A Comprehensive Stacked Ensemble Framework for ECG-Based Stress Level Classification: Integrating Base Classifiers with Meta-Learning and Uncertainty Measurement

Effective and quick medical intervention is crucial for the early detection of mental diseases, such as stress, which are among the top causes of death worldwide. For estimating stress levels, the traditional machine learning method is laborious and complicated. To address this issue, we propose a hybrid three-stage machine learning-based methodology for predicting the stress level. This hybrid model evaluates on a dataset created by the acquisition of raw ECG signals from ten healthy subjects both in relaxed and stressed conditions. The raw signal is pre-processed by a combined filter. Total 14 valuable features extracted from the clean ECG signal. This model aggregates two well known feature selection methods of Minimal-redundancy-maximal-relevance (MRMR) and Neighborhood Component Analysis (NCA), to select the best feature subset. We select the probable machine learning algorithm (parametric or non-parametric) by correlation and normality analysis of the features set. A novel three-stagebased machine learning approach has been proposed which utilizes the strength of base classifiers (KNN, SVM, DT, RF, AdaBoost) at the first stage. Bootstrap sample generates using the original dataset and do training with the base classifiers hard voted first stage matrix. Both Random forest and Support vector machine algorithms has been used to classify the stress level. The proposed model shows promising performance as it achieved a test accuracy of $93.34 \%$ when Random forest is used as a Meta classifier. This result shows how well integrating the advantages of several basic classifiers can work. Finally, to evaluate the generalizability and reliability of the proposed model, uncertainty evaluation was conducted. Moreover, this will motivate additional AI researchers to investigate alternative approaches for stress detection.