Healthcare innovations empowered by discriminative and generative artificial intelligence

Abstract

Artificial Intelligence (AI) is reshaping healthcare through discriminative and generative models, among other paradigms. Discriminative models classify data by learning boundaries, while generative models capture underlying distributions to produce new data. This dissertation advances both directions with novel computational frameworks. For discriminative models, I developed NACHOS (Nested and Automated Cross-validation with Hyperparameter Optimization on Supercomputers), an algorithm that integrates nested cross-validation, automated hyperparameter optimization, and high-performance computing (HPC) to more reliably estimate test performance and quantify uncertainty at scale. A companion algorithm, DACHOS (Deployment with Automated Cross-validation and Hyperparameter Optimization using Supercomputing) supports reproducible deployment by retraining the best-selected configuration on full datasets. To validate the NACHOS framework, I applied it to real-time medical imaging with optical coherence tomography (OCT). In the percutaneous nephrostomy study, ResNet50 achieved a mean accuracy of 82.6% ± 3.0% under NACHOS evaluation, highlighting that deeper architectures more effectively capture fine-grained tissue differences. In the epidural anesthesia study, a biologically informed sequential binary pipeline outperformed a conventional multi-class approach, reaching 96.7% ± 1.3% overall accuracy and >99% precision for detecting the epidural space—a clinically critical milestone. Together, these case studies demonstrate that rigorous evaluation strategies and thoughtful task reformulation can substantially improve robustness and safety in time-critical decision support. For generative models, I applied large language models (LLMs) to enhance a smoking cessation app whose limited corpus of ~900 messages risked repetitiveness. I systematically compared five open-source LLMs and ChatGPT, optimizing prompts and decoding strategies using perplexity and LIWC analysis, and validated outcomes through expert counselor evaluations. The results indicated that ChatGPT and the larger language models consistently generated the most credible and persuasive content. Overall, this dissertation contributes computational frameworks that advance discriminative and generative AI in healthcare, demonstrating novel pipelines for performance estimation with uncertainty quantification on HPC systems in medical imaging and message generation for patient support systems.