Presentation Authors: Hsinhsiao Wang*, Dylan Cahill, Julia Finkelstein, Julie Campbell, Carlos Estrada, Boston, MA
Introduction: Despite variable clinical concerns and visit types, ambulatory clinic appointments are typically scheduled in fixed increments, resulting in suboptimal time utilization. Advanced analytics are being utilized to improve process efficiency in many industries, but are rarely applied to address operational challenges in ambulatory clinical settings. We sought to develop a machine learning (ML) model that predicts the time pediatric urologists require to see patients and to apply this model to create a more efficient clinic schedule template.
Methods: We prospectively collected data from 294 clinic MD visits from January-April 2018. Variables collected included: demographics, physician, visit date/time, new/return visit, urologic surgical history, 1st post-op visit, same-day testing/imaging, and diagnosis. Timestamps were recorded at patient check-in, MD in and out of room, and final check-out. The primary outcome was MD face time, defined as time of MD-out minus time MD-in. Univariate analysis was performed between predictors and the outcome. Data were split into train/test in 4:1 ratio. Two separate models were created for new and return visits with 1 extra variable (time since last visit) for returns. Gradient boosted machine was chosen as prediction algorithm. Hyperparameters were tuned using 5-fold cross validations within train set. Two out-of-sample clinic days were chosen to compare the patient wait time between classic (fixed 15 min) and ML strategies. Simulation of patient punctuality was performed 1000 times in each clinic day.
Results: 256 visits (113 new/143 return) were included in the final analysis. Mean age at visit was 6.47 years. In univariate analysis, longer visits were significantly associated with new patients (p < 0.01), same-day testing (p < 0.01), older patients, and diagnoses such as voiding dysfunction, scrotal complaints, and neurogenic bladder. Conversely, morning clinic, previous urologic surgery (p < 0.01), 1st post-op visit (p < 0.01), and diagnoses such as penile complaints, undescended testis, and hydrocele were associated with shorter visits. Our ML model predicted MD in-room time accurately to within 3.5+/-2.3 minutes. Using the prediction model on 1000 simulated visit days with random patient punctuality, we were able to reduce the wait time by 24-54% (>90% simulations MLâ‰¤ classic wait time).
Conclusions: Pediatric urologist's face time can be accurately predicted with machine learning models. This insight can be incorporated into a robust dynamic scheduling model to minimize patient wait time, maximize face time, increase clinical efficiency, and likely improve family satisfaction.
Source of Funding: Dr. Wang is supported by AHRQ grant # T32-HS000063-24. The funder had no role the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to subm