Passive Digital Markers for Alzheimer’s Disease and Related Dementia Predict Mild Cognitive Impairment

Abstract

AbstractBackgroundAlzheimer’s Disease and Related Dementia (ADRD) typically progresses through three clinical stages: Normal Cognition (NC), Mild Cognitive Impairment (MCI), and ADRD. Early detection of MCI is critical for identifying individuals at high risk for ADRD for proper management and early interventions to prevent progression to ADRD. Recently, Boustani et al. identified a passive signature in unstructured Electronic Medical Records (EMR) data for early detection of ADRD. Our objective is to test whether machine learning models trained using ADRD Passive Digital Markers (PDMs) can accurately predict MCI 3 years in advance.MethodsUsing a dataset of 1379 and 5516 MCI and healthy individuals (respectively), we represented each individual using 209 markers in the form of the counts of specific words or phrases (e.g., infection, respiration disorders, anti‐anxiety agents) extracted from clinical notes. In addition, we also included three demographic variables (age at prediction time, gender, and race). Each patient was assigned an index date, where the index date for MCI patients is the MCI diagnosis date. Each MCI patient has four distinct controls with adjusted age at their MCI index date. The prediction time is set to index date – k, where k equals 1 and 3 years. The PDMs were extracted from the EMR data collected within two years before the prediction date. Four widely used machine learning algorithms were evaluated using 10‐fold cross‐validation procedures for predicting MCI 1‐year and 3‐years prior. The predictive performance was assessed using ROC curves and Area Under ROC Curve (AUC).Results Figure 1 shows that machine learning models developed using ensemble 200 decision trees trained using the gradient boosting algorithm outperform the remaining three algorithms evaluated in this study. The best AUC scores are 0.84 and 0.78 for identifying MCI patients 1 year and 3 years in advance (respectively). Table 1 summarizes the top 30 PDMs and their feature importance scores for predicting 1‐year and 3‐year MCI.ConclusionMachine learning models trained using PDMs, originally designed for early detection of ADRD, can also predict MCI. Future work will leverage structured EMR data to improve the performance of our models.