A hybrid adaptive approach for instance transfer learning with dynamic and imbalanced data.

Area of Science:

• Clinical Informatics
• Machine Learning
• Health Services Research

Background:

• Machine learning models in healthcare face performance drift due to evolving clinical practices and data distributions.
• Outdated predictive models pose risks and can lead to adverse patient outcomes and financial losses.
• Existing transfer learning algorithms like TrAdaBoost have limitations in handling domain similarity and class imbalance.

Purpose of the Study:

• To propose a novel Hybrid Adaptive Boosting (HA-Boost) approach for transfer learning in clinical risk prediction.
• To address limitations of TrAdaBoost by incorporating domain similarity and class imbalance adaptation mechanisms.
• To evaluate HA-Boost's effectiveness in predicting hospital-acquired acute kidney injury using longitudinal electronic health records.

Main Methods:

• Developed HA-Boost, a transfer learning algorithm with domain similarity-based and class imbalance-based adaptation.
• Utilized real-world longitudinal electronic health records data for predicting hospital-acquired acute kidney injury.
• Compared HA-Boost performance against baseline models over a 7-year period.

Main Results:

• HA-Boost demonstrated stable and superior performance compared to baseline methods.
• The proposed method achieved higher accuracy in terms of Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC).
• Effectiveness was consistent across a 7-year time span, indicating robustness in dynamic environments.

Conclusions:

• Transfer learning, particularly the HA-Boost approach, is an effective strategy for updating clinical risk prediction models.
• HA-Boost successfully mitigates model performance drift in dynamic healthcare data environments.
• The findings confirm the value of adaptive model updating for maintaining reliable clinical decision support.