Applied machine learning for stroke differentiation by electrical impedance tomography with realistic numerical...

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Computational Neuroscience

Background:

• Existing stroke differentiation methods face limitations in cost, speed, and mobility.
• Electrical impedance tomography (EIT) offers a potential low-cost, rapid, and mobile alternative.
• EIT imaging's nonlinear and ill-posed nature presents reconstruction challenges, often addressed by combining with machine learning.

Purpose of the Study:

• To investigate the efficacy of EIT combined with machine learning for stroke differentiation.
• To develop and validate a robust computational framework using realistic head models and clinically relevant scenarios.
• To assess the impact of various parameters on classification performance for different stroke types.

Main Methods:

• Development of 135 unique, realistic head models incorporating cerebrospinal fluid and representing normal, hemorrhagic, and ischemic brains.
• Simulation of EIT voltage data from these models under varying signal-to-noise ratios and driving frequencies.
• Application of support vector machines with nested cross-validation and principal component analysis for feature reduction and classification.

Main Results:

• Classifier accuracy at 60 dB SNR: 79.92% ± 10.82% for lesion differentiation, 74.78% ± 3.79% for lesion detection, 77.49% ± 15.90% for bleed detection, and 60.31% ± 3.98% for ischemia detection.
• Achieved 76% feature reduction using PCA, decreasing features from 208 to 50.
• Results were based on 17,280 observations across 3 independent runs with polynomial kernel functions.

Conclusions:

• EIT data combined with machine learning demonstrates significant potential for stroke differentiation.
• Classification accuracy is highly dependent on the specific scenario and stroke type.
• Further development of application-specific classifiers may be necessary to achieve optimal diagnostic accuracy.