Monitoring Depression Recovery through Cingulate Dynamics with Deep Brain Stimulation

Treatment-resistant depression (TRD) presents a significant challenge in clinical psychiatry. Deep brain stimulation (DBS) targeting the subcallosal cingulate (SCC) has emerged as a promising intervention, but its efficacy varies among individuals, necessitating trial-and-error adjustments. To address this issue, we introduced a novel approach involving electrophysiology recording during SCC DBS treatment for ten TRD patients. Our findings, detailed in this article, shed light on a potential game-changer in TRD management, leveraging artificial intelligence (AI) and objective biomarkers to guide personalized treatment.

Introduction:

Treatment-resistant depression (TRD) is a debilitating condition, often necessitating unconventional interventions. Deep brain stimulation (DBS) of the subcallosal cingulate (SCC) has shown promise but is challenging to optimize. In this article, we explore how advanced technology and AI can provide objective biomarkers to revolutionize TRD management.

Objective Biomarkers for TRD:

Current TRD management relies heavily on subjective assessments, leading to challenges in distinguishing natural mood fluctuations from genuine intervention requirements. To bridge this gap, we conducted a study using a novel device capable of electrophysiology recording during SCC DBS treatment. The study involved ten TRD participants, and at the 24-week endpoint, a remarkable 90% exhibited significant clinical improvement, with 70% achieving remission. Our innovative approach involved utilizing explainable artificial intelligence (XAI) to identify SCC local field potential changes that correlate with the patient's clinical state. This biomarker, distinct from transient stimulation effects, proved effective in guiding therapeutic adjustments and accurately tracking individual recovery states.

Individual Recovery Trajectories:

Notably, our research unveiled that recovery trajectories vary based on preoperative structural integrity and functional connectivity within the targeted white matter treatment network. Objective biomarkers helped match these trajectories, with concurrent detection of objective facial expression changes through data-driven video analysis.

Implications for TRD Treatment:

Our findings highlight the potential of objective biomarkers in personalizing SCC DBS treatment for TRD. They offer valuable insights into the complex interplay between functional, anatomical, and behavioral aspects of TRD pathology, encouraging further research to elucidate variability in depression treatment.

Conclusion:

In conclusion, our study represents a significant advancement in the management of TRD, introducing objective biomarkers and AI-driven insights into SCC DBS treatment. This transformative approach holds promise for more effective, personalized TRD management and provides fresh perspectives on the multifaceted nature of depression treatment. Further research in multimodal measurements promises to advance depressive disorder treatments even further.

Patient Cohort and Longitudinal Data:

Our study involved ten TRD patients who received an experimental DBS implanted pulse generator (IPG). DBS leads were precisely placed at the intersection of four major white matter pathways. The primary endpoint of the study was the Hamilton Depression Rating Scale (HDRS) score at 24 weeks of SCC DBS treatment. Notably, 90% of patients experienced significant reductions in their HDRS scores, indicating clinical improvement. Individual patient trajectories varied, with some responding more rapidly than others.

Leveraging Electrophysiology Data:

Electrophysiological data from six participants enabled us to develop a neural network classifier that effectively distinguished between 'sick' and 'stable response' states, showing consistent electrophysiological changes across patients. Furthermore, we identified a spectral discriminative component (SDC) that served as a low-dimensional representation of these changes, reflecting the patient's depressive or recovered state. Importantly, the SDC correlated closely with HDRS scores, demonstrating its potential as an objective biomarker.

Beta Band Dynamics:

Our analysis revealed that changes in beta and gamma band power in the SCC were indicative of recovery. Interestingly, while acute stimulation experiments had previously shown a decrease in beta band power post-stimulation, our longitudinal analysis demonstrated a sustained increase in beta band power during chronic stimulation. This suggests a distinct long-term effect of SCC DBS.

Stimulation Voltage Adjustments:

During the 24-week treatment protocol, stimulation voltage adjustments were made as needed. Our research found that increases in stimulation voltage led to a decrease in the SDC, indicating progress toward a 'well' state. In contrast, these voltage adjustments did not consistently or significantly affect HDRS scores. This reinforces the potential of the SDC as a sensitive and reliable biomarker for tracking patient progress.

Figure 3: Response to Stimulation Change and Validation in Relapsed Responder

In Figure 3a, we observe the changes in the SDC (left) and HDRS-17 (right) scores before and after a voltage adjustment in the stimulation during the week. The gray lines represent individual changes in response to stimulation voltage adjustments, while the black lines represent the average change across all adjustments. Error bars indicate standard deviations (n=8 stimulation dose changes). Notably, we found a significant improvement with a p-value of 0.04 using a one-sided Wilcoxon signed-rank test.

Figure 3b provides an illustration of SDC data from an out-of-sample participant who experienced a relapse. The blue line represents the HDRS-17 score, while the red line signifies the SDC calculated from LFP features not involved in the classifier or SDC training. The SDC shows a noticeable increase above the threshold of 0.5 (indicated by the grey dashed line), signifying relapse (red arrow) at week 12, which was observed in the HDRS-17 score at week 17 (blue arrow). Purple arrows represent changes in stimulation voltage levels. Interestingly, adjusting stimulation voltage did lead to a decrease in the SDC, as demonstrated in Figure 3a. However, the SDC only stabilized after three voltage adjustments. Remarkably, the final voltage level in this patient (4.5 V) was comparable to the average voltage used in typical responders (4.4 ± 0.57 V).

SDC Tracking Relapse in an Out-of-Sample Patient

To showcase the practical utility of the SDC in a clinical context, we conducted a retrospective analysis of LFP data from one participant (P001), who was not part of the classifier or GCE training dataset. This participant experienced a clinical relapse after four months of remission. P001 initially exhibited low HDRS-17 scores (below 8) during the active stimulation phase but later had a significant and sustained worsening of symptoms, leading to classification as a non-responder by week 16 (see Figure 3b, blue line). Using the SDC model trained on five typical responders (excluding P001), we observed that the SDC accurately captured this trend in P001, indicating a transition from a response state to a sick state (Figure 3b, red line). Intriguingly, the SDC predicted the relapse based on brain signals (Figure 3b, red arrow) approximately one month before clinical relapse was detected by the HDRS-17 (Figure 3b, blue arrow). This illustrates the SDC's potential to forecast impending instability and the need for early intervention, even before clinical symptoms become apparent. Furthermore, increases in stimulation dosage (Figure 3b, purple arrows) led to reductions in the SDC, but the effect persisted only after three adjustments. Notably, the final stable voltage for this patient (4.5 V) after the six-month study period closely matched the average dose used in typical responders.

To assess the concordance between HDRS-17 and the SDC in this out-of-sample participant, we compared the states indicated by these two measures. As the therapeutic response occurred at the beginning of the observation period, we couldn't use the criteria for a "stable response" described earlier. Nevertheless, when considering two states, "sick" and "response," denoting a change in HDRS-17 of less than 50% decrease and greater than 50% decrease, respectively, we found that the SDC state accurately predicted the HDRS state 75% of the time over the 24-week treatment course, with a p-value of 0.029 using a shuffle-based procedure.

White Matter Abnormality Correlates of Transition

Previous research has established that incomplete white matter pathway activation can influence therapeutic outcomes in SCC DBS. We hypothesized that functional and structural abnormalities in these pre-specified white matter bundles could also impact the recovery trajectory, as inferred from the SDC. Through preoperative imaging, we identified significant negative correlations between the transition weeks to reach a "stable response" state, as identified by the SDC, and white matter integrity. This was assessed using fractional anisotropy (FA) and radial diffusivity, as well as FA and axial diffusivity. Regions showing these significant correlations between structural integrity and time to recovery were associated with white matter bundles connecting the DBS target site to regions like the ventromedial frontal cortex (vmF), anterior hippocampus (aHc), insular (Ins), and dorsal anterior and posterior cingulate cortex (dACC and PCC), respectively (Figure 4a, b). These findings suggest that alterations in white matter microstructure within the targeted brain network contribute to longer treatment durations needed to achieve a stable response. Specifically, the correlation of radial diffusivity with time to recovery supports the idea that demyelination at baseline is a primary factor contributing to white matter deficits associated with variable recovery times in patients.

Moreover, we identified a significant correlation between white matter abnormalities in the dACC and functional connectivity between the SCC and MCC, emphasizing a link between functional properties within the target network and structural properties impacting prospective notions of disease severity, as measured by time to recovery. Additionally, when considering the entire cohort, we found a significant negative correlation between dACC FA and SCC-MCC functional connectivity and the number of lifetime depressive episodes experienced by each individual before SCC DBS (n=9 participants, with one excluded due to image artefacts). This correspondence suggests that structural and functional deficits within the target network are also related to a retrospective measure of disease severity based on a patient's history of chronic depression.

SDC Tracking Changes in Facial Expressions

In addition to standardized clinical rating scales, we quantified behavioral improvements by analyzing changes in facial expressions captured in videos of weekly clinical interviews. These features encompassed various aspects of facial movements, including facial action units, eye gaze, and head pose (Figure 5a). Importantly, these features were not designed to explicitly represent specific emotions (e.g., sadness). To distinguish "sick" and "stable response" states, we employed individualized classifiers for each patient, a contrast to the single LFP classifier used for the entire cohort. Random forest classifiers, based on facial expression features, successfully classified "sick" and "stable response" states for each individual participant (AUROC 0.95 ± 0.05), demonstrating consistent yet individualized differences between these states (Figure 5b).

Subsequently, we used these individual facial expression features from the intermediate period (weeks 5−20) to determine the classifier's prediction of the disease state, termed "face classifier output." As a secondary confirmation of the SDC biomarker, we compared the face classifier output to the SDC for each patient. We observed that the face classifier output's trajectory closely paralleled the corresponding participant's SDC trajectory (Figure 5d; Extended Data Fig. 6), and there was a significant relationship between the face classifier output and the SDC (Figure 5e; linear mixed model, F(1.00, 51.74) = 6.54, P = 0.01). We also found that the transition weeks from the "sick" state to the "stable response" state, as inferred from the SDC and face classifier output, were in

agreement (Figure 5f; Kendall's tau = 0.89, P = 0.037) when using a strict threshold (0.35) to binarize these measures for direct comparison. These results collectively suggest that the SDC accurately tracks changes in facial expressions associated with recovery from depression.

Discussion

This study explored the long-term multimodal changes associated with SCC DBS and introduced the SDC as an objective biomarker capable of accurately capturing "sick" and "stable response" states in all patients while responding to changes in DBS stimulation. The transition to achieving a "stable response" state, as identified by the SDC, was correlated with structural and functional abnormalities in the targeted white matter tracts. This correlation was further supported by an analysis of complex facial expressions, emphasizing the distinct phases of recovery observed with chronic DBS, as opposed to short-term stimulation effects.

Notably, the SDC demonstrated potential clinical value by predicting a relapse in an out-of-sample patient before clinical symptoms manifested, highlighting its ability to inform early intervention decisions. We also observed instances where the SDC indicated a transition to stable recovery ahead of HDRS-17 scores due to changes in anxiety symptoms without altering core depression symptoms. These observations underline the SDC's capacity to differentiate between scenarios outlined in Figure 1e, providing valuable information for clinical management decisions. Importantly, the SDC's applicability extends across participants, eliminating the need for recent individualization strategies proposed elsewhere. Future replication studies in independent cohorts will enhance the specificity and sensitivity required for implementing a "clinician-in-the-loop" DBS approach.

Our findings regarding beta band activity changes align with previous studies and suggest that chronic DBS has long-term effects that resemble the actions of slower-acting antidepressants, such as SSRIs. Furthermore, we hypothesize that beta band changes reflect network-wide alterations across multiple regions within the targeted treatment network. The correlation between white matter abnormalities and recovery time underscores the importance of structural integrity within the target network in depression pathophysiology. These findings are consistent with post-mortem observations of white matter and oligodendroglia abnormalities in TRD suicides in the SCC region and its projections, supporting the relevance of white matter microstructure alterations in depression.

Lastly, our personalized analysis of facial expressions provides an independent assessment of a patient's clinical state, closely mirroring the SDC's insights. While some facial action units are common across participants, individualized differences in facial expressions exist between "sick" and "stable response" states. The bilateral and symmetric nature of these changes is suggestive of the normalization of emotional rather than volitional facial movement, consistent with the impact of cingulum bundle lesions. This analysis adds a valuable dimension to clinical evaluations, highlighting the importance of patient appearance in diagnosis and recovery assessments.

In conclusion, our study presents a comprehensive exploration of SCC DBS treatment's long-term effects, introducing the SDC as an objective biomarker capable of monitoring patient states throughout treatment. The SDC's ability to predict relapses and track changes in facial expressions provides valuable insights for clinical decision-making. Moreover, the correlation between structural and functional white matter abnormalities and recovery times emphasizes the importance of white matter integrity in depression treatment. These findings collectively advance our understanding of SCC DBS therapy and its potential to transform the management of treatment-resistant depression.

While this study provides valuable insights into the application of SCC DBS therapy for treatment-resistant depression and introduces the SDC as a promising biomarker, it is essential to acknowledge several limitations. Firstly, the LFP analysis faced challenges due to data artifacts and protocol changes after pilot implantations, limiting the number of participants for certain analyses. However, it's worth noting that the SDC remained reliable across all participants, including the held-out patient, demonstrating its robustness.

Secondly, the study used LFP data collected with therapeutic stimulation temporarily turned off to eliminate significant stimulation-related artifacts. Future research might explore the feasibility of calculating the SDC during transient periods without the need to interrupt therapeutic stimulation, which could be more practical in a clinical setting.

Thirdly, the study did not explicitly model acute moment-to-moment distress, which could enhance the behavioral interpretability of the chronic biomarker. Future studies with increased data collection frequency may address this limitation by identifying transient mood or anxiety symptoms through LFP signatures.

Lastly, the analysis in this study is retrospective, leaving open questions about the precise timing of optimal stimulation adjustments or the introduction of adjunct rehabilitative interventions. Future research could focus on prospectively evaluating the SDC's utility in guiding clinical decision-making for individualized DBS treatment plans.

The study included ten participants with treatment-resistant major depressive disorder enrolled in a single-site clinical trial. The trial employed a prototype DBS device capable of collecting local field potentials (LFPs) from the stimulation site. Clinical symptom severity was assessed using the 17-item Hamilton Depression Rating Scale (HDRS-17), Montgomery-Åsberg Depression Rating Scale (MADRS), and Beck Depression Inventory during weekly visits. DBS therapy started at least 30 days after implantation surgery and consisted of bilateral monopolar stimulation with specific parameters.

LFP recordings were performed weekly during a 6-month observation phase, with each session divided into segments with and without stimulation. Only segments without stimulation-related artifacts were included in the analysis. Feature extraction involved estimating spectral power, coherence, and phase-amplitude coupling (PAC) from the LFP data.

LFP classification and the inference of the Stable Disease Course (SDC) involved the use of neural network models and a Gaussian Copula-based autoencoder framework. The trained feature compression network inferred the SDC from LFP data collected during months 2–5. The SDC was then transformed into the probability of belonging to the 'sick' state and averaged across 10-second segments within a week.

The study examined the transition to a 'stable response' state in participants receiving SCC DBS treatment. The transition was defined based on HDRS-17, SDC, or a face classifier output, using specific thresholds and criteria.

Image acquisition and processing included high-resolution structural T1-weighted images, resting-state functional MRI, and diffusion-weighted images. Tract-Based Spatial Statistics processing was performed to evaluate white matter microstructure. Additionally, a Ventral Tegmental Area (VTA) was generated, and white matter tracts passing through the VTA were analyzed.

Facial expression analysis involved processing video recordings of participants during clinical interviews. OpenFace software was used to extract features from facial expressions, which were then used to train binary classifiers to discriminate between 'sick' and 'stable response' states. These classifiers provided another behavioral marker for tracking the response during ongoing DBS.

Statistical analysis employed one-sided Wilcoxon signed-rank tests for hypothesis testing of changes in clinical assessments, the SDC, and individual features. Linear mixed models were used for correlation analyses between the SDC and clinical scores, as well as between the SDC and the face classifier output.

Overall, this study contributes to our understanding of SCC DBS therapy and its potential to transform the management of treatment-resistant depression. However, it is important to consider the limitations mentioned in the discussion section while interpreting the findings.