Hybrid Cross-Domain Transfer Learning for Large-Scale Healthcare Analytics in Low-Resource Settings

Abstract

In recent years, Deep Learning (DL) has shown great promise in healthcare analytics, yet its effectiveness is often constrained by the scarcity of labeled data in low-resource clinical settings. This study proposes a novel cross-domain transfer learning framework to improve in-hospital mortality prediction by leveraging knowledge from data-rich environments and adapting it to under-resourced domains. Using the publicly available MIMIC-IV ICU dataset, we simulate domain shifts by segregating patient groups based on demographics and care units to mimic real-world disparities. Our approach begins by training a deep neural network employing a bidirectional LSTM with clinical embeddings on the high-resource source domain. Subsequently, knowledge is transferred to the low-resource target domain through fine-tuning combined with Domain-Adversarial Neural Networks (DANN) and Maximum Mean Discrepancy (MMD) techniques for effective domain adaptation and distribution alignment. The model performs binary classification of mortality risk by outputting a probability score, which is thresholded at 0.5 to categorize patients into high- or low-risk groups. Performance is evaluated using accuracy, precision, recall, and F1-score, complemented by interpretability analyses through SHAP values that highlight important clinical features. Experimental results demonstrate that the proposed framework significantly outperforms baseline models trained only on limited target data, yielding substantial improvements in predictive accuracy and robustness across domains. This research underscores the potential of domain-adaptive deep learning to provide equitable and reliable clinical decision support in resource-constrained healthcare settings.

Ähnliche Arbeiten