High Classification Accuracy for Schizophrenia with Rest and Task fMRI Data

Department of CSEE, University of Maryland Baltimore County, MD, USA.
Frontiers in Human Neuroscience (Impact Factor: 2.9). 06/2012; 6:145. DOI: 10.3389/fnhum.2012.00145
Source: PubMed

ABSTRACT We present a novel method to extract classification features from functional magnetic resonance imaging (fMRI) data collected at rest or during the performance of a task. By combining a two-level feature identification scheme with kernel principal component analysis (KPCA) and Fisher's linear discriminant analysis (FLD), we achieve high classification rates in discriminating healthy controls from patients with schizophrenia. Experimental results using leave-one-out cross-validation show that features extracted from the default mode network (DMN) lead to a classification accuracy of over 90% in both data sets. Moreover, using a majority vote method that uses multiple features, we achieve a classification accuracy of 98% in auditory oddball (AOD) task and 93% in rest data. Several components, including DMN, temporal, and medial visual regions, are consistently present in the set of features that yield high classification accuracy. The features we have extracted thus show promise to be used as biomarkers for schizophrenia. Results also suggest that there may be different advantages to using resting fMRI data or task fMRI data.


Available from: Tom Eichele, Jun 10, 2015
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: Objective: The diagnostic process for schizophrenia is mainly clinical and has to be performed by an experienced psychiatrist, relying primarily on clinical signs and symptoms. Current neurophysiological measurements can distinguish groups of healthy controls and groups of schizophrenia patients. Individual classification based on neurophysiological measurements mostly shows moderate accuracy. We wanted to examine whether it is possible to distinguish controls and patients individually with a good accuracy. To this end we used a combination of features extracted from the auditory and visual P300 paradigms and the mismatch negativity paradigm. Methods: We selected 54 patients and 54 controls, matched for age and gender, from the data available at the UPC Kortenberg. The EEG-data were high- and low-pass filtered, epoched and averaged. Features (latencies and amplitudes of component peaks) were extracted from the averaged signals. The resulting dataset was used to train and test classification algorithms. First on separate paradigms and then on all combinations, we applied Naive Bayes, Support Vector Machine and Derision Tree, with two of its improvements: Adaboost and Random Forest Results: For at least two classifiers the performance increased significantly by combining paradigms compared to single paradigms. The classification accuracy increased from at best 79.8% when trained on features from single paradigms, to 84.7% when trained on features from all three paradigms. Conclusion: A combination of features originating from three evoked potential paradigms allowed us to accurately classify individual subjects as either control or patient. Classification accuracy was mostly above 80% for the machine learners evaluated in this study and close to 85% at best.
    Journal of the Neurological Sciences 10/2014; 347(1-2). DOI:10.1016/j.jns.2014.10.015 · 2.26 Impact Factor
  • [Show abstract] [Hide abstract]
    ABSTRACT: A functional connectivity (FC) analysis from resting-state functional MRI (rsfMRI) is gaining its popularity toward the clinical application such as diagnosis of neuropsychiatric disease. To delineate the brain networks from rsfMRI data, non-neuronal components including head motions and physiological artifacts mainly observed in cerebrospinal fluid (CSF), white matter (WM) along with a global brain signal have been regarded as nuisance variables in calculating the FC level. However, it is still unclear how the non-neuronal components can affect the performance toward diagnosis of neuropsychiatric disease. In this study, a systematic comparison of classification performance of schizophrenia patients was provided employing the partial correlation coefficients (CCs) as feature elements. Pair-wise partial CCs were calculated between brain regions, in which six combinatorial sets of nuisance variables were considered. The partial CCs were used as candidate feature elements followed by feature selection based on the statistical significance test between two groups in the training set. Once a linear support vector machine was trained using the selected features from the training set, the classification performance was evaluated using the features from the test set (i.e. leaveone- out cross validation scheme). From the results, the error rate using all non-neuronal components as nuisance variables (12.4%) was significantly lower than those using remaining combination of non-neuronal components as nuisance variables (13.8 ~ 20.0%). In conclusion, the non-neuronal components substantially degraded the automated diagnosis performance, which supports our hypothesis that the non-neuronal components are crucial in controlling the automated diagnosis performance of the neuropsychiatric disease using an fMRI modality.
    Proceedings of SPIE - The International Society for Optical Engineering 02/2014; DOI:10.1117/12.2042980 · 0.20 Impact Factor
  • Source
    [Show abstract] [Hide abstract]
    ABSTRACT: OBJECTIVE Developing categorical diagnoses that have biological meaning within the clinical phenotype of psychosis (schizophrenia, schizoaffective disorder, and bipolar I disorder with psychosis) is as important for developing targeted treatments as for nosological goals. The Bipolar and Schizophrenia Network on Intermediate Phenotypes (B-SNIP) was formed to examine a broad array of intermediate phenotypes across psychotic disorders and to test the hypothesis that intermediate phenotype characteristics are homogeneous within phenomenologically derived DSM-IV diagnoses. METHOD The consortium recruited 933 stable probands with schizophrenia, schizoaffective disorder, or psychotic bipolar I disorder, 1,055 of their first-degree relatives, and 459 healthy comparison subjects for clinical characterization and dense phenotyping. Clinical, psychosocial, and family characteristics were contrasted. RESULTS All proband groups showed lower psychosocial functioning than the relatives or comparison group. On average, schizophrenia probands showed more symptoms and lower psychosocial functioning than probands with psychotic bipolar disorder, but there was considerable overlap in clinical manifestations. The characteristics of schizoaffective disorder were more often similar to schizophrenia than to psychotic bipolar disorder. The rates of lifetime suicide attempts were high across all proband groups, with the highest reported frequencies in the schizoaffective and bipolar groups. Proband family lineages included both families with "pure" psychosis diagnoses and families with mixed schizophrenia-bipolar diagnoses. CONCLUSIONS Symptoms, psychosocial functioning, and familial lineage overlap across the three DSM-IV psychosis diagnoses used in B-SNIP. The comingling of psychosis diagnoses within families suggests overlapping genetic determinants across psychoses. These data provide scant evidence for distinct phenotypic clustering around traditional phenomenological diagnoses.
    American Journal of Psychiatry 07/2013; 170(11). DOI:10.1176/appi.ajp.2013.12101339 · 13.56 Impact Factor