Prognostic value of machine learning for brain computed tomography as a predictor of neurologic outcomes after cardiac arrest: a systematic review and meta-analysis.
Authors
Affiliations (6)
Affiliations (6)
- Department of Emergency Medicine, Hanyang University College of Medicine, Seoul, Republic of Korea.
- Department of Emergency Medicine, Kangnam Sacred Heart Hospital, Hallym University College of Medicine, 1 Singil-Ro, Yeongdeungpo-Gu, Seoul, 07441, Republic of Korea. [email protected].
- Hallym Biomedical Informatics Convergence Research Center, Kangnam Sacred Heart Hospital, Hallym University College of Medicine, Seoul, Republic of Korea. [email protected].
- Hallym Biomedical Informatics Convergence Research Center, Kangnam Sacred Heart Hospital, Hallym University College of Medicine, Seoul, Republic of Korea.
- Department of Digital Health, Samsung Advanced Institute for Health Science & Technology (SAIHST), Sungkyunkwan University, Seoul, Republic of Korea.
- Department of Emergency Medicine, Kangnam Sacred Heart Hospital, Hallym University College of Medicine, 1 Singil-Ro, Yeongdeungpo-Gu, Seoul, 07441, Republic of Korea.
Abstract
The gray-to-white matter ratio (GWR) on brain computed tomography (CT) is used to predict neurological outcomes after cardiac arrest. Even though automated methods, such as automatic GWR and machine learning, have been compared to manual GWR, the superiority remains unknown. Therefore, we conducted a systematic review and meta-analysis to compare the diagnostic accuracy of these three CT-based methods. We systematically searched MEDLINE, EMBASE, Web of Science, Scopus, and IEEE Xplore and included studies evaluating neurological outcomes using the Cerebral Performance Category Scale. We performed a subgroup analysis to compare machine learning with manual or automatic GWR measurements. The Prediction model Risk of Bias ASsessment Tool was used to assess the risk of bias and applicability. Diagnostic accuracy for predicting poor neurological outcomes was evaluated using the pooled diagnostic odds ratio (DOR) and pooled area under the curve (AUC). In total, 1594 patients from seven observational studies were included. Machine learning showed significantly higher diagnostic accuracy (pooled AUC, 0.813; pooled DOR, 14.02; 95% confidence interval [CI], 6.51-30.18; I<sup>2</sup> = 63.1%) than manual GWR measurement (pooled AUC, 0.755; pooled DOR, 5.16; 95% CI, 3.75-7.08, I<sup>2</sup> = 0%; p = 0.02). Machine learning showed statistically equivalent diagnostic accuracy, although it was numerically lower than automatic GWR measurement (pooled AUC, 0.832; pooled DOR, 11.92, 95% CI, 7.55-18.82; I<sup>2</sup> = 24.3%; p = 0.72) for predicting poor neurological outcomes. Machine learning in brain CT may have significant diagnostic value for predicting poor neurological outcomes in post-cardiac arrest patients. Machine learning may be comparable to automatic GWR measurement and outperform manual GWR measurement in terms of diagnostic accuracy.