Abstract
Purpose :
To build a model to predict visit lengths based on a priori known visit characteristics for academic ophthalmology outpatient clinics and to modify it for pediatric ophthalmology clinic.
Methods :
Using electronic medical record (EHR) data for visits at Oregon Health & Science University (OHSU) Casey Eye Institute between 2/1/2006 and 10/1/2015, we estimated the statistical relationship between visit length and various a priori known visit characteristics for all subspeciality clinics. Length of visits are estimated by a linear model including the following patient predictors:
- average prior visit length; - diagnosis; - travel distance; - subspecialty; - age; - visit type (new/return); - sex
- visit month, year and day of the week; - visited clinic volume
For the pediatric ophthalmology clinic we used EHR data from 1/1/2013 to 12/31/2014. We modified the initial model to look closely at the exam time rather than the whole length of the visit. For this linear model we chose the following predictors:
- age; - average prior exam length; - dilation status at visit; - visit type (new return, pre-op, and post-op); - diagnosis; - volume of clinic day
Results :
For the model including all the subspeciality clinics, average prior visit length, age (especially those > 60), and the indication of a new patient visit appear to have significant increasing effect on visit length. The model has an R-squared of 0.251. It is able to correctly predict 1 out of 4 actual visit lengths, which is considered good for healthcare prediction models. The modified pediatric model has an R-squared of 0.18. Significant predictors of the exam length are age, prior exam length, dilation, visit type, and diagnoses of strabismus, neurological disease and ROP. All of the variables have positive coefficients, except of ROP diagnosis.
Conclusions :
Both the general and modified models identified age, prior visit lengths, and visit type as main predictors of the visit and exam lengths. The modified model had a lower value of R-squared supposedly due to smaller dataset and fewer independent variables. The modified model was consistent with the provider’s experience and observations. While the current models’ predictive strength are typical for healthcare models, we feel they can be improved. We are investigating other factors, such as patient personality, as well as considering other non-linear models. Accurately estimating patient visit lengths can help anticipate clinic loads and improve clinic efficiency.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.