MULTIMODAL IDENTIFICATION OF NEUROBIOLOGICAL SUBTYPES ASSOCIATED WITH KETAMINE RESPONSE IN TREATMENT-RESISTANT DEPRESSION USING NETRAAI

Background

Ketamine has demonstrated rapid antidepressant effects in treatmentresistant depression (TRD), but its psychoactive properties complicate blinding and obscure identification of true biological response. Integrating multimodal clinical and neurobiological data may help resolve response heterogeneity and clarify ketamine’s mechanisms of action.

Methods

We reanalyzed data from a small, randomized, double-blind, placebo-controlled crossover Phase II trial of intravenous racemic ketamine (0.5 mg/kg) versus saline placebo in TRD (NCT00088699; n=33). NetraAI, an explainable AI/ML framework designed to identify low-dimensional human-interpretable feature sets, was used to analyze and integrate baseline clinical symptom scales (e.g., MADRS, HAM-A, HAMD-17, BDI), physiological measures (e.g., BMI), volumetric structural MRI features, and resting-state MEG recordings. Due to the crossover design, participant data from both infusion sessions were pooled, yielding 63 analyzable cases (3 incomplete subject datasets were excluded). Pooling was performed to increase statistical power while preserving within-subject treatment labels; analyses focused on identifying stable subgroup structure rather than independent observations. Treatment response was defined as a ≥40% reduction in MADRS scores at Day 7. The NetraAI framework partitions the patient population into explainable and unexplainable subpopulations, focusing on identifying compact (2-4 variable), multimodal feature sets defining biologically coherent Model-Derived Subgroups (MDS).

Results

Multimodal integration revealed multiple internally consistent reproducible ketamine responder MDS characterized by convergent clinical, anatomical, and functional signatures. One MDS (p=0.028, Cohen’s d=2.11) integrated elevated emotional reactivity with reduced posterior cingulate and inferior parietal white matter volumes, implicating default mode network and emotional regulation circuitry consistent with ketamine’s glutamatergic modulation of largescale networks. Another MDS (p=0.047, Cohen’s d=1.33) combined cardiovascular anxiety symptoms with elevated isthmus cingulate and lingual grey matter volumes, suggesting altered interoceptive and autonomic processing within salience-related networks. The large effect sizes of these MDS indicate clinically meaningful separation of strong ketamine responders. Additional MDS incorporating MEG-derived variables were also identified.

Conclusions

Integrating clinical symptoms, physiological measures, structural neuroanatomy, and functional network dynamics identifies distinct neurobiological pathways associated with ketamine response in TRD. These findings demonstrate how NetraAI’s multimodal data integration can link symptom dimensions to brain circuitry, support biomarker-informed patient stratification and improve interpretability of ketamine trials. Findings should be interpreted in the context of modest sample size and crossover design but collectively support the potential of explainable multimodal approaches to advance precision psychopharmacology and trial enrichment strategies in neuropsychiatric drug development.