Improving Diagnostics with Deep Forest Applied to Electronic Health Records.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Data Science

Background:

• Electronic Health Records (EHR) are complex, high-dimensional data sources crucial for patient care.
• Identifying implicit correlations within EHR data can enhance treatment and management strategies.
• Existing models face limitations in establishing stable relationships between diverse medical concepts.

Purpose of the Study:

• To introduce Patient Forest, a novel end-to-end approach for learning patient representations from tree-structured EHR data.
• To develop an accurate and reliable classifier for predicting patient readmission and mortality.
• To address the challenge of limited data sources in EHR analysis.

Main Methods:

• Developed Patient Forest, an end-to-end model utilizing statistical features from tree-structured data.
• Applied the model to predict patient readmission and mortality.
• Conducted experiments on MIMIC-III and eICU datasets.
• Performed qualitative evaluation using t-SNE for visualization of learned representations.

Main Results:

• Patient Forest demonstrated superior performance in predicting readmission and mortality compared to existing machine learning models.
• The model showed particular effectiveness when training data were limited.
• t-SNE visualization confirmed the model's capability in learning meaningful EHR representations.

Conclusions:

• Patient Forest offers an effective method for learning patient representations from EHR data.
• The model provides accurate and reliable predictions for readmission and mortality, especially in data-scarce scenarios.
• The approach enhances the utility of EHR data for clinical decision support and research.