Artificial Intelligence-Assisted Rapid On-site Evaluation: A Human Factors-Informed Patient Safety Framework for Cytopathology

Open Access
Article
Conference Proceedings
Authors: Bryan Johnston, Jay Kalra, Heather Ruff, Janine Benoit
Abstract

Artificial intelligence (AI) in cytopathology has largely focused on image classification rather than integration into real-time clinical workflows. Rapid on-site evaluation (ROSE), including endobronchial ultrasound-guided transbronchial needle aspiration and endoscopic ultrasound-guided sampling of pancreatic and other lesions, presents a more complex implementation environment. These procedures involve sequential sampling, specimen preparation, adequacy assessment, multidisciplinary communication, and immediate decisions regarding next steps in procedural management. Errors or delays may contribute to unnecessary needle passes, inadequate material, repeat procedures, and potentially delayed diagnoses. Although emerging AI studies demonstrate potential applications in adequacy assessment and malignant cell detection, current ROSE and cytopathology guidance provides limited direction regarding validation, usability, human oversight, or integration into procedural work. To address this gap, we conducted a structured, literature-informed conceptual synthesis of AI-assisted ROSE, procedural cytopathology, digital cytology, and human-factors research. We used this literature to illustrate a purpose-driven approach: define the clinical purpose, determine what would need to be established to judge whether it was achieved, identify the evidence required, and determine what must therefore remain observable before deployment. Three clinical purposes were examined: confirming representative lesional material, preserving and triaging material for molecular or ancillary testing, and determining whether additional sampling is needed. Drawing on these findings, we suggest that the same broad workflow requires different evidence depending on the clinical purpose and that model accuracy alone does not establish safe performance. We therefore recommend treating AI as a component of the procedural cytopathology work system rather than as a stand-alone image classifier. This approach may help move AI-assisted ROSE from isolated image analysis toward safer and more operationally integrated clinical systems.

Keywords: Artificial Intelligence, Cytopathology, Workflow Integration, Regulation, Prospective Evaluation, Human Factors

DOI: 10.54941/ahfe1008441

Cite this paper
Downloads
0
Visits
1
Download PDF

More from this volume

← Toward Robust Patient-Level Parkinson’s Disease Classification: Evaluating Stability Across Participant Compositions Using Wearable-Sensor and Questionnaire InformationHealth Psychology in the Digital Age: Cyberchondria, Health Anxiety, and Online Health Information-Seeking Behaviors →
View all articles in Human Factors in Design, Engineering, and Computing →