• deep learning
• explainable ai
• machine learning
• prototyping
• quantization
• sleep staging

Sleep study is a fundamental tool in the medical field for diagnosing sleep disorders, but it presents critical issues related to the duration of the examination and the consequent effort required from the physician for sleep staging, which is the process of classifying sleep epochs into five categories: Wake, NREM1, NREM2, NREM3, and REM. This brings significant costs, which may limit access to this type of study.
To overcome these issues, neural network-based models have been developed, capable of offering performance comparable to that of humans in terms of accuracy while also drastically reducing the time required for staging. The state-of-the-art model at the moment is SeqSleepNet, a hierarchical recurrent neural network that operates sequence-to-sequence. Specifically, it is a network that takes as input a sequence of epochs (in this case, sleep epochs) and produces, in a single step, the classification of those epochs. However, the use of such neural networks in the medical field is limited by their "black box" nature (the issue where, despite the model making accurate predictions, we are unable to understand exactly what led it to make those decisions), which compromises their interpretability and, consequently, their clinical usability.
This issue has been addressed in this thesis through a double quantization approach. The first step is to understand which portions of the sleep epoch the model considers most important. To achieve this, it is forced to not consider the sleep epoch in its entirety and focus instead only on a limited number of reduced-size sections obtained from it. However, not all of them are significant for the sleep staging process, and the first quantization is performed in this context. The model is forced to select the most significant sections, discarding those that provide little information for classification purposes and retaining only those that contain meaningful ones.
The second quantization occurs using prototypes, which allow the selected sections to be represented in a more compact and interpretable space. Each significant section is associated with a prototype, a simplified representation that summarizes the main information.
The ability to trace the original signal back to a limited set of prototypes makes the classification process interpretable. This is because an expert can associate the prototypes with salient characteristics of the signal, thereby understanding the link between the presence of a specific prototype and the resulting classification.
The results obtained confirm the validity of the proposed model, achieving accuracy comparable to the previously described black-box models