(132) Validated Algorithms to Identify Non-fatal Strokes using Uniform Objective Criteria: A Pooled Clinical Trial Data Analysis

Publication Number: 793

Background: In the claims databases, algorithms to identify non-fatal stroke usually use the primary diagnosis in inpatient (INP) claims but have varying sensitivity and positive predictive value (PPV). Clinical trial data (CTD) is the gold standard for clinical research. CTD typically collects information on patient characteristics, treatments, health resource utilization, and outcomes and is a rich source to evaluate strategies to improve the accuracy of algorithms to identify clinical events.

Objectives: This study aimed to utilize pooled CTD to develop and validate algorithms to identify non-fatal strokes (ischemic stroke, hemorrhagic stroke, and undetermined stroke) by adapting the objective criteria to identify a stroke.

Methods: Using anonymized phase III CTD from the Medidata Enterprise Data Store, we pooled patients over 40 with T2DM, cardiovascular (CV) diseases or risk for CV diseases across trials. Rule-based algorithms were developed to identify the occurrence and the date of non-fatal stroke using objective criteria from the literature. Algorithm 1 (A1) only used the presence of brain CT/MRI investigations to identify a stroke. A2 required CT/MRI and all-cause INP ±7 days of each other. A3 required CT/MRI and either INP ±7 days, stroke-related signs and symptoms 31 days prior, or treatments 7 days after CT/MRI. The source of truth was the adjudicated events in CTD, and the algorithm’s accuracy was assessed using PPV and sensitivity. The algorithms were considered to have accurately identified stroke if the identified event was within 15 days of the adjudicated event.

Results: A total of 40,866 patients met the study criteria, of which 4.0% of patients had ≥1 stroke (1,765 stroke events) during the CT follow-up (median: 1.55 years). In summary, all the algorithms had high sensitivity, and PPV {A1 (89.6% [95% CI: 88.2%, 91.0%], 71.5% [69.6%, 73.4%]), A2 (80.1% [78.2%, 82.0%], 76.3% [74.4%, 78.2%]), and A3 (82.0% [80.2%, 83.8%], 75.7% [73.8%, 77.6%])}. Adding all-cause INP increased the PPV at the cost of sensitivity, and adding signs, symptoms, or treatments had little impact on the two statistics.

Conclusions: This study shows the usefulness of using pooled CTD to develop algorithms to identify clinical events in the RWD. High algorithm sensitivities indicate the importance of CT/MRI in stroke identification. Lower PPV may be due to the absence of a primary diagnosis in CTD, and including it may increase algorithm’s PPV when applied to claims data. Signs, symptoms, and treatments of stroke are reimbursable events and will likely improve the algorithm's accuracy in the claims data. The prevalence of stroke in CTD was higher than in the general population, and further research is needed to assess their generalizability.