Predicting non-response to ketamine for depression: a symptom-level analysis of real-world data

Abstract

Background Ketamine helps some patients with treatment resistant depression (TRD), but reliable methods for predicting which patients will, or will not, respond to treatment are lacking. Methods This is a retrospective analysis of PHQ-9 item response data from 120 military veterans with TRD who received repeated doses of intravenous racemic ketamine or intranasal eskatamine in a real-world clinic. Regression models were fit to individual patients' symptom trajectories and model parameters were analyzed to characterize how different symptoms responded to treatment. Logistic regression classifiers were used to predict treatment response using patients' baseline depression symptoms alone. Finally, by parametrically adjusting the classifier decision thresholds, the full space of models was searched to identify the best models for predicting non-response with very high negative predictive value. Results Model slopes indicated progressive improvement on all nine symptoms, but the symptom of depressed mood improved faster than the symptom of low energy. The first principal component (PC) represented a data-driven measurement of overall treatment response, while the second PC divided the symptoms into affective and somatic subdomains. Logistic regression classifiers predicted response better than chance using baseline symptoms, but these models achieved only 60.2% predictive value. Using threshold tuning, we identified models that can predict non-response with a negative predictive value of 96.4%, while retaining a specificity of 22.1%, suggesting we could successfully identify 22% of individuals who would not respond purely based on baseline symptom scores. Conclusions We developed an approach for identifying a subset of patients with TRD who will likely not respond to ketamine. This could inform rational treatment recommendations to avoid additional treatment failures.