Seite - 340 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

The last classifier in the stacking scheme is the recording-based classifier. The features for the recording-based classifier are calculated using mean aggregation over the features of all segments in the recording. In addition, the average class probabilities for each segment, estimated by the segment-based classifier, are appended as features. Since the recording-based classifier is a meta-learner that utilizes the output of the segment-based classifier, a holdout set is required for its training. For that reason, we used a leave-one-subject-out (LOSO) technique on the training data, to generate the meta-training dataset. All models were trained using the Random Forest (RF) algorithm. We used the RF algorithm because of its robustness to noise in the input features. 4. Experimental results We evaluated the proposed approach in two experimental setups: classifying CHF patients (both decompensated and recompensated) vs. healthy, and classifying decompensated vs. recompensated recordings. LOSO cross-validation was used for the two experimental setups. The first experimental setup simulates a situation where the algorithm detects whether a new user shows symptoms of CHF. The second experimental setup is aimed at detecting worsening of the condition in patients that have already been diagnosed with CHF, and represents a first step in building a method that would be able to notify the user to seek medical help before a hospitalization is required. The results are presented in the following subsections. 4.1. CHF patients vs. healthy analysis The results of the LOSO evaluation are presented in Table 2. The table contains the results achieved by the proposed Stacking classifier and by the Segment classifier. The recording predictions for the Segment classifier are calculated by taking the majority prediction for all segments in one recording. The first two rows in Table 2 present the confusion matrices for the Stacking classifier and for the Segment classifier. The entries in the confusion-matrices represent recordings. The rest of the rows in Table 2 present the performance metrics. The majority class is 68 % i.e., 68 % of the recordings belong to healthy individuals. The Stacking approach has achieved the accuracy of 82 %, which is 14 percentage points higher that the majority class and 6 percentage points higher than the Segment-based approach. In addition, compared to the Segment-based approach, the Stacking approach has significantly higher precision, recall, and F1-score. Table 2. Confusion matrices and performance measures (recall, precision, F1-score and accuracy) Stacking Segment-based Patient Healthy Patient Healthy Patient 49 29 30 47 Healthy 14 151 17 148 Recall 0.63 0.92 0.39 0.90 Precision 0.78 0.84 0.64 0.76 F1-score 0.70 0.88 0.48 0.82 Accuracy 0.82 0.74 Majority 0.68 M.Gjoreski etal. /TowardEarlyDetectionandMonitoringofChronicHeartFailure340