AI-based multimodal treatment response prediction in schizophrenia: combining inflammation and imaging

Background:
The most dominant hypothesis for the cause of schizophrenia is excess of striatal dopamine, as suggested by the effectiveness of antipsychotic medications which act to block dopamine, mainly at the D2 receptor. However, approximately one third of patients fail to respond to D2 antagonists, whilst exhibiting similar antipsychotic dopamine receptor binding as responsive patients. Therefore, the non-responsive cohort may have illness driven by a different pathophysiological mechanism. One suggested theory is that non-responsive patients have more prominent immune system dysregulation than antipsychotic responders and that this may be related to more pronounced structural brain changes. The aim of this project was to develop Support Vector Machine (SVM) learning models of treatment response based on inflammatory markers and structural imaging data and examine their predictive utility in an external sample.

Method:
We analysed data from the Schizophrenia: Treatment Resistance and Therapeutic Advances (STRATA) consortium and the Benefit of Minocycline on Negative Symptoms of Psychosis: Extent and Mechanism (BeneMin) datasets as part of the Psychosis Immune Stratified Medicine (PIMS) collaboration. Individuals were categorised as treatment responders if they met the following criteria: (i) they had been treated with a single antipsychotic medication since the onset of their condition, or any changes in medication were solely due to side effects rather than lack of effectiveness; (ii) they achieved a CGI-SCH score lower than 4; (iii) their PANSS total score was below 60; (iv) they had a compliance rating higher than 3. Conversely, the criteria for being classified as a treatment non-responder included (i) undergoing treatment with at least two different antipsychotics for over four weeks each, at doses above the minimum therapeutic level as specified by the British National Formulary; (ii) a CGI-SCH score of more than 3; (iii) a PANSS total score of 70 or higher; (iv) a compliance rating score higher than 3. An SVM model that classified treatment response vs non-treatment response was developed using the STRATA) dataset. The model was trained in a repeated nested pooled cross-validation framework with 5 outer CV2 permutations, 5 outer CV2 folds, 5 inner CV1 permutations, and 5 inner CV1 folds. The model was then externally validated to the BeneMin dataset in order to test its predictive utility.

Results:
Combining inflammatory and imaging data the SVM model classified treatment response vs non-treatment response with a balanced accuracy of 70.8 and an area under the curve of 0.71. Patients classified in the treatment non-response group in the external validation sample showed no statistically significant differences at baseline. However, 12 months after baseline non-responders exhibited higher PANSS positive, PANSS general, and PANSS total scores.

Conclusion:
This is, to our knowledge, the first study to date to incorporate inflammation and neuroimaging data to predict treatment response in schizophrenia. Our models may be able to aid the classification of treatment response in patients with schizophrenia. Our findings contribute to a growing body of evidence suggesting that immune system dysregulation may play a role in the pathophysiological mechanisms of treatment non-response in schizophrenia, with insight for targeted approaches in the future.