C-B5-04: An Algorithm for Classifying Potential Cases of Amyotrophic Lateral Sclerosis from Electronic Health Records Using Decision Tree Analysis

Background/Aims: The CDC is developing a national amyotrophic lateral sclerosis (ALS) registry that would be based on electronically available medical records and claims. The HMORN contracted with the CDC to help examine the feasibility of doing this in locations (e.g., HMOs) where complete longitudinal records are more likely to be available. Any final system used by the CDC will require an algorithm to classify putative cases. Our aim was to develop an automated algorithm to classify putative cases of ALS and other motor neuron diseases using electronic health records. Methods: We identified potential cases of ALS by review of inpatient, outpatient, other provider services, mortality and pharmacy data for any evidence of ALS or, more generally, motor neuron disease (MND) from 2001 to 2005 at Kaiser Permanente, Northern California. The records of each individual were reviewed to classify into five categories: probable ALS, not ALS, other MND (not ALS), insufficient information, or unclear. Classification And Regression Tree (CART) analyses were used to create an algorithm based on a comprehensive set of variables that may predict the likelihood of having true ALS (or, conversely, ruling them out). Results: Multiple algorithms were developed with key variables being number times an ALS code appeared in the record, physician-type making diagnosis, alternative diagnoses, mortality records, and length of follow-up. In our best model, we were able to correctly classify 99.6% of those originally determined to be probable ALS. However, there remain a substantial number of individuals classified by review or algorithm in the unclear, insufficient, or other MND categories. In addition, with this analytic approach, the “costs” of misclassification can be modified, and we will present alternative algorithms with a discussion of the benefits and disadvantages of each. Validation analysis results conducted in a second HMO will be presented. Conclusions: An algorithm was developed that appropriately classifies the vast majority of probable ALS cases. However, given the lack of detail in some records and the difficulty in diagnosing some clinical cases, there will be some misclassification when based solely on electronic medical records.