Machine Learning Analytics of Resting-State Functional Connectivity Predicts Survival Outcomes of Glioblastoma Multiforme Patients

Abstract

Glioblastoma multiforme (GBM) is the most frequently occurring brain malignancy. Due to its poor prognosis with currently available treatments, there is a pressing need for easily accessible, non-invasive techniques to help inform pre-treatment planning, patient counseling, and improve outcomes. In this study we determined the feasibility of resting-state functional connectivity (rsFC) to classify GBM patients into short-term and long-term survival groups with respect to reported median survival (14.6 months). We used a support vector machine with rsFC between regions of interest as predictive features. We employed a novel hybrid feature selection method whereby features were first filtered using correlations between rsFC and OS, and then using the established method of recursive feature elimination (RFE) to select the optimal feature subset. Leave-one-subject-out cross-validation evaluated the performance of models. Classification between short- and long-term survival accuracy was 71.9%. Sensitivity and specificity were 77.1 and 65.5%, respectively. The area under the receiver operating characteristic curve was 0.752 (95% CI, 0.62-0.88). These findings suggest that highly specific features of rsFC may predict GBM survival. Taken together, the findings of this study support that resting-state fMRI and machine learning analytics could enable a radiomic biomarker for GBM, augmenting care and planning for individual patients.

Keywords: biomarker; brain tumor; classification; overall survival; resting state functional connectivity; short and long-term survival; support vector machine.

Figure 1

Overview of machine learning. Step 1: the rsFC matrix of size 288 × 288 was computed for each patient. The features were rsFC between pairs of ROIs (for example, the correlation between ROI X and ROI Y represented by dots in step 1). Step 2: feature selection was performed in two steps using only the lower triangular rsFC matrix to exclude self and symmetric connections. First (i), features with correlation with patients' days of overall survival OS (p < 0.05, uncorrected) were selected for heuristic filtering. Then (ii), recursive feature elimination (RFE) was used to select the best predicting feature subset. Step 3: hyperparameter optimization and final SVM model training was performed. Step 4: the trained model from the cross-validation fold was tested against the single held-out subject. Steps 1–4 were repeated for each fold of leave-one-out cross-validation. Step 5: following all cross-validation folds, accuracy, AUC, sensitivity, and specificity for the full sample set were computed.

Figure 2

Heatmaps showing the distribution of tumor density, defined by contrast-enhanced (CE) T1w boundaries, in the full sample of 64 patients.

Figure 3

Correlation between rsFC (Fisher's z transformed pairwise ROI correlation) and OS. The horizontal and vertical axis of the plot is the ROI number sorted by the resting-state network. The surface color represents the correlation between OS and ROI-to-ROI rsFC. Correlation strength between ROI pair and OS is represented by colorbar. AUD, auditory; CO, cingulo-opercular; DMN, default-mode network; DAN, dorsal attention network; FPN, frontoparietal network; MTL, medial temporal lobe; PMe, parietomedial network; REW, reward network; SAL, salience network; SMD, somatomotor dorsal network; SML, somatomotor lateral network; VAN, ventral attention network; VIS, visual network; and Cereb, cerebellum regions (all).

Figure 4

ROC analysis for the STS vs. LTS classification.

Figure 5

The top-60 frequently selected features are shown in circle. (A) Red connection (line) indicates the positive correlation of connections (i.e., rsFC) with OS and that of negatively correlated connections are in blue (B). The color of the outer sphere represents the resting-state network (RSN) that the ROIs belongs to. AUD, auditory; CO, cingulo-opercular network; DMN, default-mode network; FPN, frontoparietal network; MTL, medial temporal lobe; REW, reward network; SAL, salience network; SMD, somatomotor dorsal network; SML, somatomotor lateral network; VAN, ventral attention network; VIS, visual network; and Cereb, cerebellum regions (all).