OBJECTIVES: Patient-reported outcomes (PRO) measures, when used in routine clinical care, have been shown to benefit physician satisfaction and patient outcomes¹. Implementation into clinical work flows, however, can be challenging. We examine here Mini-Mental Status Exam (MMSE) usage in an ambulatory Electronic Medical Record (EMR) in Alzheimer’s patients, including evaluating predictors of cognitive decline².

METHODS: This study was conducted a nationally representative ambulatory EMR dataset from Jan 1, 2014 to Dec 31, 2018, in OMOP Common Data Model, v5. Analyses were conducted in SHYFT Quantum v6.7.0. Patients had either 2 diagnosis codes for Alzheimer’s disease, at least 1 Alzheimer’s treatment prescription, and at least 180 days continuous activity pre-index. Descriptive statistics were summarized for baseline demographics, treatments and MMSE scores. Changes in MMSE over time, as well as Kaplan-Meier time-to-decline were calculated. Prediction of MMSE-decline was compared using multiple models, including Linear Regression, Gaussian Naïve Bayes, Multi-level Perceptron, K-Nearest Neighbor, LDA, SVM, Decision Tree, Random Forest, Gradient Boosted, and XG Boosted Regressions. Model performance was assessed by standard metrics such as RMSE or MAE. A 3:1 training/testing split was employed using cross-validation scoring with a K-fold of 6 .

RESULTS: MMSE Score change was captured in less than a fifth of available patients. Few showed continued assessment after demonstrated decline, hampering ability to assess decline after diagnosis or treatment initiation. Among patients with at least two MMSE scores, patient demographics, pre-existing dementia and the earliest MMSE score were statistically significant predictors of MMSE decline. Common Alzheimer’s comorbidities³, such as major depressive disorder and bipolar disorder, were not statistically significant predictors of MMSE score change. Best model performance was achieved with XG Boosted Regressor.

CONCLUSIONS: Despite the sparseness of PRO assessment, data capture was sufficient to drive predictive modeling. Together, these findings highlight both the difficulties and opportunities of PRO integration into real-world clinical care.