Using model explanations to guide deep learning models towards consistent explanations for EHR data.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Science

Background:

• Deep learning (DL) models can produce inconsistent explanations when trained with varying hyperparameters.
• This lack of explainability hinders DL adoption in sensitive sectors like healthcare and finance.
• Electronic Health Records (EHR) data analysis requires high transparency.

Purpose of the Study:

• To analyze explanation inconsistency in deep learning models across various datasets, with a focus on EHR data.
• To propose a novel deep learning ensemble architecture designed to enhance explanation consistency.
• To quantify the improvement in explanation consistency achieved by the proposed method.

Main Methods:

• Evaluation of explanation inconsistency on six tabular datasets, including EHR data.
• Development of a novel deep learning ensemble architecture.
• Training sub-models within the ensemble to generate consistent explanations.
• Quantitative assessment of explanation consistency improvement.

Main Results:

• Explanation consistency improved by up to 315% on MIMIC-IV dataset (from 0.02433 to 0.1011).
• Average improvement in explanation consistency was 124% across datasets (e.g., 0.12282 to 0.4450 on BCW dataset).
• The proposed ensemble architecture effectively addresses explanation inconsistency.

Conclusions:

• The novel deep learning ensemble architecture significantly enhances explanation consistency.
• Improved explainability can facilitate wider industrial adoption of DL, particularly in healthcare.
• Results provide a foundation for future research in reliable and interpretable AI.