Predicting non-response to ketamine for depression: An exploratory symptom-level analysis of real-world data among military veterans.
Eric A. Miller, Houtan Totonchi Afshar, Jyoti Mishra, Roger S McIntyre, Dhakshin Ramanathan
Psychiatry Research May 1, 2024 DOI: 10.1016/j.psychres.2024.115858 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Retrospective analysis Peer reviewed |
|---|---|
| Sample size | 120 |
| Population | Adults with treatment-resistant depression who received intravenous racemic ketamine or intranasal esketamine in a real-world clinic |
| Intervention | intravenous racemic ketamine |
| Topics | Esketamine Ketamine Depression |
| Keywords | Predictive modeling Symptom trajectories Treatment resistant depression Treatment-resistant depression trd Refractory depression Predictive analytics Medical forecasting Veterans mental health |
| Citations | 12 |
| Key points | Using baseline symptoms alone, a logistic regression classifier identified 22% of patients who would not respond to ketamine or esketamine with a negative predictive value over 96%. |
Abstract
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. Herein, we aim to inform prediction models of non-response to ketamine/esketamine in adults with TRD. This is a retrospective analysis of PHQ-9 item response data from 120 patients with TRD who received repeated doses of intravenous racemic ketamine or intranasal eskatamine in a real-world clinic. Regression models were fit to patients' symptom trajectories, showing that all symptoms improved on average, but depressed mood improved relatively faster than low energy. Principal component analysis revealed a first principal component (PC) representing overall treatment response, and a second PC that reflects variance across affective versus somatic symptom subdomains. We then trained logistic regression classifiers to predict overall response (improvement on PC1) better than chance using patients' baseline symptoms alone. Finally, by parametrically adjusting the classifier decision thresholds, we identified optimal models for predicting non-response with a negative predictive value of over 96 %, while retaining a specificity of 22 %. Thus, we could identify 22 % of patients who would not respond based purely on their baseline symptoms. This approach could inform rational treatment recommendations to avoid additional treatment failures.