Towards artificial intelligence in mental health by improving schizophrenia prediction with multiple brain parcellation ensemble-learning

Abstract

In the literature, there are substantial machine learning attempts to classify schizophrenia based on alterations in resting-state (RS) brain patterns using functional magnetic resonance imaging (fMRI). Most earlier studies modelled patients undergoing treatment, entailing confounding with drug effects on brain activity, and making them less applicable to real-world diagnosis at the point of first medical contact. Further, most studies with classification accuracies >80% are based on small sample datasets, which may be insufficient to capture the heterogeneity of schizophrenia, limiting generalization to unseen cases. In this study, we used RS fMRI data collected from a cohort of antipsychotic drug treatment-naive patients meeting DSM IV criteria for schizophrenia (N = 81) as well as age- and sex-matched healthy controls (N = 93). We present an ensemble model -- EMPaSchiz (read as 'Emphasis'; standing for 'Ensemble algorithm with Multiple Parcellations for Schizophrenia prediction') that stacks predictions from several 'single-source' models, each based on features of regional activity and functional connectivity, over a range of different a priori parcellation schemes. EMPaSchiz yielded a classification accuracy of 87% (vs. chance accuracy of 53%), which out-performs earlier machine learning models built for diagnosing schizophrenia using RS fMRI measures modelled on large samples (N > 100). To our knowledge, EMPaSchiz is first to be reported that has been trained and validated exclusively on data from drug-naive patients diagnosed with schizophrenia. The method relies on a single modality of MRI acquisition and can be readily scaled-up without needing to rebuild parcellation maps from incoming training images.

Fig. 1

Negative correlation between sample size and cross-validated (CV) accuracy for predicting schizophrenia using resting-state brain patterns (studies cited in Table 1)

Fig. 2

Comparison of 5×10-fold cross-validation prediction accuracies for stacked learners in EMPaSchiz model. The comprehensive ensemble of EMPaSchiz “stacked-multi” is shown in red. Stacked-multi shows the best performance and performs significantly better than all other stacked models (all p < 0.05). The dotted line is the majority class baseline predictor. SEM standard error of mean

Fig. 3

Comparison of 5×10-fold cross-validation prediction accuracies for single-source and multi-source models. The comprehensive ensemble model of EMPaSchiz “stacked-multi” is shown in red. (Horizontal dotted line, at 0.53, is the majority class baseline predictor. SEM Standard error of mean)

Fig. 4

Comparison of 5×10-fold cross-validation prediction accuracies for stacked-multi models with various levels of feature selection and PCA

Fig. 5

Key pathological alterations in schizophrenia suggested by top-most reliable features—elevated (red) and suppressed (blue) changes in regional activity. Panels show top 98th percentile of top regional features. a Higher ALFF in right caudate and right superior temporal pole (aal). b Higher ALFF in lateral aspect of left superior temporal gyrus and horizontal ramus of the right lateral sulcus, and lower ALFF in left posterior-dorsal cingulate gyrus (destrieux). c Higher fALFF in left putamen, right caudate and lower fALFF in right anterior cingulum (aal). d Higher ReHo in left superior temporal pole, right inferior temporal gyrus, and lower ReHo in left inferior parietal lobule and right superior temporal gyrus (basc_multiscale_197)

Fig. 6

Key pathological alterations in schizophrenia suggested by top-most reliable features—network edges show elevated (red) and suppressed (blue) changes in functional connectivity. Panels show top 99th percentile of top functional connectivity features using dosenbach and msdl atlases. a Decreased functional connectivity between regions—left ventral frontal cortex and left lateral cerebellum, left occipital and left angular gyrus, left middle insula and right fusiform gyrus, and lastly left post parietal cortex with three nodes namely right frontal gyrus, left parietal, left precentral gyrus. Increased interhemispheric functional connectivity between left superior frontal gyrus and the right anterior insula. b Decreased functional connectivity between regions—striatum and posterior occipital lobe, right intraparietal sulcus and right frontal pole, ventral anterior cingulate cortex and medial default mode network, left temporo-parietal junction and right parietal cortex, right superior temporal sulcus and Broca’s area. Increased functional connectivity between regions—right anterior insula and striatum, right insula and left auditory cortex, and left anterior intraparietal sulcus and posterior occipital lobe

Fig. 7

Schematic representation for performance of learned EMPaSchiz model a and process of learning that model b, c—see main text for explanation

Fig. 8

Schematic representation for evaluation of the model with cross-validation. We use 5×10-fold cross-validation to evaluate each learner (the original EMPaSchiz-Learner, and its six variants). Here, we first divide the entire dataset of 174 subjects into 10 sets; we then use 9/10 of them to train the EMPaSchiz model (see Fig. 7b, c); we then run that model on the remaining 1/10 of the data (see Fig. 7a). We then compute accuracy as the number of correctly labelled instances, over all 10 folds, and use this as an estimate of the score of the learned EMPaSchiz -Performance system. We run this entire process five times—over five different partitionings and compute the overall accuracy of predictions over these 50 train-test splits. Trained models are depicted in green lines and predictions are depicted in red lines