A Novel Predictive Model for Agitation in Alzheimer’s Dementia Using Administrative Claims Data

OBJECTIVES: Agitation in Alzheimer’s dementia (AAD) is a novel set of debilitating symptoms for which relatively little research and few treatments exist. AAD-specific diagnosis codes were introduced only in August 2022, thus real-world data on this patient population was previously limited. This study developed a predictive model for classifying AAD using administrative claims data on demographics, treatment and health care utilization history.

METHODS: Using Merative Marketscan data from 2015-2022, we conducted a case-control study among n=968 cases newly diagnosed with Alzheimer’s dementia who had confirmed agitation and n=968 controls with Alzheimer’s dementia without agitation, matched on duration of Alzheimer’s dementia. Features included birth year and sex, and presence/absence of each of 35,719 unique diagnosis (ICD-9/10), procedure (CPT), and drug product (NDC) codes in the patient’s historical data. We tested bivariate associations between each feature and agitation using logistic regression, adjusting for demographic characteristics. Statistically significant features were used to build a final prediction model using LASSO regression with 5-fold cross-validation. Model fit and classification performance were evaluated using AUC and sensitivity/specificity, respectively.

RESULTS: The final prediction model comprised 148 features and had good model fit (AUC: 93.4%) and classification performance (sensitivity: 83.5%, specificity: 90.3%). Features retained in the final model included need for moderate-to-intensive evaluation and care in the home or nursing home setting, and history of other psychiatric conditions including depressive and delusional disorders. Antipsychotics drugs, notably quetiapine, were more common among Alzheimer’s patients with vs. without agitation, despite matching on time since diagnosis.

CONCLUSIONS: This study demonstrated that a predictive model using broad data on historical treatment and healthcare utilization history can achieve excellent classification performance. This approach has the potential to improve research on novel health conditions prior to implementation of comprehensive diagnostic codes or where confirmed diagnosis is otherwise unavailable.