Machine learning and large language models for appendicitis risk estimation and hemoglobinopathy screening

A. Schipper

• Promotor: B. van Ginneken and G. Kusters
• Copromotor: S. Kurstjens and M. Rutten
• Graduation year: 2026
• Radboud University

Abstract

This thesis examines through two representative clinical use cases, whether AI can operate as practical, trustworthy adjuncts to clinical reasoning during the earliest and most uncertain phases of diagnostic decision-making: 1. Hemoglobinopathy screening in the clinical laboratory; 2. Appendicitis risk stratification in the emergency department. Both domains represent common diagnostic challenges where both rule-based (knowledge-based) and AI-based (non-knowledge-based) CDSS exist. However, existing systems are mostly developed in singlecenter settings, often lack external validation and benchmarking against current clinical practice, and are rarely designed to operate in parallel with routine diagnostic workflows, leaving their practical value uncertain. This thesis therefore evaluates: whether AI-based CDSS built on routine data can achieve reliable and generalizable performance across hospitals and patient populations, how they compare with existing rule-based diagnostic tools and physicians' diagnostic performance, and how such systems might operate as components embedded in established workflows without imposing additional administrative or cognitive burden, while providing real-time decision support.