardiology, 3(5), 409-416.