The rapid adoption of large language models (LLMs) in healthcare presents significant opportunities to enhance clinical decision support, patient education, and care coordination. However, concerns related to data privacy, computational demands, and secure deployment remain major barriers to implementation. This project proposes the development of privacy-preserving, lightweight language models specifically designed for healthcare applications to support both providers and patients while maintaining strict data protection standards. The framework will leverage model compression, parameter-efficient fine-tuning, and edge-deployable architectures to create computationally efficient models capable of operating within health system environments without reliance on external cloud-based processing of sensitive data. Privacy-preserving techniques, including federated learning, differential privacy, and secure on-device inference, will ensure compliance with healthcare regulations and institutional governance requirements. For providers, the system will deliver clinical documentation assistance, evidence-based information retrieval, and workflow support. For patients, it will generate accessible, health-literacy–aligned educational content and guided self-management support. Performance will be evaluated using clinical accuracy, bias mitigation, usability, security robustness, and computational efficiency benchmarks in collaboration with healthcare partners. By integrating privacy, efficiency, and clinical reliability into a unified framework, this project will enable scalable, trustworthy AI solutions that strengthen patient engagement, improve provider efficiency, and expand responsible AI adoption across healthcare systems.
