Can Attention Be Used to Explain EHR-Based Mortality Prediction Tasks: A Case Study on Hemorrhagic Stroke

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Stroke is a leading cause of death and disability.
• Existing predictive models (APACHE II, SAPS III) lack accuracy and interpretability.
• Early prediction is crucial for risk mitigation.

Purpose of the Study:

• Develop an interpretable, attention-based transformer model for early stroke mortality prediction.
• Enhance model fidelity and interpretability for better clinical understanding.
• Compare the novel model's performance against traditional methods.

Main Methods:

• Designed an attention-based transformer model for stroke mortality prediction.
• Utilized Shapley values and attention scores to evaluate model fidelity and interpretability.
• Compared performance metrics against established scoring systems.

Main Results:

• The proposed transformer model offers improved accuracy and interpretability.
• Feature importance was successfully derived, enhancing model explainability.
• Demonstrated potential to outperform traditional stroke risk assessment tools.

Conclusions:

• Interpretable AI models can significantly advance early stroke mortality prediction.
• The attention-based transformer model provides a promising, explainable alternative for clinical decision support.
• Further research can refine AI-driven stroke risk stratification.