Predicting antipsychotic responsiveness using a machine learning classifier trained on plasma levels of inflammatory markers in schizophrenia

Abstract

We apply machine learning techniques to navigate the multifaceted landscape of schizophrenia. Our method entails the development of predictive models, emphasizing peripheral inflammatory biomarkers, which are classified into treatment response subgroups: antipsychotic-responsive, clozapine-responsive, and clozapine-resistant. The cohort comprises 146 schizophrenia patients (49 antipsychotics-responsive, 68 clozapine-responsive, 29 clozapine-resistant) and 49 healthy controls. Protein levels of immune biomarkers were quantified using the Olink Target 96 Inflammation Panel (Olink®, Uppsala, Sweden). To predict labels, a support vector machine (SVM) classifier is trained on the Olink®data matrix and evaluated via leave-one-out cross-validation. Associated protein biomarkers are identified via recursive feature elimination. We constructed three separate predictive models for binary classification: one to discern healthy controls from individuals with schizophrenia (AUC = 0.74), another to differentiate individuals who were responsive to antipsychotics (AUC = 0.88), and a third to distinguish treatment-resistant individuals (AUC = 0.78). Employing machine learning techniques, we identified features capable of distinguishing between treatment response subgroups. In this study, SVM demonstrates the power of machine learning to uncover subtle signals often overlooked by traditional statistics. Unlike t-tests, it handles multiple features simultaneously, capturing complex data relationships. Chosen for simplicity, robustness, and reliance on strong feature sets, its integration with explainable AI techniques like SHapely Additive exPlanations enhances model interpretability, especially for biomarker screening. This study highlights the potential of integrating machine learning techniques in clinical practice. Not only does it deepen our understanding of schizophrenia's heterogeneity, but it also holds promise for enhancing predictive accuracy, thereby facilitating more targeted and effective interventions in the treatment of this complex mental health disorder.

Fig. 1. Experimental approach and experimental design.

A Hierarchical approach used for model development to determine schizophrenia status, antipsychotic response, and clozapine response. HCL refers to healthy control, ARE refers to antipsychotic responsive, CRE refers to clozapine responsive, CRT refers to clozapine resistant. B Validation pipeline using train-test split.

Fig. 2. Difference in outcomes from the ML models and standard differential expression analysis using statistical techniques.

A Model performance for each classification task to determine pharmacological status. ROC curves for each model - (left) Status classification (HCL v. people with schizophrenia), (middle) Antipsychotic response classification (antipsychotic responsive v. antipsychotics resistant), and (right) Clozapine response classification (clozapine responsive v. clozapine resistant). B Bar graph showing the statistically significant fold change in protein expression of individuals with schizophrenia, compared to HCL. C Bar graph showing the statistically significant fold change in protein expression of individuals on clozapine, compared to individuals on antipsychotics except clozapine.

Fig. 3. Beeswarm plot of SHAP-calculation for protein markers identified via ML for predictive models to differentiate treatment response in individuals with schizophrenia.

Each datapoint represents a prediction from the test set, the x-axis represents selected proteins, and the y-axis represents the respective impact on the prediction outcome by the model (e.g. towards treatment resistance or treatment positive). Since the datapoints represent the contribution of each sample by each feature to its prediction, they are color coded to its relative protein expression value in the test set. Grey hue along the x-axis denotes proteins that are deemed to be statistically significant. There are no statistically significant features detected between clozapine responsive and clozapine resistant samples. (top) HCL from participants with schizophrenia, (middle) Antipsychotic responsiveness (ARE vs TRS), and (bottom) Individuals with TRS (CRE vs. CRT). Each datapoint represents a sample of each feature, which represents its SHAP value and how it contributes to the final prediction outcome (refer to Supplementary S3 for a comprehensive list of fold changes, P and SHAP values).