(A-205) Electronic Health Record (EHR) Enhanced Signal Detection Using Tree-Based Scan Statistic Methods

Background: Tree-based scan statistics (TBSS) are data mining methods that screen thousands of hierarchically related health outcomes to detect unsuspected adverse effects of medical products. Traditionally, TBSS have been used to analyze claims data where outcomes are defined via diagnosis codes linked to each other using a tree structure based on the International Classification of Diseases. TBSS have not been previously applied to the rich clinical information in Electronic Health Records (EHR), including laboratory results, vital signs, and clinical notes.

Objectives: We propose a generalization of TBSS that allows the analysis of structured and unstructured EHR data. Outcomes from EHR data include laboratory results and outcomes derived from applying natural language processing (NLP) to clinical notes.

Methods: To integrate EHR data in TBSS, we addressed three main challenges: i) defining incident outcomes with a multipath tree based on the Medical Dictionary for Regulatory Activities (MedDRA, www.meddra.org); ii) designing a novel algorithm for generating the expected distributions of test statistics under the null with an outcome-specific definition of washout; iii) integrating different outcome types, including continuous lab results. We evaluate our method in four scenarios that sequentially add sources of outcome data to the TBSS tree: 1) diagnosis code data, 2) NLP-derived outcomes, 3) binary outcomes derived from laboratory results, and 4) continuous lab results, considering analyses where incident outcomes are identified within inpatient or emergency department versus any care setting. We compared the results of TBSS under these scenarios using Medicare and Medicaid claims data deterministically linked to EHRs for patients receiving care at the Mass General Brigham. We considered a new-user cohort study comparing exposure to second-generation sulfonylureas (SUs, 2820 new initiators) to dipeptidyl peptidase 4 (DPP-4, 985 new initiators) inhibitors in adults with type-2 diabetes. A p-value < 5% is used to define a statistical alert.

Results: No statistical alerts were observed when comparing SUs and DPP-4 users with outcomes derived from diagnosis code data in inpatient or emergency department settings. Adding NLP-derived outcomes resulted in an alert for "Headaches" with p-values of 0.047 and a relative risk (RR) of 6.5. Progressively adding binary and continuous lab results produced the same alert with p-values of 0.045 (RR 6.5) and 0.045 (RR 6.5), respectively. Headaches are a non-specific symptom of hypoglycemia and are clinically plausible in the considered setting. No alerts occurred when incident outcomes were defined using any care setting.

Conclusions: Integrating EHR data in TBSS may be useful for the detection of safety signals.