Using Optimal Feature Selection and Continuous Learning to Implement Efficient Model Arrays for Predicting Daily Clinical Radiology Workload.

Abstract

Rationale and objective: Clinical workload can fluctuate daily in radiology practice. We sought to design, validate, and implement an efficient and sustainable machine learning model to forecast daily clinical image interpretation workload.

Materials and methods: A year of radiology exam volume data at two academic medical centers was analyzed with an optimal feature selection algorithm and several machine learning models, to produce the most accurate and explainable prediction of the next weekday's clinical workload. Continuous learning was used to maintain high model quality over time.

Results: After evaluating several AI models of differing complexity on a large set of 707 workflow features, a continuously learning linear regression model array was selected based on three optimal features: the current number of unread exams, the number of exams scheduled to be performed after 5 pm, and the number of exams scheduled to be performed the next day. The model array had an average R² of 0.83 (IQR 0.13) across the tested radiology divisions; it significantly outperformed trivial estimates and provided an accurate daily prediction pattern. The solution was successfully implemented into an online dashboard, displaying the forecasted clinical volume as a percentile in reference to the past year's daily clinical volume. Retraining the model on a weekly basis using live data resulted in high, and sometimes increased, model quality.

Conclusion: An AI model can be developed and implemented to forecast daily clinical radiology workload, as a practice management tool.