Feature Selection Techniques for a Machine Learning Model to Detect Autonomic Dysreflexia.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Physiological Signal Processing

Background:

• Feature selection is vital for machine learning (ML) model performance, especially in healthcare with limited data.
• Autonomic Dysreflexia (AD) is a serious neurological condition requiring accurate detection in individuals with spinal cord injuries.
• Understanding feature impact and physiological relevance improves ML model efficacy in disease state identification.

Purpose of the Study:

• To explore feature selection methods for enhancing ML model performance in detecting Autonomic Dysreflexia (AD).
• To identify optimal metrics and techniques for AD onset detection using multi-modal physiological data.
• To demonstrate the applicability of these techniques to other healthcare datasets for improved ML development.

Main Methods:

• Utilized a dataset comprising 36 features from electrocardiograms, skin nerve activity, blood pressure, and temperature.
• Applied and evaluated various feature selection techniques to identify the most impactful features for AD detection.
• Trained and compared different ML algorithms, including a 5-layer neural network.

Main Results:

• A 5-layer neural network model achieved 93.4% accuracy in detecting the onset of Autonomic Dysreflexia (AD).
• The best performing model identified five key relevant features from the initial thirty-six.
• The study highlights the effectiveness of critical feature selection in improving diagnostic accuracy.

Conclusions:

• Critical feature selection significantly improves ML model performance for detecting niche disease states like AD, even with small datasets.
• The presented techniques offer a framework for developing more robust and accurate ML algorithms in various healthcare applications.
• Optimized feature selection leads to more efficient and effective diagnostic tools in clinical settings.