Multi-dimensional patient acuity estimation with longitudinal EHR tokenization and flexible transformer networks.

Area of Science:

• Artificial Intelligence
• Clinical Informatics
• Natural Language Processing

Background:

• Transformer models excel in Natural Language Processing (NLP) but are underexplored for Electronic Health Record (EHR) data.
• Traditional NLP models like recurrent and convolutional neural networks have shown utility in health forecasting.
• Existing methods lack health-specific tokenization for heterogeneous EHR data.

Purpose of the Study:

• To propose a dynamic tokenization method for discrete and continuous EHR data.
• To develop a transformer-based classifier integrating temporal patient measurements.
• To evaluate the clinical AI framework for predicting patient acuity and outcomes.

Main Methods:

• Developed a dynamic tokenization scheme for diverse EHR data types.
• Implemented a transformer classifier with a joint embedding space for temporal data integration.
• Conducted multi-task learning for simultaneous prediction of six mortality and readmission outcomes.

Main Results:

• The proposed longitudinal EHR tokenization and transformer modeling achieved higher prediction accuracy than baseline machine learning models.
• Demonstrated the feasibility of the clinical AI framework for patient acuity estimation.
• Showcased improved performance in predicting mortality and readmission.

Conclusions:

• The novel tokenization and transformer approach enhances clinical outcome prediction from EHR data.
• Suggests potential for multimodal data integration and advanced clinical decision support tools.
• Highlights the value of tailored transformer networks in healthcare.