Sparse learned kernels for interpretable and efficient medical time series processing

Area of Science:

• Medical Signal Processing
• Machine Learning
• Wearable Technology

Background:

• Accurate interpretation of medical time series signals is critical for clinical decisions.
• Deep learning models excel in performance but are computationally intensive and lack interpretability.

Purpose of the Study:

• To propose SMoLK (sparse mixture of learned kernels), an interpretable and efficient architecture for medical time series processing.
• To evaluate SMoLK's performance against larger models in real-world wearable applications.

Main Methods:

• Developed SMoLK, a single-layer sparse neural network using lightweight, flexible kernels.
• Implemented parameter reduction techniques to optimize SMoLK's size and maintain performance.
• Tested SMoLK on photoplethysmography artifact detection and atrial fibrillation detection from electrocardiograms.

Main Results:

• SMoLK achieved performance comparable to models orders of magnitude larger.
• Demonstrated efficiency, robustness, and generalization to new data distributions.
• Validated SMoLK's suitability for real-time applications on low-power devices.

Conclusions:

• SMoLK offers an interpretable and efficient alternative for medical time series analysis.
• The architecture is well-suited for wearable devices and high-stakes clinical decision-making.
• Interpretability of SMoLK aids in understanding and trusting model outputs in critical scenarios.