AI for AD/ADRD Care: Essentials to Improve Learning Health Systems’ Management of Aging People

Introduction: Primary care providers (PCPs) are the first line of medical care, and as such, they often are the first to hear of concerns regarding cognitive decline. Yet, under-diagnosis of Alzheimer's disease and related disorders (ADRD) in the primary care setting is widely recognized, and PCPs are uncertain about which patients to assess, which tools to use, and how to use them. This project aims to develop a brief digital cognitive screening tool that analyzes cognitive and voice data using a machine learning framework to offer a practical and efficient solution for cognitive health screening to the PCP.

Methods: Our study group comprised 53 cognitively normal and 51 cognitively impaired older adults. Each completed a risk assessment task, a symbol matching task, and four speech/language tasks, followed by a second administration of working memory to investigate the added utility of practice effects. The speech/language tasks were processed to extract acoustic and linguistic features. Bayesian adaptive regression trees were used to test 11 models. Traditional clinical cognitive evaluations for mild cognitive impairment were used as the gold standard.

Results: The top three models included the results of the symbol matching tasks alone (classification accuracy of c = 0.91) or in combination with either the personal narrative task (c = 0.94) or the counting backward task (c = 0.90). These results were similar to or slightly better than the traditional Quick Mild Cognitive Impairment (QMCI) screen.

Conclusion: The combination of features from the working memory task and voice tasks can accurately classify cognitively normal versus cognitively impaired older adults. Future work is being conducted to understand this tool's usability in the real-world clinical setting, and we are looking to develop a Spanish version of this tool and evaluate its effectiveness in a Spanish cohort. 

Fifty percent of people living with dementia are undiagnosed, delaying access to treatment, education, and support. There is a need for large-scale, low-cost strategies to improve dementia detection. The electronic health record (EHR) Risk of Alzheimer's and Dementia Assessment Rule (eRADAR) was developed to identify older adults at risk of having undiagnosed dementia using routinely collected clinical data. In this webinar, we will present two studies aimed at evaluating the potential impact of using eRADAR to improve dementia diagnosis. We will present external validation of the eRADAR risk score in two real-world healthcare systems, Kaiser Permanente Washington and the University of California at San Francisco. This talk will highlight statistical considerations for validating a prediction model intended for prospective clinical implementation, including temporal validation and auditing algorithmic fairness.  

The U.S. Medicare Minimum Data Set (MDS) is a federally mandated standardized clinical assessment tool administered by the United States Centers for Medicare and Medicaid Services to facilitate care management for residents in Medicare and Medicaid certified nursing homes. However, longitudinal assessments in these real-world data are irregular, and their timing is likely informative and outcome dependent. To address this problem, we propose a semiparametric joint model that handles time-varying covariates and includes a shared random effect (latent variable) with a time-varying coefficient. We also extend the model in two ways: 1) inclusion of inverse conditional intensity rate ratio weights to handle auxiliary covariates, and 2) time-varying coefficients for measured covariates. We demonstrate the estimators' asymptotic consistency and normality. Extensive simulation studies show excellent finite-sample properties. The proposed methods are applied to 40,713 assessments of 9,545 older adults living with Alzheimer's Disease who were part of the MDS during their first 180 days after discharge from hospital to nursing home after hip fracture. The simulations and MDS data application show that the proposed methods are feasible for use with real-world data.