High-order brain interactions during ketamine-induced state changes: A functional marker of response in late-life treatment-resistant depression?

Ketamine is recognised as a rapid-acting option for treatment-resistant depression, but only a subset of patients show a robust clinical response, and the neural mechanisms that support benefit remain uncertain. The paper places this problem in the context of prior EEG findings suggesting that ketamine alters oscillatory activity and may shift the brain towards a more flexible functional state. The authors argue that conventional pairwise connectivity measures may miss more complex, nonlinear relations between brain regions, and that high-order interactions (HOI) derived from multivariate information theory could better capture these network changes. Shah and colleagues set out to test whether ketamine produces temporally evolving changes in HOI in late-life veterans with treatment-resistant depression, and whether those changes relate to antidepressant response and acute dissociation. They specifically hypothesised that ketamine-induced changes would be associated with increases in brain redundancy, as measured by the O-information, and that these effects would differ from those seen with midazolam. The study is presented as a secondary analysis intended both to extend earlier HOI work in healthy volunteers and to explore whether HOI might serve as a clinically useful biomarker in depression treatment.

This was a secondary analysis of EEG data from a randomised, double-blind, midazolam-controlled, multi-arm trial of intravenous ketamine in late-life treatment-resistant depression (ClinicalTrials.gov NCT02556606). The analysed sample comprised 30 veterans: 18 received ketamine and 12 received midazolam. The ketamine arm pooled three doses, 0.1, 0.25, and 0.5 mg/kg, on the basis of prior work suggesting dose-invariant EEG effects; the authors nevertheless ran dose-sensitivity analyses to check whether pooling was reasonable. Missing data were excluded. EEG was recorded from a 64-channel system. For resting-state analysis, alternating eyes-open and eyes-closed segments were collected, but the analysis was restricted to the first 100 seconds to focus on the eyes-closed condition and reduce visual variability. The data were filtered, downsampled, cleaned using artefact subspace reconstruction and independent component analysis, and then analysed on a reduced 15-electrode montage because the original HOI method required a comparable electrode set. For mismatch negativity (MMN), data were epoched around stimulus onset, cleaned similarly, and then concatenated into continuous time series for standard and deviant conditions. The main outcomes were high-order information-theoretic metrics: O-information, which reflects whether interactions are more redundancy-dominant or synergy-dominant, and S-information, which reflects overall higher-order interdependence. Analyses focused on baseline, 1 hour, 24 hours, and 7 days after infusion for resting-state data, and 2 hours and 7 days for MMN data. Post-infusion values were expressed as change from baseline. The authors first used effect-size-based feature selection to identify the strongest electrode combinations and then evaluated those features in linear mixed-effects models with subject-level random intercepts. Model selection used Bayesian Information Criterion and considered age, sex, BMI, baseline HOI, group, time, and group-by-time interactions. Associations with clinical change were then examined using Spearman correlations, permutation testing, false discovery rate correction, and ordinary least squares models adjusted for age, sex, BMI, and baseline HOI. Clinical outcomes were MADRS change from baseline to day 7 and CADSS change from baseline to 1 hour.

In resting-state EEG, ketamine produced temporally dynamic changes in HOI, with the largest single effect seen for alpha-band O-information at 1 hour after infusion (Cohen's d = 2.57). Across time, the effect pattern shifted: alpha effects were strongest early, theta effects were more evident at 24 hours, and gamma effects partially re-emerged by day 7. Across all 32,647 analysed n-plets, distribution-level summaries showed a similar early positive shift, while S-information tended to become negative by day 7, particularly in alpha and theta bands. In the mixed-effects models, the best-fitting structure included group, timepoint, baseline HOI, and a group-by-timepoint interaction. Significant group effects survived false discovery rate correction in 7 of 8 band-by-metric families; theta S-information was the only family that did not survive correction. The strongest effects were for alpha O-information (χ² = 48.73, q < .001), beta O-information (χ² = 24.70, q < .001), and gamma O-information (χ² = 23.41, q < .001). Baseline HOI was significant across all band-by-metric families. Significant group-by-time effects were present for alpha O-information (χ² = 25.50, q < .001) and gamma O-information (χ² = 10.27, q = .024). The maximal electrode configurations were not stable across timepoints, suggesting that different n-plets carried the peak effects at different stages. During MMN processing, ketamine also altered HOI, with early increases followed by attenuation. Mixed-effects modelling of the maximal n-plets supported a model including group, timepoint, and baseline HOI. FDR-corrected group effects remained significant for O-information in both deviant and standard responses (q = .001), whereas S-information did not survive correction. The strongest task-evoked effects shifted spatially from central-posterior sites early to more frontal-central configurations later. Clinical associations were selective. For depressive symptoms, only one feature remained significant after correction in the covariate-adjusted model: 24-hour alpha O-information predicted day 7 MADRS improvement (β = 69.31, SE = 25.77, p = .007, q = .050). This meant that greater increases in alpha-band redundancy at 24 hours were associated with greater symptom improvement at day 7. Other candidate associations were present but did not survive correction. Associations with dissociative symptoms were stronger. In the primary model, four 24-hour features survived FDR correction: alpha O-information (β = 6.10, q = .007), alpha S-information (β = 5.48, q = .019), beta O-information (β = 34.11, q < .001), and beta S-information (β = 6.00, q = .002). All were positive, indicating that greater increases in HOI were associated with greater CADSS scores. The authors note that some earlier candidate associations, including gamma-band features, did not survive correction in the final models.

Shah and colleagues interpret the findings as evidence that ketamine induces large-scale, frequency-specific reorganisation of higher-order brain dynamics in late-life treatment-resistant depression. They emphasise that the effects were temporally structured rather than static: alpha-band redundancy rose earliest and most strongly, theta effects were more apparent later, and gamma-related effects persisted to day 7. In their view, this pattern supports the idea that ketamine changes not just the strength of brain activity, but the way information is distributed across the network over time. They also highlight that these effects were detectable with a relatively low-density electrode montage, which they present as encouraging for potential clinical scalability. The authors relate their results to earlier HOI work in healthy volunteers, stating that the present study extends those findings into a controlled clinical sample. They argue that the observed increase in alpha-band redundancy is notable because it does not simply mirror conventional oscillatory power changes; instead, it may reflect an aspect of network organisation that is not captured by spectral measures alone. They suggest that this could be relevant to models in which ketamine reduces hierarchical control and increases flexibility in information processing. For clinical outcomes, the authors take a cautious view. They report that the antidepressant association was limited to a single 24-hour alpha O-information feature, and they describe this as selective and exploratory rather than a broad HOI signal. By contrast, associations with acute dissociation were stronger and more consistent, with multiple 24-hour alpha and beta features surviving correction. They interpret this as suggesting that HOI-clinical relationships differ by outcome and are not reducible to one shared mechanism within this dataset. The authors acknowledge several limitations. The single-infusion design limits generalisability to repeated dosing, and the sample was small, late-life, predominantly male, and drawn from veterans, which reduces representativeness. The modest ketamine subgroup size limited power to detect dose heterogeneity, and the small midazolam arm together with adaptive randomisation may have created imbalance. The reduced electrode montage improves scalability but lowers spatial resolution and may miss finer network structure. They also note that HOI metrics are undirected, so they cannot determine the direction of information flow, and that some spatial differences from earlier studies suggest variability in exact electrode contributions. They further state that future work should assess temporal trajectories more explicitly, integrate HOI with spectral power, refine spatial modelling, and test predictive utility in larger longitudinal samples. The authors conclude that HOI may be a useful and potentially scalable framework for studying ketamine-related changes in brain organisation, while still requiring further validation.