The underdiagnosis of Alzheimer's disease and related dementias (ADRD) has significant consequences, including missed intervention opportunities, increased healthcare costs, and an underestimation of the disease burden. In addition, current studies relying on diagnoses and medications for defining cognitive impairment (CI) in electronic health records (EHR) are limited by biases and inaccuracies in these data. To address this critical need for accurate identification of cognitive impairment in EHR, this study proposes the development and evaluation of a deep learning algorithm that utilizes EHR data from three large healthcare institutions: Mass General Brigham (MGB), University of Wisconsin-Madison (UW-Madison), and University of Texas Health Science Center at San Antonio (UTHSCSA). We seek to leverage the wealth of information in EHR including clinical notes, patient health history, and health system interactions that often contain signs of cognitive decline. Deep learning algorithms can leverage and learn from these complex text and data patterns in EHR. In this proposal, we aim to develop and evaluate a deep learning algorithm, Decipher-AI, to improve the detection of CI due to underlying ADRD pathophysiology (including cognitive concerns, mild cognitive impairment, and dementia). Decipher-AI is a two- level hierarchical model consisting of a note-level natural language processing (NLP) model that predicts the probability of CI for a clinical note and a patient-level model that combines structured data and aggregated NLP outputs to predict the probability of CI for a patient. For training and evaluation, we will use a “seed” reference standard set with detailed chart review and adjudication of cognitive diagnosis by an expert clinician (n=1,000), and then apply active learning strategies to iteratively increase sample size to n=20,000 (oversampling to include 50% from minority populations). Specifically, the study aims to develop Decipher-AI for identifying any indicators of cognitive impairment and detect ADRD subtypes (Aim 1); validate its performance at two other healthcare institutions (Aim 2); and examine biases and differences in predictors of CI across race/ethnic, sex, and socio-economic subgroups (Aim 3). While previous research has demonstrated success in identifying cognitive dysfunction in patients without dementia diagnoses or medication, further efforts are required to build inclusive reference datasets, integrate structured and unstructured data, and evaluate the algorithm in diverse healthcare settings across multiple institutions. Cutting-edge deep learning algorithms have been applied to many real-world tasks but in a limited manner to ADRD. We anticipate that our state-of-the-art deep learning algorithm will more efficiently and accurately detect signs of CI, and after validation be deployed in primary care to screen for patients with undiagnosed cognitive impairment. In addition, improved screening of cognitive impairmen...