Explainable Deep Learning for Human Behaviour Understanding: Sleep Monitoring, Human Activity Recognition, and Future Opportunities for Healthcare

Abstract

This Ph.D. thesis investigates the transformative potential of explainable deep learning in human behavior analysis, focusing on sleep monitoring and human activity recognition. The research addresses a critical need in contemporary healthcare for AI systems that combine high predictive accuracy with interpretability, thereby enabling clinicians to understand and trust the predictions generated by these models. Central to this thesis is the introduction of novel, state-of-the-art deep learning models designed to categorize behaviors, specifically in accurately classifying sleep stages and recognizing human activities from video data. The thesis presents several significant technical contributions. In the area of sleep monitoring, this work introduces advanced deep learning models, including the time-distributed convolutional neural network (TDConvNet), the one-dimensional convolutional autoencoder (1D-CAE), and the explainable SleepXAI framework. These models are designed to streamline the classification process by utilizing singlechannel physiological signals, thus eliminating the need for manual feature engineering. Furthermore, they incorporate methods such as gradient-weighted class activation mapping (Grad-CAM) to enhance the interpretability of model predictions. These innovations provide a robust and transparent tool for diagnosing and treating sleep disorders, offering insights critical for informed clinical decision-making. This thesis further advances the human activity recognition field by developing two key contributions. The first is the modular deep learning framework for videobased fall detection, which integrates privacy-preserving techniques, selective frame analysis using Short-Time Fourier Transform (STFT), and Grad-CAM for visual explanations. This framework is designed to ensure accuracy and interpretability in real-time fall detection, addressing the critical need for reliable monitoring in elderly care. The second contribution is the development of the Interpretable Feature Reduction Function (IFRF), which significantly enhances the accuracy and transparency of activity classification by optimizing the selection and utilization of features. The resulting models are efficient and capable of providing clear and actionable insights into specific movements and behaviors, which is essential in healthcare applications where precise activity recognition is vital to patient care. The empirical results demonstrate that the models developed in this thesis could substantially improve sleep stage classification and human activity recognition. The sleep monitoring models have the potential to enhance the accuracy of detecting challenging sleep stages while providing interpretable insights that may surpass the capabilities of traditional methods. Similarly, the human activity recognition models, particularly in fall detection and multi-activity recognition, show promise in identifying complex activities, potentially increasing the transparency and reliability of predictions. The broader implications of this research may extend beyond its immediate applications. By prioritizing explainability, this thesis aims to lay the groundwork for the practical deployment of AI in healthcare, offering models that could be seamlessly integrated into medical settings. These advancements have the potential to enhance patient monitoring, improve diagnostic accuracy, and support better clinical outcomes. Moreover, the methodologies introduced in this research are versatile, potentially paving the way for future innovations that could further expand the role of AI across various medical domains. In conclusion, this thesis significantly contributes to the intersection of artificial intelligence and healthcare. It demonstrates how deep learning can be harnessed for its predictive power and its capacity to augment and support human expertise in a transparent and trustworthy manner. This work represents a critical step toward integrating explainable AI into clinical practice, offering tools that have the potential to improve patient care while maintaining the confidence of healthcare professionals in automated decision-making processes.