Occlusion-Based Explanations in Deep Recurrent Models for Biomedical Signals.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Data Science

Background:

• The biomedical field generates vast amounts of sequential data, including biosignals like blood pressure and brain activity.
• Machine learning (ML) is increasingly used for predictive analysis of biosignals, but model opacity hinders trust in high-stakes decisions like clinical diagnosis.

Purpose of the Study:

• To develop a model-agnostic explanation method for ML models analyzing time-series biosignals.
• To enhance trust and adoption of AI in clinical settings by providing interpretable predictions.

Main Methods:

• An occlusion-based explanation method was developed to identify input influence on ML model predictions.
• The method is specifically designed for time-series data and recurrent neural networks (RNNs).

Main Results:

• The approach provides two types of explanations: one for technical experts and another for physicians.
• Effectiveness was demonstrated across various physiological datasets for both classification and regression tasks.

Conclusions:

• The proposed explanation method enhances the interpretability of ML models for biosignal analysis.
• This facilitates informed decision-making for physicians and aids in ML model verification by experts.