Predicting Primary Care Visit Length in Israel: A Machine Learning Approach to Optimize Clinic Operations

Michael Hauzer, Lihi Eisen, Assi Albert Cicurel, Uri Gottlieb

Keywords: machine learning; primary care; visit length prediction; healthcare operations; scheduling optimization; clinical efficiency

Background:

: Inefficient clinic scheduling in primary care contributes to extended wait times, compromising patient satisfaction and physician productivity. While previous research has identified visit length variability as a key factor, traditional prediction methods fail to capture the complex interplay of factors affecting visit duration. Machine learning approaches offer potential solutions for optimizing clinical operations through improved prediction accuracy.

Research questions:

Can machine learning algorithms effectively predict primary care visit length using electronic health record data, and what are the key determinants influencing visit duration in Israel's primary care setting?

Method:

A retrospective cohort study analyzed electronic health record data from Clalit Health Services between January 2021 and August 2023. The dataset included adult patient visits (n=1,500,000) to primary care physicians. Features included socio-demographic data, digital literacy indicators, visit characteristics, medical history, and physician visit length patterns. LightGBM regressor algorithm was employed with 50-100 predictors, using root mean squared error (RMSE) as the primary evaluation metric. Models were trained on 70% of data (2021-mid-2022) and validated on 30% (mid-2022-2023).

Results:

The predictive model demonstrated improved accuracy over the baseline constant prediction (RMSE 5.27 vs 6.09 for frontal visits; 4.08 vs 5.62 for telephone visits). Physician-specific historical visit patterns emerged as the strongest predictors. Mean visit durations were 10.1±6.1 minutes for frontal visits and 6.5±4.3 minutes for telephone consultations. Model performance varied by visit duration category, with better predictions for visits between 5-15 minutes compared to extremely short (<5 minutes) or long (>20 minutes) visits.

Conclusions:

Machine learning approaches can improve visit length prediction accuracy compared to standard scheduling methods, though prediction remains challenging for visits at duration extremes. Findings suggest that physician practice patterns are more influential than patient characteristics in determining visit length, highlighting the importance of provider-specific scheduling strategies.

Points for discussion:

How might these findings inform the development of personalized scheduling systems in primary care?

What role could automated prediction tools play in optimizing clinic workflow without compromising care quality?

How can we balance improved operational efficiency with the need for flexible visit durations in complex cases?