Modelling techniques to assess surgeons' cognitive workload: a systematic review
File(s)
Accepted version
Supporting information
Author(s)
Chen, Billy
Majumdar, Arnab
Boyce, Niki
Escribano, Jose
Type
Journal Article
Abstract
Background: Excessive intraoperative cognitive workload can impair surgical performance and threaten patient safety. Machine learning (ML) offers potential for real-time objective monitoring, yet existing studies vary widely in design, input data, and modelling strategies. This systematic review synthesises current evidence on ML-based approaches to estimate surgeon intraoperative workload.
Methods: Searches of nine databases (inception-March
2025) followed PRISMA 2020. Eligible studies applied ML to model cognitive workload in surgical or simulated
settings. Data were extracted on participants, input
modalities, labelling strategies, preprocessing, modelling approaches, and performance metrics. Quality was assessed using the Mixed Methods Appraisal Tool (MMAT).
Results: Fifteen studies were included. Electroencephalography (EEG, 60%), electrocardiography (40%), and eye-tracking (40%) were used most frequently, with 60% adopting multimodal configurations. Workload was labelled using questionnaires (40%), task difficulty (40%), or hybrid approaches (13%). Most studies (87%) addressed classification tasks, with reported accuracy ranging from 54% to 99.9%. Both classical algorithms (e.g., support vector machines, random forests) and deep learning
architecture (e.g., CNNs, LSTMs) achieved competitive
results. However, all selected studies relied on small cohorts (fewer than 30 participants) in simulated environments, with inconsistent reporting of preprocessing framework.
Conclusion: ML-based cognitive workload modelling is
feasible and shows strong performance. Adoption in clinical practice will require validated labelling frameworks, larger and more diverse cohorts, standardised preprocessing pipelines, and ecologically valid operating room datasets. Sensor practicality, comfort, and acceptability remain critical considerations for real-world deployment.
Methods: Searches of nine databases (inception-March
2025) followed PRISMA 2020. Eligible studies applied ML to model cognitive workload in surgical or simulated
settings. Data were extracted on participants, input
modalities, labelling strategies, preprocessing, modelling approaches, and performance metrics. Quality was assessed using the Mixed Methods Appraisal Tool (MMAT).
Results: Fifteen studies were included. Electroencephalography (EEG, 60%), electrocardiography (40%), and eye-tracking (40%) were used most frequently, with 60% adopting multimodal configurations. Workload was labelled using questionnaires (40%), task difficulty (40%), or hybrid approaches (13%). Most studies (87%) addressed classification tasks, with reported accuracy ranging from 54% to 99.9%. Both classical algorithms (e.g., support vector machines, random forests) and deep learning
architecture (e.g., CNNs, LSTMs) achieved competitive
results. However, all selected studies relied on small cohorts (fewer than 30 participants) in simulated environments, with inconsistent reporting of preprocessing framework.
Conclusion: ML-based cognitive workload modelling is
feasible and shows strong performance. Adoption in clinical practice will require validated labelling frameworks, larger and more diverse cohorts, standardised preprocessing pipelines, and ecologically valid operating room datasets. Sensor practicality, comfort, and acceptability remain critical considerations for real-world deployment.
Date Acceptance
2025-11-29
Citation
Annals of Surgery
ISSN
0003-4932
Publisher
Lippincott, Williams & Wilkins
Journal / Book Title
Annals of Surgery
Copyright Statement
Subject to copyright. This paper is embargoed until publication. Once published the author’s accepted manuscript will be made available under a CC-BY License in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy).
License URL
Publication Status
Accepted