AI approaches for phenotyping Alzheimer's disease and related dementias using electronic health records

• Published in: Alzheimer's & dementia : translational research & clinical interventions Vol. 11; no. 2; pp. e70089 - n/a
• Main Authors: Knox, Sara, Aghamoosa, Stephanie, Heider, Paul M., Cutty, Maxwell, Wright, Andrew, Scherbakov, Dmitry, Hood, Gabriel, Nolin, Sara A., Obeid, Jihad S.
• Format: Journal Article
• Language: English
• Published: United States Wiley 01.04.2025
• Subjects: Alzheimer's disease and related dementias; artificial intelligence; digital phenotype; electronic health records; large language models; Alzheimer's disease and related dementias; large language models; electronic health records; artificial intelligence; digital phenotype
• ISSN: 2352-8737; 2352-8737
• DOI: 10.1002/trc2.70089

INTRODUCTION The current standard electronic (e‐)phenotype for identifying patients with Alzheimer's disease and related dementias (ADRD) from medical claims data yields suboptimal diagnostic accuracy. This study leveraged artificial intelligence (AI)–based text‐classification methods to improve the identification of patients with dementia due to ADRD using clinical notes from electronic health records (EHRs). METHODS EHR data for patients aged ≥ 64 (N = 4000) from an academic medical center were used. The cohort included 1000 patients with ADRD per the Chronic Conditions Warehouse (CCW) algorithm for ADRD (i.e., at least one ADRD International Classification of Diseases, Tenth Revision codes [ICD‐10 code]) and 3000 matched controls without ADRD (i.e., no CCW codes). We trained several AI‐based text‐classification models, including bag‐of‐words models, deep learning, and large language models (LLMs), to make ADRD determinations from clinical notes. The performance of each model was evaluated against “gold standard” manual chart review. RESULTS A foundational LLM derived from Llama 2 demonstrated superior performance in identifying patients with ADRD (area under the curve [AUC] = 0.9534, F1 score 0.8571) compared to both the current standard CCW algorithm (AUC = 0.8482, F1 score 0.8323, although only the AUC was statistically significantly different) and other AI‐based models. Several of the AI‐based models, including convolutional neural networks, also outperformed the CCW algorithm. DISCUSSION These findings highlight the potential of AI‐based text‐classification methods to optimize the automated identification of patients with ADRD using rich EHR data. However, the success of this approach depends on the quality of clinical notes, and more work is needed to refine and validate these methods across more diverse data sets. Highlights The current e‐phenotype for patients with Alzheimer's disease and related dementias (ADRD) in electronic health records has suboptimal diagnostic accuracy. The study used artificial intelligence (AI)–based text classification methods to improve the detection of patients with ADRD. AI‐based models, including convolutional neural networks, outperformed the Chronic Conditions Warehouse algorithm. The current standard electronic (e‐) phenotype for identifying patients with Alzheimer's Disease and Related Dementias (ADRD), the Chronic Conditions Warehouse (CCW) algorithm for ADRD, yields suboptimal diagnostic accuracy. We leveraged Artificial Intelligence (AI)‐based text‐classification methods, to improve the identification of patients with dementia due to ADRD using clinical notes from electronic health records (EHR). Using EHR data for patients aged 64 and older (N = 4000) from an academic medical center, we trained several AI‐based text‐classification models, including bag‐of‐words models, deep learning, and large language models (LLMs), to make ADRD determinations from clinical notes. The performance of each model was evaluated against “gold standard” chart review. A foundational LLM derived from Llama 2 demonstrated superior performance in identifying patients with ADRD dementia (area under the curve: AUC 0.95) compared to the current standard CCW algorithm (AUC = 0.85) and other AI‐based models. Several of the AI‐based models, including convolutional neural networks, also outperformed the CCW algorithm. These findings highlight the potential of AI‐based text‐classification methods to optimize the automated identification of patients with ADRD using rich EHR data. However, this approach depends on the quality of clinical notes and more work is needed to refine and validate these methods across more diverse data sets.