Abstract
Purpose :
There is long standing use of low birthweight and gestational age to screen and predict retinopathy of prematurity (ROP). The role of other neonatal comorbidities in ROP severity is unclear. Developing a generalizable predictive model of ROP severity with high sensitivity and specificity is made difficult by the lack of knowledge of the role of these comordibities. This study investigated the hypothesis that comorbidity data from state insurance claims can be used to create an effective predictive model of ROP.
Methods :
We analyzed Rhode Island’s All-Payers Claim Database (HealthFacts RI) which contains insurance claims data from 2011 to 2018. Patients were identified by ICD-10-CM diagnosis codes for ROP stages 0 to 5. Claims pertaining to previous or concurrent comorbidities for these patients were included and those after the last ROP diagnosis were excluded. Data was processed using the computer programming language, Julia (v1.1.0, Lexington, MA). ICD-10-CM codes were categorized by clinical phenotypes (PheCodes). Patients were grouped by ROP stages 0 to 2 and stages 3 to 5. Statistically significant ROP comorbidities were identified and a predictive model was made using PredictMD (v0.32.0, Providence, RI). Through a licensing agreement between the Brown Center for Biomedical Informatics and the Rhode Island Department of Health, an IRB was not required for this project.
Results :
A total of 645 patients met study inclusion criteria. Statistically significant neonate comorbidities based on PheCodes included perinatal disorders of the digestive system, endocrine and metabolic disorders, apnea of newborn, and congenital abnormalities of the respiratory system, among others. A random forest classifier predictive model for ROP severity was created. Model performance metrics include an Area Under the Curve Receiver Operating Characteristic (AUROC) curve of 0.7385, a 95.83% sensitivity, and a 14.59% specificity.
Conclusions :
Our random forest classifier model may be an effective tool to predict ROP stages 3 to 5. Furthermore, the results imply that the features used to create the model are strong predictors of ROP severity. Future work will incorporate more nuanced clinical data into the development of a more precise ROP predictive model. This study highlights the potential use of data science, specifically state insurance claims data, and machine learning in clinical decision making.
This is a 2020 ARVO Annual Meeting abstract.