Unsupervised extraction of knowledge from multimodal biomedical data

Artificial intelligence (AI) is transforming healthcare by improving diagnosis, treatment, and patient outcomes through precision medicine. Indeed, AI shows particular promise in leveraging complex medical data to support clinical decision-making. In this work, we focus on multimodal learning by combining two distinct types of data: tabular and electroencephalography (EEG) data. Specifically, we use cognitive data from neuropsychological tests and questionnaires alongside EEG recordings for pathology classification. Notably, both modalities are informative, quick to collect, and essentially cost-free. To address tabular data, we developed a new clustering algorithm called Repeated Spectral Clustering (RSC), that was applied to several important use cases. First, we clustered stroke patients using the NIHSS, correctly recovering three known clusters (L/R motor and language deficits) and identifying a new one linked to lacunar strokes. Second, we clustered eating disorder patients using psychological tests and demographics. The algorithm recovered two clusters matching DSM-5 diagnoses (bulimia and anorexia) and two others without defined diagnoses but with consistent pathology profiles, supporting the need for a transdiagnostic approach. Finally, the algorithm was extended to longitudinal data. Using a scalar feature, the recovery ratio (based on NIHSS at two time points), we identified six clusters representing patients with different recovery trajectories, ranging from full recovery to deterioration 90 days post-stroke. Concerning EEG data, we first built BIDSAlign, a library designed to preprocess and merge multiple EEG repositories, ensuring sufficient data for deep-learning (DL) experiments. We then evaluated the impact of preprocessing on classification performances. Results showed that extensive pipelines do not necessarily improve performance, while minimal preprocessing can be sufficient. To ensure fair performance estimates, we adopted a Nested Leave-N-Subjects-Out strategy and examined why it provides more reliable generalization than alternative training approaches. Building on these insights, we proposed xEEGNet, a lightweight and fully interpretable neural network inspired by ShallowNet, with 200 fewer parameters and only a 1.5% lower median performance. Finally, using a multimodal dataset, we examined how cognitive and EEG data contribute to distinguishing different forms of dementia by comparing established fusion strategies. Results showed that, although EEG was more informative, combining both modalities consistently improved performance. In conclusion, this work makes several contributions. First, it provides strategies for handling tabular data and heterogeneous modalities, with a particular focus on EEG preprocessing and data partitioning. Second, it introduces two novel methods: the RSC algorithm for tabular data and the xEEGNet architecture for EEG analysis. Finally, it demonstrates that integrating cognitive and EEG data, in a multimodal framework, can enhance classification performance.