5.4 INDIVIDUALIZED PREDICTION OF FUNCTIONAL OUTCOMES IN SCHIZOPHRENIA PATIENTS IN RESPONSE TO NEURO-COGNITIVE INTERVENTION: A MACHINE LEARNING ANALYSIS

Abstract

Background: Computer-based neurocognitive interventions (cNCI) represent an innovative way to improve cognitive functioning and functional outcome in schizophrenia patients. However, evidence suggests that cNCI can be highly effective in some, though not all patients. This variability is most likely due to intermediate neurocognitive and brain phenotypes that moderate the neuroplastic response induced by the respective training paradigms. Previous studies also suggested that prospective short- and long-term functional outcomes can be individually approximated in early psychosis using not only clinical, but also structural neuroimaging data (sMRI) (Kambeitz- Ilankovic et al., 2015, Koutsouleris et al., 2018). We employed multivariate pattern analysis (MVPA) to identify patterns in sMRI data predictive of functional outcome after 40 hours of cNCI. Method(s): The support vector machine (SVM) pipeline wrapped grey matter (GM) volume matrices into repeated nested, cross-validation that was used to train a multi-modal prediction of good (18) vs. poor functional outcome (17) that underwent 40 hours of neuro-cognitive training including social cognition exercises. The generalization capacity of the model was evaluated by applying this model to 18 unseen schizophrenia patients that underwent 50 hours cognitive intervention. Finally, we used General Assessment of Functioning Scale (GAF) to determine good or poor outcome status at baseline and follow up. We aimed to set a clinically meaningful cut-off by applying median split to GAF scores to differentiate poor and good "outcomers" of both samples. Result(s): Volume-based pattern classification predicted good vs. poor outcome status at follow-up in with a balanced accuracy of 74. 5% (sensitivity 64.7%, specificity 84.2%) as determined by nested cross-validation. Neuroanatomical prediction of functional outcome was nearly as accurate when predicting the difference in global level of functioning at two different time points (baseline and follow-up) with balanced accuracy of 65 % (sensitivity 66.7%, specificity 64.7%). Both models were significant (p<0.05) and generalizable onto independent intervention sample with balanced accuracy of 68.8 % and 58%, respectively. The neuroanatomical markers that showed diagnostic specificity for prediction of poor functional outcome in schizophrenia patients after cNCI involved medial and dorsolateral prefrontal and temporo-parietal GM volume reductions, as well as GMV increments in cerebellar and occipital regions. Conclusion(s): Identifying basic neuroimaging biomarkers that reveal neurobiological endophenotype possibly responsible for the degree of response to neuro-cognitive intervention would help to individualize and efficiently administer cNCI to patients with the highest estimated benefit. Furthermore, the prevention of disease progression and reoccurrence of further psychotic episodes could be further strengthened by multimodal treatment stratification based on individuals' neurobiological characteristics, including neuro-cognitive measurements.