Regression-Based Estimation of Wearable Sensor Signal Quality Using Statistical and Spectral Features

Abstract

Physiological signals collected from wearable sensors are usually prone to noise due to motion artifacts, sensor displacement, and other sources. Since this signal degradation might cause problems in signal reliability, signal quality needs to be checked and determined. Although machine learning techniques have been successfully applied to signal quality assessment, the dominant features and their effects on different physiological modalities still need to be investigated. In this study, I performed an analysis on modality-specific signal degradation mechanisms in electrocardiography, electrodermal activity, and respiration signals using interpretable machine learning techniques including Random Forest, eXteme Gradient Boosting, and Light Gradient Boosting Machine. The models were trained using Leave-One-Subject-Out Cross-Validation to estimate the signal quality levels. In addition, feature importance analysis and SHapley Additive exPlanations were employed to investigate model interpretability. The results showed high prediction performance across all physiological modalities. EDA signals performed the highest coefficient of determination (R2 = 0.996) followed by respiration (R2 = 0.954) and ECG (R2 = 0.908). Feature importance and SHAP analysis were also confirmed these results, in addition they reflected the fact that the most dominant feature differs for EDA and ECG/respiration signals. These findings demonstrated that physiological signal degradation exhibited distinct modality-specific characteristics and interpretable machine learning techniques could provide valuable insight into the mechanisms of signal degradation.

Keywords

• machine learning
• regression
• signal quality
• wearable sensors
• biomedical signals

İstatistiksel ve Spektral Özellikler Kullanılarak Giyilebilir Sensör Sinyal Kalitesinin Regresyon Tabanlı Tahmini

Abstract

Giyilebilir sensörlerden toplanan fizyolojik sinyaller genellikle hareket artefaktları, sensör yer değiştirmesi ve diğer kaynaklar nedeniyle gürültüye eğilimlidir. Bu sinyal bozulması sinyal güvenilirliğinde sorunlara yol açabileceğinden, sinyal kalitesinin kontrol edilmesi ve belirlenmesi gerekmektedir. Makine öğrenme teknikleri sinyal kalitesi değerlendirmesinde başarıyla uygulanmış olsa da, baskın özellikler ve bunların farklı fizyolojik modaliteler üzerindeki etkileri hala araştırılmaya ihtiyaç duymaktadır. Bu çalışmada, yorumlanabilir makine öğrenme teknikleri (Random Forest, eXteme Gradient Boosting ve Light Gradient Boosting Machine) kullanarak elektrokardiyografi, elektrodermal aktivite ve solunum sinyallerindeki modaliteye özgü sinyal bozulma mekanizmalarını analiz ettim. Modeller, sinyal kalitesi seviyelerini tahmin etmek için Leave-One-Subject-Out Cross-Validation yöntemi kullanılarak eğitildi. Ek olarak, model yorumlanabilirliğini araştırmak için özellik önem analizi ve SHapley Additive exPlanations kullanıldı. Sonuçlar, tüm fizyolojik modalitelerde yüksek tahmin performansı gösterdi. EDA sinyalleri en yüksek belirleme katsayısını (R2 = 0,996) gösterirken, bunu solunum (R2 = 0,954) ve EKG (R2 = 0,908) izledi. Özellik önemi ve SHAP analizi de bu sonuçları doğruladı ve ayrıca en baskın özelliğin EDA ve EKG/solunum sinyalleri için farklı olduğunu yansıttı. Bu bulgular, fizyolojik sinyal bozulmasının farklı modaliteye özgü özellikler sergilediğini ve yorumlanabilir makine öğrenme tekniklerinin sinyal bozulma mekanizmalarına değerli bilgiler sağlayabileceğini göstermiştir.

Keywords

• biyomedikal sinyaller
• regresyon
• sinyal kalitesi
• giyilebilir sensörler

References

1. Y. Jiang, C. Spies, J. Magin, S. J. Bhosai, L. Snyder, and J. Dunn, “Investigating the accuracy of blood oxygen saturation measurements in common consumer smartwatches,” PLOS Digital Health, vol. 2, no. 7, p. e0000296, Jul. 2023, doi: 10.1371/JOURNAL.PDIG.0000296.
2. A. Anma and M. Komori, “The Efficacy of a Wearable Device-based Vital Sign Monitoring System in a Japanese Home Medical Care Clinic,” Internal Medicine, vol. 64, no. 24, pp. 3502–3508, Dec. 2025, doi: 10.2169/INTERNALMEDICINE.5213-24.
3. M. Joshi et al., “Performance of Continuous Digital Monitoring of Vital Signs with a Wearable Sensor in Acute Hospital Settings,” Sensors 2025, Vol. 25, Page 2644, vol. 25, no. 9, p. 2644, Apr. 2025, doi: 10.3390/S25092644.
4. M. J. M. Breteler et al., “Reliability of an all-in-one wearable sensor for continuous vital signs monitoring in high-risk patients: the NIGHTINGALE clinical validation study,” Journal of Clinical Monitoring and Computing 2025 39:5, vol. 39, no. 5, pp. 1087–1100, Mar. 2025, doi: 10.1007/S10877-025-01279-X.
5. A. Angelucci, M. Greco, M. Cecconi, and A. Aliverti, “Wearable devices for patient monitoring in the intensive care unit,” Intensive Care Medicine Experimental 2025 13:1, vol. 13, no. 1, pp. 26-, Feb. 2025, doi: 10.1186/S40635-025-00738-8.
6. S. Fallet et al., “Assessing ECG signal quality indices to discriminate ECGs with artefacts from pathologically different arrhythmic ECGs,” Physiol. Meas., vol. 37, no. 8, p. 1370, Jul. 2016, doi: 10.1088/0967-3334/37/8/1370.
7. R. He et al., “Reducing false arrhythmia alarms in the ICU using novel signal quality indices assessment method,” Comput. Cardiol. (2010)., vol. 42, pp. 1189–1192, Feb. 2015, doi: 10.1109/CIC.2015.7411129.
8. S. Fallet et al., “Assessing ECG signal quality indices to discriminate ECGs with artefacts from pathologically different arrhythmic ECGs,” Physiol. Meas., vol. 37, no. 8, p. 1370, Jul. 2016, doi: 10.1088/0967-3334/37/8/1370.

9. C. Orphanidou, “Signal Quality Assessment in Physiological Monitoring: Requirements, Practices and Future Directions,” pp. 1–14, 2018, doi: 10.1007/978-3-319-68415-4_1.
10. A. Ignácz, S. Földi, P. Sótonyi, and G. Cserey, “NB-SQI: A novel non-binary signal quality index for continuous blood pressure waveforms,” Biomed. Signal Process. Control, vol. 70, p. 103035, Sep. 2021, doi: 10.1016/J.BSPC.2021.103035.
11. M. Elgendi, “Optimal Signal Quality Index for Photoplethysmogram Signals,” Bioengineering 2016, Vol. 3, Page 21, vol. 3, no. 4, p. 21, Sep. 2016, doi: 10.3390/BIOENGINEERING3040021.
12. A. Mondal, M. S. Manikandan, and R. B. Pachori, “Automatic ECG signal quality assessment using convolutional neural networks and derivative ECG signal for false alarm reduction in wearable vital signs monitoring devices,” Biomed. Signal Process. Control, vol. 108, p. 107876, Oct. 2025, doi: 10.1016/J.BSPC.2025.107876.
13. Y. ; Liu et al., “An Automatic ECG Signal Quality Assessment Method Based on Resnet and Self-Attention,” Applied Sciences 2023, Vol. 13, Page 1313, vol. 13, no. 3, p. 1313, Jan. 2023, doi: 10.3390/APP13031313.
14. F. Mohagheghian et al., “Optimized Signal Quality Assessment for Photoplethysmogram Signals Using Feature Selection,” IEEE Trans. Biomed. Eng., vol. 69, no. 9, pp. 2982–2993, Sep. 2022, doi: 10.1109/TBME.2022.3158582.
15. E. Idrobo-Avila, G. Bognar, D. Krefting, T. Penzel, P. Kovacs, and N. Spicher, “Quantifying the Suitability of Biosignals Acquired During Surgery for Multimodal Analysis,” IEEE Open J. Eng. Med. Biol., vol. 5, pp. 250–260, 2024, doi: 10.1109/OJEMB.2024.3379733.
16. Y. Jiang et al., “An Optimized Signal Quality Assessment Method for Noncontact Capacitive ECG,” IEEE Trans. Instrum. Meas., vol. 74, 2025, doi: 10.1109/TIM.2025.3533644.
17. P. Schmidt, A. Reiss, R. Duerichen, and K. Van Laerhoven, “Introducing WeSAD, a multimodal dataset for wearable stress and affect detection,” ICMI 2018 - Proceedings of the 2018 International Conference on Multimodal Interaction, pp. 400–408, Oct. 2018, doi: 10.1145/3242969.3242985.
18. J. Stritzelberger et al., “Validity of Empatica E4 Wristband for Detection of Autonomic Dysfunction Compared to Established Laboratory Testing,” Diagnostics, vol. 15, no. 20, p. 2604, Oct. 2025, doi: 10.3390/DIAGNOSTICS15202604.
19. A. Simons, T. Doyle, D. Musson, and J. Reilly, “Impact of Physiological Sensor Variance on Machine Learning Algorithms,” Conf. Proc. IEEE Int. Conf. Syst. Man Cybern., vol. 2020-October, pp. 241–247, Oct. 2020, doi: 10.1109/SMC42975.2020.9282912.
20. L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, Oct. 2001, doi: 10.1023/A:1010933404324.
21. “XGBoost.” Accessed: Jun. 10, 2026. [Online]. Available: https://xgboost.ai/
22. “Welcome to LightGBM’s documentation! — LightGBM 4.6.0 documentation.” Accessed: Jun. 10, 2026. [Online]. Available: https://lightgbm.readthedocs.io/en/stable/
23. “LeaveOneOut — scikit-learn 1.9.0 documentation.” Accessed: Jun. 10, 2026. [Online]. Available: https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.LeaveOneOut.html
24. B. Rosner, R. J. Glynn, and M. L. T. Lee, “The Wilcoxon signed rank test for paired comparisons of clustered data,” Biometrics, vol. 62, no. 1, pp. 185–192, 2006, doi: 10.1111/J.1541-0420.2005.00389.X.
25. “Welcome to the SHAP documentation — SHAP latest documentation.” Accessed: Jun. 10, 2026. [Online]. Available: https://shap.readthedocs.io/en/latest/