GenHPF: General Healthcare Predictive Framework for Multi-task Multi-source Learning.

Area of Science:

• Computational biology and bioinformatics
• Medical informatics
• Machine learning in healthcare

Background:

• Predictive models in healthcare show promise but face challenges in large-scale application due to data heterogeneity across electronic health records (EHRs).
• Existing algorithms often fail to generalize across different tasks or databases because of variations in data formats and schemas.
• Significant preprocessing and feature engineering are typically required, hindering the widespread adoption of predictive algorithms.

Purpose of the Study:

• To introduce the General Healthcare Predictive Framework (GenHPF), a novel approach designed for broad applicability across diverse EHRs and multiple prediction tasks.
• To address the challenge of data heterogeneity and reduce the need for extensive preprocessing and feature engineering in healthcare predictive modeling.
• To facilitate the scalable deployment and utilization of predictive algorithms in clinical settings.

Main Methods:

• GenHPF converts EHR data into a hierarchical textual representation, resolving heterogeneity in medical codes and schemas while preserving maximal features.
• Multi-task learning experiments were conducted using single-source and multi-source settings on three distinct, publicly available EHR datasets.
• The framework was evaluated on 12 clinically relevant prediction tasks, comparing its performance against baseline models, including those utilizing domain knowledge.

Main Results:

• GenHPF significantly outperformed baseline models in multi-source learning scenarios, achieving average Area Under the Receiver Operating Characteristic Curve (AUROC) improvements of 1.2%P in pooled learning and 2.6%P in transfer learning.
• The framework demonstrated comparable performance to single-EHR-trained models when applied to individual datasets.
• Self-supervised pretraining combined with GenHPF further enhanced performance, yielding a 0.6%P improvement.

Conclusions:

• GenHPF offers a robust framework for multi-task and multi-source learning in healthcare, effectively handling data heterogeneity with minimal preprocessing.
• The proposed method significantly improves the generalization and performance of predictive models across different EHR datasets and tasks.
• GenHPF streamlines the development and deployment pipeline, accelerating the scaling and application of predictive algorithms in healthcare.