Accurate Heart Beat Detection with Doppler Radar using Bidirectional GRU Network

Machine Learning in RADAR-based Physiological Signals Sensing: A Scoping Review of the Models, Datasets and Metrics

“…3) Physiological event detection: In addition to extracting simple vital signs within a certain observation window, the ability to extract the onset and interval of specific events related to the physiology of breathing and heartbeats allows for a more detailed assessment of the patient's condition. With R-peak detection, for example, as proposed in [51], [80], we can not only obtain the frequency with which an event occurs, but also its variability over time, which may provide a way to detect imbalances in the regulation of the autonomic nervous system [105], [106]. Another example of a clinically relevant task is electrocardiogram segmentation, since the ability to extract the onset and offset of the ECG major waves, as proposed in [95], opens up new possibilities for the diagnosis of short-QT and long-QT syndromes in a completely unobstrusive way.…”

“…signal reconstruction (n=3) [30], [68], [94] accuracy= T P +T N T P +F P +T N +F N physiological event detection (n=3) [59], [78], [79] F1-score= 2×T P 2×T P +F P +F N physiological event detection (n=4) [78]- [80], [95] the ground truth signal and the reconstructed one [63], [70]. With regard to physiological event detection, two types of metrics can be distinguished: time error metrics applied to vital signs or event detection and classification metrics, such as accuracy, f1 score, recall and precision, defined on the basis of a tolerance for predicting a true positive around the event to be detected.…”

“…Regarding classification metrics, based on a defined tolerance around a physiological event, we can define a true positive when the prediction is within the tolerance, a false positive when the prediction is outside the tolerance, and a false negative when there are no predictions within the tolerance. The employed tolerances change depending on the event to be detected with a tolerance of 125 ms for systolic segments classification [81], a tolerance of 70 ms [79] or 75 ms [78] in heart sound detection, a tolerance of 75 ms for R-peak detection [80] and a tolerance of 10 ms for ECG segmentation (QR, RS, ST, TQ) [95].…”

“…The standard then proceeds to define Sensitivity (Se) as TP/(TP+FN) and positive predictive value (+P) as TP/(TP+FP); sensitivity is often called recall and positive predictive value is called precision. In the included papers regarding physiological event detection the researchers usually employ tolerances that are more strict than 150 ms, such as 75 ms for R-peak and heart sounds [78], [80], 70 ms for heart sounds [79] and 10 ms for ECG waveforms segmentation [95]. Moreover, some of the papers include value estimation error metrics to report the distance between estimated vital signs and ground truth vital signs, such as MAE for IBI [51], and also metrics of distance between the predicted event and the actual event such as RMSE for heart sounds prediction [59], MRE for systole/diastole segmentation [81] and MRPE on ECG waveforms [95].…”

“…In the signal reconstruction task all the papers dealing with ECG or cardiac waveform reconstruction employ an ECG as reference sensor [32], [36], [41], [63], [73]. In the physiological event detection task 4 papers employ a public dataset with reliable ECG recordings [51], [59], [80], [95] and other 3 record data from an ECG as ground truth [78], [79], [81]. The lack of reliable reference sensor impairs the positive results of many of the reviewed literature in the context of heart rate estimation.…”