Validação do PPG de dispositivos vestíveis para estimação e análise da variabilidade da frequência cardíaca

Abstract

HRV is a time series formed by the intervals between heartbeats, from which metrics are calculated for the evaluation of the cardiovascular system. Furthermore, it contributes to the detection of various anomalies in the body, such as arrhythmias and heart failure. The gold standard for estimating HRV metrics is the 12-lead ECG, a test that requires the active participation of the patient and professional supervision for its execution. However, PPG has emerged as a viable alternative to the ECG due to its advantages, such as easier and passive data collection. This aspect is especially relevant in the context of mobile devices, which enable continuous monitoring of the signal, making the analysis more accessible and integrated into daily life, as seen with smartwatches. Thus, this study presents an investigation into the feasibility of using mobile device PPG for HR estimation and HRV analysis. The study was conducted using both manual annotation of PPG peaks and annotation performed by algorithms. The validation of peak detection by the algorithms was carried out using metrics such as sensitivity, PPV, and EDR, comparing the results with manual annotation. Additionally, the heart rate estimated from simultaneously collected ECG and PPG was compared using RMSE. Finally, HRV metrics obtained from PPG were compared to those from ECG, employing metrics such as Spearman’s correlation coefficient, NRMSE, Pearson’s correlation coefficient(R), and determination coefficient (R2 ). The results indicated the feasibility of using PPG for HR estimation, with an RMSE of 0.12 bpm for 1-minute signals and 2.49 bpm and 2.89 bpm for 5-minute signals, corresponding to the LED infrared and LED red sensors, respectively, of the Samsung Galaxy Watch 5. For HRV analysis, the manually annotated 1-minute signals showed a strong correlation with ECG, with Spearman’s coefficients above 0.80 and linear correlation above 0.90 in all three domains, as well as mean NRMSE values of 2.59, 3.93, and 2.83 for the time, frequency, and nonlinear domains, respectively. The algorithms also achieved consistent results for 1-minute signals, with RMSE below 1 bpm and correlations close to those of manual annotation. For 5-minute signals, the correlation between metrics remained high, but NRMSE increased, with means of 8.10, 13.43, and 12.53 for the infrared sensor and 10.27, 13.54, and 16.14 for the red sensor.