Resilience is defined as the ability to modify thoughts to cope with stressful events. Patients with schizophrenia (SCZ) having higher resilience (HR) levels show less severe symptoms and better real-life functioning. However, the clinical factors contributing to determine resilience levels in patients remain unclear. Thus, based on psychological, historical, clinical and environmental variables, we built a supervised machine learning algorithm to classify patients with HR or lower resilience (LR).

SCZ from the Italian Network for Research on Psychoses (N = 598 in the Discovery sample, N = 298 in the Validation sample) underwent historical, clinical, psychological, environmental and resilience assessments. A Support Vector Machine algorithm (based on 85 variables extracted from the above-mentioned assessments) was built in the Discovery sample, and replicated in the Validation sample, to classify between HR and LR patients, within a nested, Leave-Site-Out Cross-Validation framework. We then investigated whether algorithm decision scores were associated with the cognitive and clinical characteristics of patients.

The algorithm classified patients as HR or LR with a Balanced Accuracy of 74.5% (p < 0.0001) in the Discovery sample, and 80.2% in the Validation sample. Higher self-esteem, larger social network and use of adaptive coping strategies were the variables most frequently chosen by the algorithm to generate decisions. Correlations between algorithm decision scores, socio-cognitive abilities, and symptom severity were significant (pFDR < 0.05).

We identified an accurate, meaningful and generalizable clinical-psychological signature associated with resilience in SCZ. This study delivers relevant information regarding psychological and clinical factors that non-pharmacological interventions could target in schizophrenia.

The study aimed to subtype patients with schizophrenia on the basis of social cognition (SC), and to identify cut-offs that best discriminate among subtypes in 809 out-patients recruited in the context of the Italian Network for Research on Psychoses.

A two-step cluster analysis of The Awareness of Social Inference Test (TASIT), the Facial Emotion Identification Test and Mayer–Salovey–Caruso Emotional Intelligence Test scores was performed. Classification and regression tree analysis was used to identify the cut-offs of variables that best discriminated among clusters.

We identified three clusters, characterized by unimpaired (42%), impaired (50.4%) and very impaired (7.5%) SC. Three theory-of-mind domains were more important for the cluster definition as compared with emotion perception and emotional intelligence. Patients more able to understand simple sarcasm (⩾14 for TASIT-SS) were very likely to belong to the unimpaired SC cluster. Compared with patients in the impaired SC cluster, those in the very impaired SC cluster performed significantly worse in lie scenes (TASIT-LI <10), but not in simple sarcasm. Moreover, functioning, neurocognition, disorganization and SC had a linear relationship across the three clusters, while positive symptoms were significantly lower in patients with unimpaired SC as compared with patients with impaired and very impaired SC. On the other hand, negative symptoms were highest in patients with impaired levels of SC.

If replicated, the identification of such subtypes in clinical practice may help in tailoring rehabilitation efforts to the person's strengths to gain more benefit to the person.