Applying instance-based techniques to prediction of final outcome in acute stroke

Area of Science:

• Neurology
• Medical Imaging
• Machine Learning

Background:

• Acute cerebral stroke is a leading cause of death and disability.
• Thrombolysis is the primary treatment for ischemic stroke but carries bleeding risks.
• Predicting final infarct extent is crucial for guiding thrombolysis decisions.

Purpose of the Study:

• To investigate the applicability of instance-based (IB) learning methods for predicting final infarct size in acute stroke.
• To assess the performance of different IB implementations in this prediction task.

Main Methods:

• Magnetic resonance imaging data from 14 acute stroke patients were analyzed.
• Three IB methods were explored: k-Nearest Neighbors (k-NN), Gaussian weighted, and constant radius search.
• Performance was evaluated using Receiver Operating Characteristic (ROC) analysis and Area Under the Curve (AUC).

Main Results:

• k-NN (k=164, AUC=0.814) and Gaussian weighted (sigma=0.17, AUC=0.813) IB methods performed similarly and significantly better than constant radius search (R=0.28, AUC=0.809).
• Qualitative analysis revealed distinct clustering of non-infarcted tissue but dispersed patterns for infarcted tissue.
• No stringent boundary was found between infarcted and non-infarcted areas in the feature space.

Conclusions:

• Instance-based methods are potentially advantageous for predicting final infarct extent in acute stroke.
• Further research is needed to refine these methods for clinical application in stroke management.