O four) has shown to enhance the performance of our cross-subject modelO 4) has shown

O four) has shown to enhance the performance of our cross-subject model
O 4) has shown to improve the performance of our cross-subject model by three (F1-score), in comparison with using exclusively the 3D-ACC signal. When once again, a fusion of PPG and 3D-ACC ML-SA1 Epigenetics signals has shown to not yield performance improvements (Scenario five) to educated 3D-ACC models. Interestingly, combining each ECG and PPG (Scenario six) yields a model with much better overall performance than the models educated exclusively with ECG (Scenario two) and PPG (Situation three), even if it nonetheless underperforms against the purely 3D-ACC educated model (Situation 1). In the end, we conclude that the ECG signal can be complement nicely the 3D-ACC signal in HAR systems, though PPG did not offer informative data to our cross-subject models.Sensors 2021, 21,15 ofThree signals fusion. As soon as we fuse all 3 sources of signals, we BSJ-01-175 Protocol observe a lower within the model’s performance in comparison to just combining ACC-3D and ECG signals (Situation four). This further corroborates using the notion that adding the PPG signal for the combination of 3D-ACC and ECG signals is disadvantageous. Per activity overall performance. Figure eight shows the overall performance of your models broken down per activity. First, we note that “cycling”, “table-soccer” and “sitting” activities remain rather stable in all models trained with 3D-ACC, exclusively or in mixture. Second, cross-subject models generally miss-classify “stairs” and “walk” activities. After we fuse each the 3D-ACC and ECG signal, the model is superior capable to distinguish between the two activities, which explains the gain in the overall overall performance with the model. Third, models educated exclusively with bio-signals have really distinct functionality profiles per activity. Note that PPG is reasonably good at distinguishing the “sitting” activity, as this really is the least physically demanding activity in our dataset (decrease heart rate). ECG models, on the other hand, outperform PPG models in all other activities, further corroborating that the ECG signal is, on typical, much more informative for cross-subject HAR models than the PPG signal.Figure 7. Cross-subject model final results.Figure eight. Cross-subject model results per activity.Sensors 2021, 21,16 of6. Discussion To elaborate a lot more around the impact on the ECG signal in the efficiency of HAR models, we dive in to the confusion matrices associated to all subjects in each subject-specific and cross-subject models. 6.1. Subject-Specific As stated earlier, considering only 3D-ACC signals, models already reach a higher recognition performance of 94.07 F1-score. Therefore, the majority of the instances in confusion matrices are labeled accurately. On the other hand, the activities of using “Stair” and “Walking” were often confused with each other. For that reason, we contrast the confusion matrices of the model which involve only 3D-ACC (Situation 1) against the models which includes both 3DACC and ECG signals (Scenario four). We observe important improvement in distinguishing the pointed out activities following adding the ECG signal. Figure 9 presents confusion matrices related to subject number 8 inside the subject-specific model. On the left side of Figure 9, we can observe the model efficiency when considering only 3D-ACC. Note the instances which are miss-classified and confused between “Stairs” and “Walking” activities. On the right side of Figure 9, even so, it is clear that immediately after adding the ECG signal, the pointed out confusions are solved.Figure 9. Comparison in between confusion matrices in subject-specific models. On the left: the model overall performance when contemplating only 3D-ACC. On the proper: th.