In this study, a classifier (hyperplane) is determined to distinguish the neural responses during emotion regulation versus viewing images in healthy adults and then applied to determine (i) the effectiveness of the emotion regulation response (defined as emotion regulation distance from the hyperplane [DFHER]) in independent samples of healthy adults, patients with BD, and the patients’ unaffected relatives (URs) and (ii) the association of DFHER with the duration of future (hypo)manic and depressive episodes for patients with BD over a 16-month follow-up period.

Study participants (N = 226) included 65 healthy adults (35 used for support vector machine [SVM] learning [HCTrain] and 30 kept as an independent test sample [HCTest]), 87 patients with newly diagnosed BD (67% BD type 2) and 74 URs. BOLD response data came from an emotion regulation task. Clinical symptoms were assessed at baseline fMRI and after 16 months of specialized treatment.

The SVM ML analysis identified a hyperplane with 75.7% accuracy. Patients with BD showed reduced DFHER relative to the HCTest and UR groups. Reduced DFHER was associated with reduced improvement in psychosocial functioning during the 16-month follow-up time (B = −1.663, p = 0.02).

The neural response during emotion regulation can be relatively well distinguished in healthy adults via ML. Patients with newly diagnosed BD show significant disruption in the recruitment of this emotion regulation response. Disrupted may indicate a reduced capacity for functional improvement during specialized treatment in a mood disorder clinic.

Machine learning (ML) has developed classifiers differentiating patient groups despite concerns regarding diagnostic reliability. An alternative strategy, used here, is to develop a functional classifier (hyperplane) (e.g. distinguishing the neural responses to received reward v. received punishment in typically developing (TD) adolescents) and then determine the functional integrity of the response (reward response distance from the hyperplane) in adolescents with externalizing and internalizing conditions and its associations with symptom clusters.

Two hundred and ninety nine adolescents (mean age = 15.07 ± 2.30 years, 117 females) were divided into three groups: a training sample of TD adolescents where the Support Vector Machine (SVM) algorithm was applied (N = 65; 32 females), and two test groups– an independent sample of TD adolescents (N = 39; 14 females) and adolescents with a psychiatric diagnosis (major depressive disorder (MDD), generalized anxiety disorder (GAD), attention deficit hyperactivity disorder (ADHD) & conduct disorder (CD); N = 195, 71 females).

SVM ML analysis identified a hyperplane with accuracy = 80.77%, sensitivity = 78.38% and specificity = 88.99% that implicated feature neural regions associated with reward v. punishment (e.g. nucleus accumbens v. anterior insula cortices). Adolescents with externalizing diagnoses were significantly less likely to show a normative and significantly more likely to show a deficient reward response than the TD samples. Deficient reward response was associated with elevated CD, MDD, and ADHD symptoms.

Distinguishing the response to reward relative to punishment in TD adolescents via ML indicated notable disruptions in this response in patients with CD and ADHD and associations between reward responsiveness and CD, MDD, and ADHD symptom severity.