Short Summary/Abstract:

Schizophrenia is a common and devastating illness that affects 1 in 100 people. It has an early onset and results in substantial disability. Despite this, the cause of schizophrenia is still not known, there are no objective biomarkers and current interventions do not reduce this disability. For the first time we will apply mechanistic interpretability techniques to multimodal deep learning models trained on linked structural MRI (sMRI) and clinical data to better understand the develiopment of psychotic disorders.  We will train multimodal deep learning models on linked sMRI and clinical data and evaluate their performance at classifying a subject into either a healthy control (HC), at clinical high risk of psychosis (CHR) or having the diagnosis of First Episode Psyhcosis (FEP). We will then use mechanistic interpretability techniques to identify the fundamental component's of the highest performing neural networks and reverse engineer them into human-interpretable algorithms. In doing so we can understand why the models made their predictions and hope to uncover novel imaging biomarkers, risk and/ or protective factors contributing to the development of psychosis.