Authors: Emeneker W, et al.
Cureus 17(9): e91800. doi:10.7759/cureus.91800
This study introduces a generative deep learning model to solve one of anesthesia’s most persistent challenges: shift scheduling. Traditional rule-based scheduling often fails when faced with complex, competing demands such as equitable call distribution, vacation requests, and compliance with institutional rules. Manual or software-assisted systems frequently collapse under these constraints, contributing to inefficiencies, provider dissatisfaction, and burnout.
The Reno-Tahoe Anesthesia group provided a dataset of real-world schedules and 31 codified rules that guided training and testing. A long short-term memory (LSTM) architecture was applied, embedding assignments and learning rules directly from historical schedules rather than requiring explicit codification by experts. The trained model achieved a Matthews Correlation Coefficient of 0.9776 and a balanced accuracy of 0.9531. Generated schedules complied with 29 of 31 rules, demonstrating near-complete rule adoption. Training costs were minimal (<$100 in cloud resources), and the model could produce 3 months of schedules for 20 providers in under one minute.
What You Should Know:
This work demonstrates that deep learning can replicate and even surpass traditional scheduling methods in anesthesia practices. By learning directly from past schedules, the model eliminates the need for extensive rule coding and reduces staff burden. Importantly, it maintains fairness and compliance while dramatically decreasing scheduling time and cost.
Clinical Relevance:
For anesthesia groups facing staffing shortages, rising workload, and increasing provider dissatisfaction, this generative deep learning approach offers a scalable, cost-effective solution. It allows schedules to be produced rapidly, equitably, and flexibly, aligning with real-world practice needs while reducing the risk of burnout linked to scheduling inequities.
Thank you to Cureus for allowing us to use this article.