GRAPH-ENHANCED MONOTONIC NEURAL NETWORKS FOR HEALTHCARE OUTCOME REGRESSION
Abstract
Having the ability to estimate healthcare outcomes based on patient data is a significant undertaking in clinical decision-making. Although powerful, conventional regression approaches do not always work to model complexizations of nonlinear connection in medical isotope and deep neural networks, and are often structure-insensitive and uninterpretable. The given paper introduces a novel Graph-Enhanced Monotonic Neural Network (GEMNet) model specifically tailored to work with regression of healthcare outcomes on structured Tabular data. GEMNet provides a trade-off between interpretability and predictive attributes through the embedding of graph neural networks (GNNs) to make predictions of inter-feature connection and by enforcing monotonic implicit constraints based on clinical knowledge. The layers based on the graph convolution are tied to the domain-sensitive domain monotonic activation dominated by the model architecture in such a way that directionally consistency is attained with known risk factors (e.g., age, blood pressure, cholesterol). It has been experimented on a variety of real-world medical datasets (including medical cost prediction and cardiovascular risk estimation) demonstrating that GEMNet tends to perform better than other current regressors, including conventional models, multilayer perceptrons and gradient boosting, in terms of mean squared error (MSE) and R-squared. Better still, the model provides us with interpretable attribution of features and it generalizes better depending on the folds of validation. The results reveal the potential of monotonic graph-based neural models as a scaled-up, clinically-based solution to structured healthcare prediction tasks.
Keywords: Healthcare Outcome Prediction, Graph Neural Networks (GNNs), Monotonic Neural Networks, Tabular Data Regression, Clinical Interpretability, Feature Dependency Modeling, Structured Data Learning, Medical Risk Modeling, Deep Learning in Healthcare, Explainable AI (XAI).
