Computational Phenotyping of Treatment-Resistant Depression prior to Electroconvulsive Therapy

KEY POINTS

QUESTION : Can neurocomputational depictions of learning and affective behavior characterize patients with treatment-resistant depression before electroconvulsive therapy?

FINDINGS : In this observational study, computational models were used to quantify the behavioral dynamics of 1) adaptive choice behavior as individuals learned from feedback and 2) associated changes in affective self-report. These models provided quantitative parameters that were associated with specific neural and behavioral changes in patients with treatment-resistant depression and may be sufficient to independently identify patients with depression.

MEANING : Computational models that describe hypothesized mechanisms underlying adaptive behavior and affective experience may provide a means to quantitatively phenotype individual differences in major depression pathophysiology.

IMPORTANCE

Globally, treatment-resistant depression affects approximately one-third of all patients diagnosed with major depressive disorder. Currently, there are neither behavioral nor neural measures that quantitatively phenotype characteristics underlying treatment-resistant depression.

OBJECTIVE

Determine whether neurocomputational models that integrate information about adaptive behavior and associated self-reported feelings can characterize differences in patients with treatment-resistant depression.

DESIGN

In this observational study, data were collected over two research visits from 2020-2023 that occurred before and after standard-of-care electroconvulsive therapy (ECT) for treatment-resistant depression. This report focuses on “visit 1”, which occurred after patients consented to ECT but before their initial treatment.

SETTING

Wake Forest University School of Medicine; Atrium Health Wake Forest Baptist Psychiatric Outpatient Center; Atrium Health Wake Forest Hospital.

PARTICIPANTS

Participants planning to receive ECT for depression (“pre-ECT”) and participants not planning to receive ECT with (“non-ECT”) or without depression (“no-depression”), were recruited from the Psychiatric Outpatient Center and community.

EXPOSURES

Computerized delivery of a ‘Probabilistic Reward and Punishment with Subjective Rating’ task during fMRI.

MAIN OUTCOMES AND MEASURES

Computational modeling of choice behavior provided parameters that characterized learning dynamics and associated affect dynamics expressed through intermittent Likert scale self-reports. Multivariate statistical analyses relating model parameters, neurobehavioral responses, and clinical assessments.

RESULTS

Pre-ECT (N=29; 55.2% female), non-ECT (N=40; 70% female), and no-depression (N=41; 65.9% female). Parameters derived from computational models fit to behavior elicited during learning and the expression of affective experiences clearly differentiates the three groups. Reinforcement Learning model parameters alone do not perform as well as models that incorporate affective self-reports. Notably, the set of model parameters that include learning and affective dynamics demonstrated excellent, cross-validated, diagnostic classification of depression diagnosis. Prior to ECT, neurobehavioral responses associated with learning and affective experiences about ‘punishing’ events were significantly impaired in pre-ECT compared to non-ECT and no-depression cohorts.

CONCLUSIONS AND RELEVANCE

Computational models of behavioral dynamics associated with learning and affect can describe specific hypotheses about neurocomputational-mechanisms underlying treatment-resistant depression. The present work suggests differences in processing of emotionally negative states and suggests a potential model-based behavioral diagnostic for individuals with major depression. Such models may eventually be used to augment the diagnosis of treatment-resistant depression or possibly determine phenotype-genotype relationships for disease status and progression.