Quantifying Uncertainty in Biomedical and Machine Learning Models Using Artificial Sampling and Inversion Techniques
Year: 2026
Type: dissertation
Abstract: Modern biomedical and machine-learning problems increasingly require reliable uncertainty quantification when relevant information about the data is restricted or only partially observed, or when the inferential targets are discrete or non-numeric. In such settings, standard large-sample approximations may be difficult to apply. This dissertation develops finite-sample methods for uncertainty quantification across three problems, (i) inference for population rankings, (ii) prediction risk control in federated learning and (iii) causal hypothesis testing. First, this dissertation develops a general framework for constructing joint confidence sets for population ranks. Rank parameters are discrete and can be highly sensitive to sampling variability, particularly when populations are closely related, making standard asymptotic inference difficult. The proposed framework provides finite-sample coverage guarantees for true ranks across a broad class of statistical and machine-learning models. Its performance is studied through simulations and applications involving quantile based comparisons, Plackett--Luce models, EPL 2023 soccer team rankings, and hospital performance. Second, a framework for the control of prediction risk is proposed in federated learning, where calibration data is distributed between local servers and cannot be shared directly due to privacy or regulatory constraints. The proposed approach requires only limited information from individual servers while providing global risk control across the federated system, both in expectation and with high probability. The framework accommodates three federated data structures, (i) independent and identically distributed samples across local servers, (ii) partially exchangeable samples across servers, and (iii) settings with overlapping samples. Applications to biomedical prediction, image classification, recommendation systems, and hallucination control in large-language models demonstrate its broad applicability. Third, a Bayesian--Frequentist framework is presented for causal hypothesis testing that incorporates prior information while retaining frequentist error guarantees. The framework provides finite-sample Type I error control under the specified working models and addresses three causal hypotheses: the Fisher sharp null, the finite-population weak null, and the superpopulation average-treatment-effect null. Together, these contributions address uncertainty arising from discrete inferential targets, decentralized data, and partially observed causal outcomes, providing finite-sample guarantees tailored to each setting.
DOI: 10.7282/t3-6p3p-q239
Author: Chandra, Onrina
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235001542
