Utilizing electronic medical record data to predict onset of Alzheimer's disease and related dementias: Epidemiology / Risk and protective factors in MCI and dementia. (7th December 2020)

• Record Type: Journal Article
• Authors: Marra, David E; Miller, Amber H.; Li, Qian; Yang, Xi; Smith, Glenn E.; Wu, Yonghui; Bian, Jiang; Maraganore, Demetrius M
• Is Part Of: Alzheimer's & dementia. Volume 16(2020)Supplement 10
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-12-07
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.041233
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: As the population ages, the incidence of Alzheimer's disease and related dementias (ADRD) are expected to rise and cause a significant burden to the world economy and healthcare systems. Pharmaceutical trials to cure ADRD have largely proven unsuccessful. Therefore, there is a critical need for early identification and intervention for individuals at risk of ADRD to prevent or delay disease onset. Method: Using electronic medical record (EMR) data from a large, academic medical center, a historical cohort of Floridians aged 50 and older who developed ADRD were identified, along with an age and sex‐matched referent group. Discrete diagnostic codes, medication history, and biomarkers were used to develop 1, 3, and 5‐year ADRD machine learning‐based prediction models through theory‐driven and data‐driven approaches. Result: A cohort of 59, 799 subjects were identified as meeting study criteria, including 5, 711 diagnosed with ADRD during the study period. Approximately five controls (n = 18, 765) for each patient were matched by birth year and sex. In the theory‐driven experiment, the logistic regression model achieved the area under the curve (AUC) of .664 for 1‐year ADRD prediction, .635 for 3‐year prediction, and .650 for 5‐year prediction. Further analysis exhibited that age, ethnicity, stroke, and history of traumatic brain injury had the highest odd's ratios for ADRD development; whereas, use of NSAIDs and hormone‐replacement therapy was associated with reduced risk of ADRD (See Table 1). Compared to the theory‐driven models, the data‐driven models (gradient boosting trees) improved the AUC by ∼0.06 for 1‐ and 3‐year prediction. Conclusion: Using widely available EMR data, a large cohort of adults who developed ADRD and matched controls were used to create 1, 3, and 5‐year prediction models. These models can be adopted in decision‐support systems for detecting individuals at high‐risk of developing ADRD, allowing for interventions to be implemented in a timely manner. Our future research will replicate these models with data from the OneFlorida Clinical Research Consortium, a data repository including 15 million patients, covering more than 60% of the Florida population. These models will be externally validated via neuropsychological assessment before EMR‐tool kits are disseminated to medical practitioners across the state of Florida.