Abstract: Understanding the progression from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) is crucial, and electronic health records (EHRs) offer a rich longitudinal dataset to study this transition. However, leveraging EHR data poses challenges due to inconsistencies in diagnostic coding practices, missing data, and the inherent subjectivity involved in MCI and AD diagnoses. This paper critically examines methods for constructing cohorts to analyze MCI-to-AD progression using EHRs from the Fairview Health System.

Varying inclusion criteria from published studies were applied to generate three distinct cohorts (A, B, and C), with increasing stringency in confirming MCI onset. Analyses revealed substantial impacts of cohort definitions on sample size, overall progression rates, and the effects of key predictors like age and initial diagnosis type. Cohort C, with stricter MCI confirmation criteria, yielded a cohort stratified by inherent AD risk but reduced sample size by over 50%.

The results highlight the subjectivity challenges in clinically diagnosing MCI/AD, leading to potential inconsistencies and incompleteness in EHR data capture. Cox regression models demonstrated cohort-specific effects, with initial MCI diagnosis exhibiting a greater impact on AD hazard in cohorts with stricter criteria.

Based on the findings, recommendations are provided for cohort utilization based on research objectives. Cohort A (larger sample, less stringent criteria) may be suitable for predictive modeling requiring extensive data, while Cohort C could facilitate inference on MCI-to-AD progression factors by representing a more confident MCI population.

This study underscores the significant impact of cohort selection criteria on the quality, validity, and generalizability of EHR-based findings for MCI-to-AD progression research. Careful consideration of inclusion criteria is imperative to ensure accurate population representation and reliable inferences from EHR analyses in this domain.