Research Projects

This project aims to develop a trustworthy and interpretable AI framework by integrating advanced generative modeling with risk-controlled Uncertainty Quantification (UQ), tailored for biomedical applications. The focus lies on addressing a fundamental limitation in current AI systems i.e., reliable decision-making under uncertainty, especially when applied to high-dimensional, multimodal clinical inputs such as biomedical images, textual data, and omics. While powerful generative models like diffusion and transformer-based architectures are increasingly used to synthesize or impute biomedical information, they often operate as black-box systems and lack calibrated uncertainty estimates critical for high-stakes applications such as cancer diagnosis, prognosis, and treatment planning. To tackle these challenges, this proposal aims to leverage recent advancements of Conformal Prediction (CP) that enables distribution-free, loss-aware uncertainty quantification. For example, Risk-Controlling Prediction Sets (RCPS), a principled extension of CP that incorporates domain-specific loss control while maintaining rigorous statistical guarantees. The methodology further emphasizes the need for individualized clinical reasoning by designing biologically-informed nonconformity scores and subgroup- or instance-level calibration strategies, thereby going beyond the population-level guarantees typically offered by CP. Drawing upon my prior research experience in uncertainty-aware modeling, including Bayesian ensemble learning and feature relevance estimation using ARD-based Gaussian process regression, the proposed approach aims to build personalized prediction sets that are statistically valid and interpretable. The framework will be evaluated using public biomedical datasets e.g., pan-cancer datasets TCGA, CPTAC, that provide multimodal records, including imaging and omics data. Generative components will be trained to infer missing omics information from visual or contextual inputs, with a special emphasis on interpretability and robustness under incomplete data. Explainable AI (XAI) methods will be integrated to clarify not just the top prediction but also why other plausible alternatives appear in the model’s output set. Importantly, this research holds direct relevance for resource-constrained clinical environments, such as rural and semi-urban healthcare centers in India and other 1developing countries, where access to expensive molecular diagnostics is limited. By offering cost-effective surrogate predictions with quantified reliability, the proposed system can contribute toward democratizing precision medicine, bridging the gap between high-end AI innovation and practical, equitable healthcare delivery.