PARSE: A personalized clinical time-series representation learning framework via abnormal offsets analysis.

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Clinical Decision Support

Background:

• Clinical risk prediction is crucial for patient care, utilizing Electronic Health Records (EHRs).
• Existing deep learning methods often overlook physiological index deviations and stability, limiting prediction performance.

Purpose of the Study:

• To develop a personalized clinical time-series representation learning framework for improved risk prediction.
• To address limitations of current methods by incorporating abnormal physiological index offsets.

Main Methods:

• Proposed PARSE (Personalized clinical time-series representation learning framework via abnormal offsets analysis).
• Extracted temporal features from EHR data.
• Captured abnormal condition features using absolute and relative offsets of physiological indexes.
• Utilized an adaptive fusion module for personalized patient representations.

Main Results:

• Achieved highest F1 scores (48.1% and 40.3%) on in-hospital mortality prediction tasks.
• Outperformed state-of-the-art methods by 2.4% and 6.2% on two datasets.
• Ablation experiments confirmed the contribution of abnormal offsets and the adaptive fusion method.

Conclusions:

• PARSE effectively extracts risk-related information from EHRs.
• Improved personalization of patient representations enhances clinical risk prediction.
• Each component of PARSE independently contributes to improved prediction performance.