Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session A: Tuesday, August 12, 1:30 – 4:30 pm, de Brug & E‑Hall

Enhancing OCD classification with Transformer-based deep learning on resting-state fMRI: insights from the ENIGMA-OCD cohort and UK Biobank pretraining

Maria Pak¹, Youngchan Ryu¹, Sangyoon Bae¹, Willem B. Bruin, Guido Van Wingen², Odile A. van den Heuvel, Jiook Cha¹; ¹Seoul National University, ²Amsterdam UMC

Presenter: Maria Pak

Obsessive-compulsive disorder (OCD) remains challenging to classify due to its heterogeneous clinical presentation and the limitations of static brain connectivity metrics. To address these hurdles, we applied a Transformer-based deep learning model to a record-sized dataset of resting-state fMRI from 2,094 individuals in the ENIGMA-OCD consortium. By pretraining on an extensive UK Biobank dataset and using dynamic connectivity measures across multiple frequency bands, our approach achieved higher predictive performance for OCD than conventional methods. We further conducted uncertainty quantification, revealing a marked reduction in calibration error for the pretrained model. Finally, self-attention-based interpretation pinpointed reduced connectivity within sensorimotor networks in patients with OCD, consistent with prior literature. These findings underscore the value of large-scale pretraining and dynamic rs-fMRI data in enhancing model generalizability, highlighting a promising avenue for more robust OCD classification and, by extension, clinical decision-making.

Topic Area: Memory, Spatial Cognition & Skill Learning

Extended Abstract: Full Text PDF