Cingulate dynamics track depression recovery with deep brain stimulation

Treatment-resistant depression (TRD) presents significant challenges due to its debilitating symptoms and unpredictable response to treatments. While subcallosal cingulate (SCC) deep brain stimulation (DBS) offers durable symptom relief for some TRD patients, managing this treatment is complex. Antidepressant response is nonlinear and varies significantly between individuals, with periods of mood fluctuation that are difficult to interpret clinically. The lack of objective markers forces psychiatrists to rely on clinical intuition to adjust stimulation parameters or use a watchful waiting approach. Current methods, such as the Hamilton Depression Rating Scale (HDRS), are subject to recall biases and environmental influences, confounding the assessment of core depression symptom changes. This study aims to address this critical gap by developing an objective, brain-based biomarker to guide personalized SCC DBS treatment and improve understanding of the variable recovery trajectories observed in TRD patients. The study leverages advances in neurotechnology (long-term electrophysiology monitoring), high response rates in a clinical cohort, and explainable artificial intelligence (XAI) to derive a data-driven biomarker to differentiate acute clinical scenarios from normal transient distress.

Previous research has established the efficacy of SCC DBS in providing long-term symptom relief for TRD (Crowell et al., 2019; Holtzheimer et al., 2017). However, the non-linear and individualized nature of recovery trajectories, often involving periods of mood fluctuation, makes clinical management challenging (Crowell et al., 2015). The absence of objective biomarkers to guide treatment decisions leads to reliance on subjective clinical assessments, which are prone to various biases (Urban et al., 2018; Solhan et al., 2009). Studies have explored the use of acute electrophysiological measurements to understand SCC dynamics during DBS (Smart et al., 2018; Sendi et al., 2021; Choi et al., 2015; Ramasubbu et al., 2013), and the role of white matter pathway activation in therapeutic outcomes (Riva-Posse et al., 2014; Riva-Posse et al., 2018; Howell et al., 2019). Recent advances in XAI provide tools for the discovery of effective biomarkers from complex datasets (O’Shaughnessy et al., 2020). This study builds upon these previous findings by employing a longitudinal approach using a novel neurotechnology platform to derive a data-driven biomarker that can be used to inform clinical decision-making and personalize treatment.

Ten participants with TRD were enrolled in a single-site clinical trial using a prototype DBS device that enabled both stimulation and recording of local field potentials (LFPs). DBS leads were implanted at the intersection of four major white matter pathways. Stimulation was initiated after a 4-week postsurgical recovery period, and the primary endpoint was the HDRS-17 score at 24 weeks. Chronic electrophysiological data were available for six participants, five of whom exhibited typical response trajectories. LFP data, acquired weekly with stimulation off, were used to classify ‘sick’ versus ‘stable response’ states using a neural network classifier. A generative causal explainer (GCE) identified a spectral discriminative component (SDC), a low-dimensional latent representation of spectral features capturing the difference between these states. The SDC served as an LFP biomarker, with higher values indicating a ‘sick’ state. Preoperative structural and functional MRI data were acquired to investigate the relationship between white matter integrity and recovery trajectories. Facial expression data from weekly video recordings were analyzed using a data-driven approach to quantify behavioral changes. Random forest classifiers were trained on facial expression data to classify 'sick' and 'stable response' states. Statistical analyses included Wilcoxon signed-rank tests, linear mixed models, and correlation analyses.

At the 24-week endpoint, 90% of participants demonstrated a robust clinical response, and 70% achieved remission. Individual recovery trajectories varied considerably. The neural network classifier successfully distinguished ‘sick’ and ‘stable response’ states (AUROC: 0.87 ± 0.09). The GCE identified the SDC, an LFP biomarker that accurately reflected the clinical state (AUC: 0.94 ± 0.036), predicting transitions to ‘stable response’ with high accuracy. The SDC tracked changes in beta band power, showing an initial decrease followed by a sustained increase with chronic stimulation, distinct from acute stimulation effects. Increases in stimulation voltage resulted in a decrease in SDC (P=0.039), indicating a move towards the ‘well’ state. Analysis of a held-out participant correctly predicted a relapse based on SDC changes five weeks before clinical observation. Significant negative correlations were found between the time to reach ‘stable response’ and white matter integrity within the targeted network (Fig. 4a, b). A significant correlation between dACC FA and functional connectivity between SCC and MCC with the number of episodes in a lifetime indicated that structural and functional deficits in the target network were related to disease severity (Fig. 4c, d, e). A data-driven analysis of facial expressions showed that the face classifier output accurately tracked the SDC, demonstrating concordance between brain-based and behavioral measures of recovery (Fig. 5).

This study provides compelling evidence for the utility of the SDC as a novel biomarker for guiding personalized SCC DBS treatment for TRD. The SDC accurately reflects clinically defined states, responds to stimulation adjustments, and predicts relapses, offering a significant advance over existing subjective assessments. The observation that the transition to a stable response is associated with changes in beta band activity, specifically an initial decrease followed by an increase with chronic stimulation, suggests that sustained antidepressant responses are mediated by different mechanisms than transient effects. The relationship between white matter integrity and recovery trajectories underscores the importance of considering network-level effects in TRD. The concordance between the SDC and facial expression analysis provides a robust independent validation of the biomarker. The study highlights the potential for using multimodal measurements to gain a more comprehensive understanding of the mechanisms underlying TRD and to personalize treatment strategies.

This study demonstrates the feasibility and potential clinical utility of the SDC, a novel brain-based biomarker for guiding personalized SCC DBS treatment in TRD. The SDC provides objective information to support clinical decision-making, improving treatment outcomes and enhancing our understanding of TRD pathophysiology. Future research should focus on validating the SDC in larger, independent cohorts and exploring its integration into closed-loop DBS systems. Investigating the mechanisms underlying the observed changes in beta band activity and white matter integrity is also crucial.

The study has several limitations. The LFP analysis included only six of the ten participants due to technical challenges with the prototype device. LFP data were collected with therapeutic stimulation temporarily turned off. The analysis was retrospective, limiting its capacity to determine the precise timing of optimal stimulation adjustments. Further research is needed to fully explore the interaction between the SDC, acute mood fluctuations, and other factors influencing treatment response.