Automated Identification of Tachyarrhythmia from Different Datasets of Heart Rate Variability Using a Hybrid Deep Learning Model

Machine learning workflow for edge computed arrhythmia detection in exploration class missions

“…Their work could perhaps provide inspiration for a next stage of our research since their model tracked 12 arrhythmia pathologies thanks to the size of their proprietary dataset and the implicit control they have over it resulting in few rhythms with F1 scores under 0.9. 30,59,60,[62][63][64][65] Given that the macro F1 score of this work is above most values reported in the literature (cf. Table 4), we can also conclude that this overall approach which leverages methods in nested cross-validation loops, denoising and lightweight optimized edge inference through the ONNX format-a deployment process synthesized in Fig.…”

“…On the same dataset, Plesinger et al 60 and Kamaleswaran et al [61][62][63][64][65] both used a Convolutional Neural Network approach but the latter with a Bagged Tree Ensemble approach, which respectively provided macro F1 scores of 0.70 and 0.81. Ojha et al 30 used a DL Long Short-Term Memory (LSTM) approach to obtain a macro F1 score of 0.927, slightly above the performance of our model. However, ML models typically require less computing power than DL models 66 , thus our approach stays competitive in our edge computing context where slightly better inference cannot be at a much higher computing cost.…”

“…For example, while deep learning (DL) has shown promising results in detecting tachycardia in ECG signals 30 , it often lacks this crucial transparency in its categorization processes, frequently functioning as a 'black box' despite recent advances in interpretability 31 . In our context, the goal was not to discover new relationships through black box algorithms, but rather to enable the inference of tachycardia from interpretable features whose relationship could be visualized and understood by the CMO for reasons outlined above.…”