• automatic sleep staging
• clinical applications
• concept annotations
• concept-based deep learning models
• explainability
• prototypical networks
• self-supervised learning
• sleep patterns
• transparency
•  concept labeling
•  interpretability
•  machine learning models

Understanding sleep patterns plays a vital role in automatic sleep staging and diagnosing sleep-related disorders. While current Machine Learning models excel in classifying sleep stages, they often operate as black-box systems, lacking the interpretability required for clinical applications. In medical settings, the models' transparency in decision-making is as important as prediction accuracy. To address this challenge, we explore the integration of Concept-Based Deep Learning Models, a class of approaches designed to enhance model interpretability by identifying and utilizing user-defined concepts that guide final predictions. These models provide a level of reasoning that could helps clinicians understand model outputs. In this work, we aim to improve the interpretability of automatic sleep staging systems by incorporating Concept-Based Models, ensuring that high classification accuracy is maintained while offering clinicians meaningful insights into the decision-making process.
The reason why similar studies are not present in the literature is the lack of datasets with annotated concepts that support the training of Concept-Based Models. We propose several methods to automate learning of concepts. The idea is to capture time-frequency characteristics in physiological signals that are useful for sleep staging and assign them to the corresponding dataset records. The most promising approach draws inspiration from Prototypical Networks and Semi-Supervised Learning, providing a way to automatically create concept annotations.

In this way, we were able to extend existing datasets and train Concept-Based Models on them. The models achieve accuracy on the sleep stage classification task comparable to that of models focused only on classification, but they are also able to predict concept activations with an average error of approximately 0.05 (on activation values ranging from 0 to 1), thereby enhancing the interpretability.