Personalized Prediction of Clozapine Treatment Response Using Therapeutic Drug Monitoring Data in Japanese Patients with Treatment-Resistant Schizophrenia

TL;DR

This study uses patient data and machine learning to predict how well Japanese patients with schizophrenia will respond to clozapine treatment.

Contribution

A random forest model was developed to personalize clozapine treatment response prediction using TDM data and clinical factors.

Findings

Men received higher clozapine doses than women, and glucose levels were elevated in both sexes.

The model achieved high AUC (0.986) in training and moderate performance (0.852) in testing for predicting treatment response.

Baseline BPRS score, treatment duration, age, and clozapine concentration were key predictors identified by SHAP analysis.

Abstract

Background: Clozapine is the only antipsychotic medication proven effective in patients with treatment-resistant schizophrenia (TRS). However, many patients have serum concentrations outside the recommended therapeutic window, and clozapine exhibits substantial interindividual variability. This study aimed to (1) examine clozapine dosage and blood concentrations in patients with TRS; (2) investigate the effects of sex and age on dosage and blood concentrations; (3) assess clinical response to clozapine treatment; and (4) develop a random forest (RF) model to predict therapeutic response using clinical and therapeutic drug monitoring (TDM) data. Methods: Dried blood spots were used to measure concentrations of clozapine, norclozapine, and clozapine N-oxide. Clinical symptoms were assessed using the Brief Psychiatric Rating Scale (BPRS). The RF algorithm was applied to analyze the relationships between biochemical and demographic factors and clinical response to clozapine. Results: A total of 754 blood samples from 167 patients were analyzed. Men received higher doses than women, and glucose levels were elevated in both sexes. The area under the curve (AUC) of the receiver operating characteristic (ROC) curve was 0.986 for the training set and 0.852 for the testing set. Accuracy, precision, recall, and F1-score (training/testing) were 0.938/0.786, 0.936/0.736, 0.934/0.780, and 0.935/0.757, respectively. The SHapley Additive exPlanations (SHAP) analysis indicated that baseline BPRS score, treatment duration, age, and clozapine concentration were important variables contributing to the output of the model. Conclusions: Our model achieved satisfactory predictive performance for clinical response and provides valuable insights into personalized prediction of clozapine efficacy.

Figures