Similarities and differences between multivariate patterns of cognitive and socio-cognitive deficits in schizophrenia, bipolar disorder and related risk

Abstract

Cognition and social cognition anomalies in patients with bipolar disorder (BD) and schizophrenia (SCZ) have been largely documented, but the degree of overlap between the two disorders remains unclear in this regard. We used machine learning to generate and combine two classifiers based on cognitive and socio-cognitive variables, thus delivering unimodal and multimodal signatures aimed at discriminating BD and SCZ from two independent groups of Healthy Controls (HC1 and HC2 respectively). Multimodal signatures discriminated well between patients and controls in both the HC1-BD and HC2-SCZ cohorts. Although specific disease-related deficits were characterized, the HC1 vs. BD signature successfully discriminated HC2 from SCZ, and vice-versa. Such combined signatures allowed to identify also individuals at First Episode of Psychosis (FEP), but not subjects at Clinical High Risk (CHR), which were classified neither as patients nor as HC. These findings suggest that both trans-diagnostic and disease-specific cognitive and socio-cognitive deficits characterize SCZ and BD. Anomalous patterns in these domains are also relevant to early stages of disease and offer novel insights for personalized rehabilitative programs.

Fig. 1. Outline of the study design.

To detect cognitive and socio-cognitive similarities and differences between schizophrenia and bipolar disorder, we employed a “reversal discovery-validation strategy”, consisting of two phases: (i) in Phase 1, we generated unimodal and multimodal signatures based on cognitive and socio-cognitive features aimed at discriminating two independent groups of Healthy Controls from two groups of patients suffering from Bipolar Disorder and Schizophrenia, respectively; (ii) in Phase 2, we applied the disease-related models generated in Phase 1 in each cohort to the other one data. In Phase 3, to test the generalizability of disease-related signatures on populations at risk or at early stages of psychosis, we applied the discovery models generated in Phase 1 also to cognitive and socio-cognitive data collected on Clinical High-Risk and First Episode of Psychosis individuals. BD = patients with Bipolar Disorder; CHR = participants with Clinical High Risk for psychosis; FEP = participants at First Episode of Psychosis; HC1 = Healthy Controls (group 1); HC2 = Healthy Controls (group 2); SCZ = patients with Schizophrenia.

Fig. 2. Performance metrics (classification plots, Receiver Operating Characteristic curves, and confusion matrices) of the classifiers discriminating between Healthy Controls (group 1) and Bipolar Disorder patients.

First row: cognitive classifier; second row: socio-cognitive classifier; third row: stacking-based classifier.

Fig. 3. Probability of each feature for being selected in the Machine Learning Cross-Validation framework for the cognitive and the socio-cognitive classifiers in the cohort including Healthy Controls (group 1) and Bipolar Disorder patients.

Scores closer to 1 represent a higher probability of being selected for decisions by the Support Vector Machine algorithm. Only features with a selection probability >0.5 are shown (complete selection probability results for the whole pool of cognitive and socio-cognitive features are available in SF1A and 1B, respectively). FEIT = Facial Emotion Identification Test; N = number; TASIT II = The Awareness of Social Inference Test – Section II; WMS = Wechsler Memory Scale.

Fig. 4. Performance metrics (classification plots, Receiver Operating Characteristic curves, and confusion matrices) of the classifiers discriminating between Healthy Controls (group 2) and Schizophrenia patients.

First row: cognitive classifier; second row: socio-cognitive classifier; third row: stacking-based classifier.

Fig. 5. Probability of each feature for being selected in the Machine Learning Cross-Validation framework for the cognitive and the socio-cognitive classifiers in the cohort including Healthy Controls (group 2) and Schizophrenia patients.

Scores closer to 1 represent a higher probability of being selected for decisions by the Support Vector Machine algorithm. Only features with a selection probability >0.5 are shown (complete selection probability results for the whole pool of cognitive and socio-cognitive features are available in SF2A and 2B, respectively). FEIT = Facial Emotion Identification Test; N = number; RAVLT = Rey Auditory Verbal Learning Test; SEM. FLUENCY = Semantic Fluency; TASIT II/III = The Awareness of Social Inference Test – Section II/Section III; WMS = Wechsler Memory Scale.

Fig. 6. Between-groups comparisons.

ANOVA analysis were conducted to compare decision scores from the stacking-based models discriminating Healthy Controls (group 1) vs. Bipolar Disorder patients (panel 6A) and Healthy Controls (group 2) vs. Schizophrenia patients (panel 6B) and decision scores extracted for Clinical High Risk and First Episode of Psychosis individuals after the Out-Of-sample-Cross-Validation procedure. Error bars represent standard error.