A Multi-Modal Deep Learning Framework for Healthcare Cost Prediction: Towards More Effective Value-Based Care Implementation

Abstract

Rising healthcare expenditures have become a pressing concern for policymakers and healthcare organizations. In the United States, Medicare spending reached approximately 1.1 trillion dollars in 2024, and projections suggest this trend will continue. While numerous studies have attempted to predict healthcare costs using machine learning, most approaches focus on single data modalities. This limitation prompted us to explore whether integrating multiple data sources could yield better results. This paper proposes a framework combining hierarchical attention networks for processing clinical text, temporal fusion transformers for time-series analysis, and graph neural networks for capturing relational patterns in healthcare delivery. Through extensive experiments on Medicare claims (covering around 12 million beneficiaries), commercial insurance data (8 million members), and electronic health records from 152 hospitals, we found that our approach achieves 94.7% accuracy for identifying the top 10% high-cost patients at a calibrated threshold (AUC=0.958; PR-AUC=0.61). Specifically, the framework reached 94.7% accuracy in cost categorisation, with a mean absolute error of approximately $1,247 for 90-day predictions. What makes this particularly interesting is that we integrated explainability mechanisms directly into the model architecture, rather than adding them afterwards. During a prospective trial with 50,000 patients, we observed a 28% reduction in unnecessary emergency visits and about 19% fewer preventable readmissions. While these results are encouraging, we acknowledge that deployment will face various challenges. Conservative estimates suggest potential Medicare savings could exceed 40 billion dollars annually, though real-world factors will inevitably affect these projections.