Automatic Bayesian classification of healthy controls, bipolar disorder, and schizophrenia using intrinsic connectivity maps from FMRI data

Abstract

We present a method for supervised, automatic, and reliable classification of healthy controls, patients with bipolar disorder, and patients with schizophrenia using brain imaging data. The method uses four supervised classification learning machines trained with a stochastic gradient learning rule based on the minimization of Kullback-Leibler divergence and an optimal model complexity search through posterior probability estimation. Prior to classification, given the high dimensionality of functional MRI (fMRI) data, a dimension reduction stage comprising two steps is performed: first, a one-sample univariate t-test mean-difference T(score) approach is used to reduce the number of significant discriminative functional activated voxels, and then singular value decomposition is performed to further reduce the dimension of the input patterns to a number comparable to the limited number of subjects available for each of the three classes. Experimental results using functional brain imaging (fMRI) data include receiver operation characteristic curves for the three-way classifier with area under curve values around 0.82, 0.89, and 0.90 for healthy control versus nonhealthy, bipolar disorder versus nonbipolar, and schizophrenia patients versus nonschizophrenia binary problems, respectively. The average three-way correct classification rate (CCR) is in the range of 70%-72%, for the test set, remaining close to the estimated Bayesian optimal CCR theoretical upper bound of about 80% , estimated from the one nearest-neighbor classifier over the same data.

Fig. 1

Automatic fMRI Bayesian classification system block diagram for 25 healthy controls (HC, class 1), 15 bipolar disorder (BI, class 2) and 21 schizophrenia (SC, class 3) patients. The tools used as well as the dimension of data are indicated outside the corresponding system block boxes.

Fig. 2

DMN and Temporal lobe group ICA mean spatial maps for healthy controls (HC), bipolar disorder (BI), and schizophrenia (SC) class subjects. Ordered from left to right, and top to bottom: (a) DMN HC, (b) DMN BI, (c) DMN SC, (d) Temporal HC, (e) Temporal BI, and (f) Temporal SC. Images were thresholded (T_score = 4).

Fig. 3

3-way classifier Receiver Operation Characteristics and AUC for the healthy vs. non healthy controls (HC) binary problem (test set) for n_cic = 200 independent random runs with 10 samples in test set, totaling 2000 test samples.

Fig. 4

3-way classifier Receiver Operation Characteristics and AUC for the bipolar vs. non bipolar disorder (BI) binary problem (test set) for n_cic = 200 independent random runs with 10 samples in test set, totaling 2000 test samples.

Fig. 5

3-way classifier Receiver Operation Characteristics and AUC for the schizophrenia vs. non schizophrenia patients (SC) binary problem (test set) for n_cic = 200 independent random runs with 10 samples in test set, totaling 2000 test samples.

Fig. 6

3-way classifier learning machines 3 × 3 confusion matrix including CCR, computed over 2000 test samples: (a) PPMS, (b) CV, (c) AIC and (d) MDL. The estimated class is given in rows and the true class in columns; class 1 healthy controls (HC), class 2 bipolar disorder (BI), class 3 schizophrenia (SC). n_cic = 200 independent random runs with 10 samples in test set, totaling 2000 test samples.