Comparison of Prognostic Performance Between a Machine Learning Model and Manually Measured Grey-White-Matter Ratio...

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Early prediction of neurological outcomes in out-of-hospital cardiac arrest (OHCA) is vital for treatment decisions.
• Machine learning (ML) models and grey-white matter ratio (GWR) from brain CT scans are potential predictors.
• The comparative performance of ML and GWR, and the benefit of incorporating prehospital data, remain unclear.

Purpose of the Study:

• To compare the early predictive performance of an ML model versus GWR for poor neurological outcomes in OHCA patients post-ROSC.
• To evaluate the enhanced predictive ability by combining the ML model with prehospital information.

Main Methods:

• Retrospective study of 143 adult OHCA patients undergoing brain CT within two hours post-ROSC.
• Developed an ML model using Residual Network 101 (ResNet-101) with transfer learning on three CT slice levels.
• Calculated GWR from the same CT slices; endpoint was persistent coma post-ROSC.

Main Results:

• The ML model (AUC: 0.796) and GWR (AUC: 0.821) showed comparable predictive performance for persistent coma (p=0.121).
• A model using prehospital information achieved an AUC of 0.846.
• Combining the ML model score with prehospital information significantly improved prediction to an AUC of 0.905.

Conclusions:

• The ML model offers moderate predictive performance, similar to the conventional GWR method in the early post-ROSC phase.
• Integrating ML predictions with prehospital data substantially enhances the accuracy of predicting neurological outcomes in OHCA patients.