Predicting the recovery of consciousness after severe acute brain injury using resting state networks: a machine learning approach (N3.001)

Abstract

<h3>Objective:</h3> In this work, we evaluate the prognostic utility of resting state functional Magnetic Resonance Imaging (rs-fMRI) in predicting meaningful functional recovery with patients in the intensive care unit (ICU) who have sustained a severe acute brain injury using a machine learning approach. <h3>Background:</h3> Predicting the re-emergence of conscious awareness and subsequent functional recovery for patients in the ICU is a substantial clinical challenge. Rs-fMRI may be of use in this regard by evaluating network connectivity within the brain. These measures may serve as objective markers of cortical integrity in this patient population. <h3>Design/Methods:</h3> A 5.5-minute rs-fMRI sequence was acquired in 25 acutely unresponsive patients in the ICU. We extracted ten canonical resting-state networks from these patients, where spatial correlation to standardized templates indexed the cortical integrity of these networks. The Nearest Centroid machine learning algorithm was used to map spatial correlations across these networks to a patient’s neurologic outcome. A good functional outcome was a score &gt; 3 on the Glasgow Outcome Scale, whereas poor functional outcome was a score ≤ 3. <h3>Results:</h3> This approach could predict good functional outcome (10/12 correctly identified) versus poor (12/15) outcome. The medial visual, lateral visual, occipital pole visual, left frontoparietal, right frontoparietal and default mode networks (DMN) were the strongest predictor of meaningful recovery. Clinical behavioural indicators did not improve the algorithms’ performance but assisted in decreasing variance in performance. <h3>Conclusions:</h3> Rs-fMRI measures predicted the recovery of consciousness and good functional outcome with a higher sensitivity (83%) than traditional prognostic measures. Multiple resting state networks, rather than the previous exclusive examination of the DMN, should be used to inform acute cortical integrity to predict subsequent recovery. This work adds to a growing body of literature demonstrating that both rs-fMRI and machine learning approaches may complement current prognostic tools by identifying patients who will have a good recovery. <b>Disclosure:</b> Mr. Kazazian has nothing to disclose. Mr. Kolisnyk has nothing to disclose. Loretta Norton has nothing to disclose. The institution of Dr. Gofton has received research support from Government of Canada. The institution of Dr. Gofton has received research support from Canadian Institutes of Health Research. The institution of Dr. Gofton has received research support from Academic Medical Organisation of Southwestern Ontario. Dr. Gofton has a non-compensated relationship as a Medical Advisory Board Member with NORSE Institute that is relevant to AAN interests or activities. Dr. Debicki has received personal compensation in the range of $500-$4,999 for serving on a Speakers Bureau for UCB Canada Inc.. The institution of Dr. Debicki has received research support from Academic Medical Association of Southwestern Ontario. The institution of Dr. Debicki has received research support from Western BrainScan. The institution of Dr. Debicki has received research support from Lawson Research. The institution of Dr. Debicki has received research support from Canadian Institutes of Health Research. Dr. Owen has nothing to disclose.