Neurosymbolic Reasoning with Decision Transparency in Clinical Triage Systems: Benchmark Study

Abstract

The integration of artificial intelligence into clinical environments has historically been challenged by the tension between predictive accuracy and decision transparency. Clinical triage systems, which dictate the prioritization of patient care in high-stakes emergency settings, require both high fidelity in their predictive capabilities and rigorous interpretability to foster clinician trust. Purely neural approaches often behave as opaque systems, precluding clinical validation, while traditional symbolic systems lack the robust generalization needed to handle noisy, unstructured clinical data. This study presents a comprehensive benchmark of neurosymbolic systems in clinical triage, evaluating their capacity to harmonize the robust perceptual abilities of deep neural networks with the explicitly verifiable logic of symbolic reasoning. Through extensive experimentation on retrospective electronic health record datasets, we analyze the performance of hybrid neurosymbolic architectures against purely neural and purely symbolic baselines. Our methodology evaluates not only standard classification metrics but also introduces novel quantifiable measures for decision transparency and logical consistency. The results demonstrate that neurosymbolic systems achieve predictive performance comparable to state-of-the-art neural architectures while maintaining the strict, rule-based interpretability characteristic of expert systems. This paper contributes a foundational framework for evaluating hybrid reasoning models in healthcare, ultimately facilitating the deployment of safer and more accountable artificial intelligence in acute care environments.