Personalized one-shot local adaptation federated learning for mortality prediction in multi-center Intensive Care Unit

The increasing volumes of electronic healthcare records (EHR) encourage the development of the application and research of machine learning (ML) in digital health. Promoting ML in healthcare based on EHR can enhance health management for increased intelligence, safety, and efficiency. Unlike traditional data-centralized ML methods, federated learning (FL) provides a novel and promising distributed learning scheme to promote ML in multiple healthcare institutions while preserving data privacy. Nevertheless, the EHR datasets preserved by independent institutions differ considerably in both quantity and characteristics, making the overall data non-independently and identically distributed (non-IID). Coupled with effect of its skewness, this non- IID data can cause a significant reduction in model performance, especially prediction accuracy, rendering FL devoid of practical significance. This study explored the problem with an in-hospital mortality prediction task under an actual multi-center Intensive Care Unit (ICU) EHR database that preserves the original non-IID data distributions. It argues that, the mechanism of generating a unified sharing model for all participants in conventional FL is no longer feasible. Instead, Personalized federated learning (PFL) focusing on generating a tailored model for each client is applicable. However, current PFL techniques still face limitations in enhancing model predictive accuracy, including the inadequate performance of the global model as a basis for personalized models, the challenge of integrating global data knowledge into highly personalized models, and the difficulty of generating optimal highly personalized models for all independent clients. Accordingly, this study proposes a PFL method, Personalized One-shot Local Adaptation (POLA), to tackle these problems progressively through a threestep optimization. Step 1 involves obtaining a well-performing global model as a teacher by modifying the baseline FL with a selection criterion and a data estimation strategy. Step 2 enables highly personalized models as students to rebalance global and local data knowledge through knowledge distillation optimizations. Step 3 automatically evolves the best-fitting parameters for the highly personalized model at each center using an adapted genetic algorithm. Ultimately, this typical "FL training + local adaptation" PFL method can enhance model predictive accuracy by automatically generating a highly personalized model for each ICU center, guided by the FL sharing model. To demonstrate its effectiveness, this study experimentally evaluated POLA in diverse multi-center ICU scenarios with varying non-IID skewness. After conducting stepwise experiments and comparing with three benchmark methods - a baseline FL method and two PFL approaches in similar optimization scheme - the results shows that POLA not only systematically optimizes its internal outputs but also outperforms the comparison methods. Overall, it showcases notable outcomes, attaining 8.32% and 0.94% enhancements in ROC AUC accuracy at the 10th and 100th training rounds in slightly non-IID data, and 8.41% and 1.04% improvements in highly non-IID data, compared to the top-performing benchmark method. Individually, it obtains performance gains for the maximum number of independent centers, benefiting 100% of them when compared to baseline FL. As these expected results show, POLA is a promising option for promoting FL in the multi-center ICU scenario.

Text 119935.pdf Download (1MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18489

Actions (login required)

View Item