Abstract
Early identification of Alzheimer's disease (AD) and AD-related dementias (ADRD) has high clinical significance, both because of the potential to slow decline through initiating FDA-approved therapies and managing modifiable risk factors, and to help persons living with dementia and their families to plan before cognitive loss makes doing so challenging. However, substantial racial and ethnic disparities in early diagnosis currently lead to additional inequities in care, urging accurate and inclusive risk assessment programs. In this study, we trained an artificial intelligence foundation model to represent the electronic health records (EHR) data with a vast cohort of 1.2 million patients within a large health system. Building upon this foundation EHR model, we developed a predictive Transformer model, named TRADE, capable of identifying risks for AD/ADRD and mild cognitive impairment (MCI), by analyzing the past sequential visit records. Amongst individuals 65 and older, our model was able to generate risk predictions for various future timeframes. On the held-out validation set, our model achieved an area under the receiver operating characteristic (AUROC) of 0.772 (95% CI: 0.770, 0.773) for identifying the AD/ADRD/MCI risks in 1 year, and AUROC of 0.735 (95% CI: 0.734, 0.736) in 5 years. The positive predictive values (PPV) in 5 years among individuals with top 1% and 5% highest estimated risks were 39.2% and 27.8%, respectively. These results demonstrate significant improvements upon the current EHR-based AD/ADRD/MCI risk assessment models, paving the way for better prognosis and management of AD/ADRD/MCI at scale.
Introduction
Early diagnosis of Alzheimer’s Disease and Related Dementias (ADRD) and Mild Cognitive Impairment (MCI) is a critical goal in geriatric medicine. Timely detection enables access to disease-modifying therapies, management of modifiable risk factors like hypertension, and proactive care planning. However, current screening paradigms face significant hurdles. High-fidelity methods such as neuroimaging and biomarker analysis are prohibitively expensive for population-wide screening. Conversely, existing EHR-based statistical models, such as eRADAR, often rely on a limited set of pre-defined risk factors and can exhibit racial and ethnic biases. There is a pressing need for scalable, accurate, and inclusive tools that leverage existing medical records to identify high-risk individuals.
Methods
This study utilizes a two-stage "Foundation Model" approach to represent longitudinal patient data, similar to the architecture of Large Language Models.
Data and Cohorts
The researchers utilized a de-identified dataset from NYU Langone Health spanning January 2013 to January 2023. The Pre-training Cohort included 1,030,438 unique patients of all ages, while the Fine-tuning Cohort focused on 142,702 patients aged 65 and older. Features included demographics, ICD-10 codes, LOINC-coded lab values, and therapeutic-level medications, totaling 57,735 distinct variables.
Model Architecture: TRADE
- Pre-training: A Transformer architecture was trained on the longitudinal EHR data using a "masked token modeling" strategy. This allowed the model to learn the complex relationships between clinical events across a patient's history.
- Fine-tuning: The pre-trained foundation model was fine-tuned using Low-Rank Adaptation (LoRA) to create TRADE (Transformer for Risks of AD/ADRD based on EHR). TRADE was specifically optimized to predict AD/ADRD/MCI onset within 1, 2, or 5 years.
Prediction Framework
To address the "diagnosis gap"—the delay between disease onset and clinical recording—the study employed a structured timeline approach. This included a feature window for historical data, a gap window to ensure predictions were truly prospective, and outcome windows for future risk assessment.
Results
TRADE demonstrated superior performance compared to the clinical benchmark, eRADAR.
- Predictive Accuracy: The model achieved an AUROC of 0.772 for 1-year predictions and 0.735 for 5-year predictions (outperforming eRADAR’s 0.688).
- Clinical Utility: For the top 1% and 5% of patients identified as highest risk, the Positive Predictive Value (PPV) was 39.2% and 27.8%, respectively, over a 5-year horizon.
- Validation: To confirm the model captured cognitive decline, researchers used GPT-4 to extract Mini-Mental State Exam (MMSE) scores from unstructured notes. High-risk flags from TRADE correlated strongly with lower MMSE scores.
- Feature Importance: Using "Integrated Gradients," the authors identified key risk precursors. Beyond memory loss, TRADE flagged gait abnormality, sleep apnea, urinary incontinence, and specific medications (e.g., escitalopram) as high-importance features.
Discussion
While TRADE significantly improves upon existing models and narrows performance gaps across racial groups, disparities persist, likely reflecting systemic inequities in diagnosis timing. The retrospective nature of the study and the inherent "diagnosis gap" in EHR data remain challenges. Future work will focus on integrating multi-modal data, including medical imaging and unstructured clinical notes, to further enhance predictive capabilities and clinical management of AD/ADRD/MCI at scale.
Summary updated