Objective:

To develop a model using administrative claims data to predict Parkinson disease (PD) risk.

Background:

We previously used Current Procedural Terminology (CPT) and ICD-9-CM codes from 2004-2009 Medicare data to build a predictive model to identify incident PD patients in 2009. While the model performed well with an area under the receiver operating characteristic curve (AUC) of 0.86, the conversion from ICD-9-CM to ICD-10-CM in the U.S. requires that this be redeveloped with ICD-10-CM codes.

Design/Methods:

Among beneficiaries age-eligible for Medicare, we used 2014-2018 Medicare data to identify all incident PD patients in 2018 (N=93,238) and 1:4 frequency matched with controls (N=372,952) on year and month of diagnosis. We created predictor variables for demographics (age, sex, race/ethnicity, and smoking), all ICD-10-CM diagnosis codes that occurred from October 2015, and all procedure codes (CPT, Healthcare Common Procedure Coding System, and ICD-10 procedure codes) from 2014 to the beneficiary diagnosis or reference date. We randomly divided the dataset into 80% training and 20% validation, then used elastic net algorithm to determine which ICD-10-CM and procedure codes best predicted a PD case. With the training data, we used a five-fold cross validation to select the two hyper-parameters (alpha and lambda), and the predictive accuracy of the model was evaluated using AUC and 95% confidence intervals (CIs). We also performed tests for multicollinearity and removed collinear variables and sex-specific codes.

Results:

Using only demographic variables, the AUC was 0.659 (95% CI: 0.657-0.661). In our elastic net model with diagnosis and procedure codes, the best alpha was 0.5 with 481 predictors, resulting in an AUC in the validation set of 0.871 (95% CI: 0.868-0.874).

Conclusions: