Advisor(s)

Ryan Melvin

Committee Member(s)

John Osborne
Sandeep Bodduluri

Document Type

Thesis

Date of Award

6-18-2026

Degree Name

Master of Science in Biomedical Engineering (MSBME)

School

Joint Health Sciences (Interdisciplinary)

Department

Biomedical Engineering

Abstract

Accurate clinical risk prediction plays an important role in improving patient outcomes, particularly in perioperative settings where early intervention can influence care decisions. Although machine learning has shown promising results using electronic health record (EHR) data, its use in clinical practice is often limited by complex development processes, the need for hyperparameter tuning, and variability in performance across datasets. These challenges are especially relevant in data-scarce clinical settings. This thesis evaluates TabPFN-2.5, a transformer-based tabular foundation model designed to perform classification without dataset-specific training. The model was applied to two real-world perioperative prediction tasks: unplanned care escalation (UCE) following post-anesthesia care unit (PACU) discharge and 30-day postoperative in-hospital mortality. Its performance was compared with logistic regression, XGBoost, and two clinical scorecard models. Across both datasets, TabPFN-2.5 showed comparable and, in some cases, improved performance. On the PACU UCE task, it achieved an AUROC of 0.814, slightly higher than XGBoost and the scorecard model. On the 30-day mortality dataset, it achieved an AUROC of 0.893, similar to the scorecard baseline, despite requiring no hyperparameter tuning and operating under feature constraints. The model also maintained consistent performance in smaller sample settings. These findings suggest that tabular foundation models may offer a practical alternative to traditional machine learning approaches. By reducing the need for extensive preprocessing and tuning, TabPFN-2.5 has the potential to support more accessible clinical decision-making, particularly in settings with limited data or technical resources.

Download

Share

COinS