Patient Embeddings From Diagnosis Codes for Health Care Prediction Tasks: Pat2Vec Machine Learning Framework

Area of Science:

• Health Informatics
• Machine Learning
• Computational Medicine

Background:

• Diagnosis codes in healthcare claims and EHRs are crucial for data-driven decision-making.
• Existing methods like binary encoding struggle with high variability, data gaps, and the large number of diagnoses.
• A robust numerical representation of patient diagnosis profiles is essential for machine learning applications in healthcare.

Purpose of the Study:

• To introduce Pat2Vec, a self-supervised machine learning framework for embedding complete patient diagnosis profiles into numerical vectors.
• To leverage natural language processing-inspired techniques for creating compact, real-valued representations of patient health data.

Main Methods:

• Developed an optimal vectorization embedding model using Bayesian optimization on German outpatient claims data (ICD-10 codes).
• Calibrated the model using diverse regression and classification tasks across multiple machine learning algorithms.
• Validated against a baseline binary encoding model using over 10 million patient records (2016-2019), including clustering and 2D visualization of patient vectors.

Main Results:

• Pat2Vec models outperformed the baseline binary encoding model in equal dimensions, demonstrating superior robustness to missing data.
• Significant performance gains were observed, especially in lower dimensions, indicating effective compression of nonlinear information.
• The framework is extensible to integrate additional healthcare data sources and can be applied to existing diagnosis datasets.

Conclusions:

• Pat2Vec offers a powerful tool for improving healthcare quality through personalized prevention and patient surveillance signal detection.
• The data-driven machine learning framework facilitates effective healthcare resource planning by identifying distinct patient subcohorts.
• The publicly shared embedding model enables broader application and research in analyzing patient diagnosis data.