Summary:
UCLA researchers in the Department of Neurology have developed a software to predict the diagnosis of Alzheimer’s Disease without racial or ethnic biases.
Background:
Alzheimer’s Disease (AD) represents a major health and economic challenge in the United States. Despite its prevalence, Alzheimer’s remains significantly underdiagnosed, primarily among underrepresented populations. Studies have shown that the sensitivity of Medicare claims for detecting Alzheimer’s is only 50-65%, compared to gold standard diagnoses from longitudinal cohorts, which can achieve up to 90% sensitivity. This challenge is even more pronounced in understudied populations, where systemic disparities further exacerbate the issue. Prior solutions have leveraged electronic health records (EHR) data to predict undiagnosed AD. A supervised learning model, eRADAR, was developed to predict undiagnosed dementia using clinical features from electronic health records data linked to participants in the Adult Changes in Thought study. This software, however, was not evaluated for the accuracy in underrepresented groups. These limitations in the field underscore the need for a predictive model that can accurately and precisely detect neurological diseases in underrepresented groups and reduce healthcare biases.
Innovation:
UCLA researchers have developed an innovative software designed to reduce biases in predicting undiagnosed Alzheimer’s Disease (AD). This software uses a semi-supervised positive unlabeled learning (SSPUL) framework, utilizing positive and unlabeled data from diverse, real-world EHRs. To address algorithmic racial and ethnic bias, race- and ethnicity-specific probabilistic criteria were assigned to positive and negative labels for a subset of unlabeled instances with minimal noise. Classification cutoffs were optimized to enhance group benefit equality across all groups. Model predictions were rigorously validated using proxy ICD codes, medications for dementia, and polygenic risk scores as benchmarks. This software successfully achieved a superior sensitivity for Non-Hispanic White, NH-African American, and Hispanic Latino groups, compared to baseline models. Additionally, it demonstrated enhanced fairness, reflected in the lowest parity loss for equal opportunity, group benefit equality, and specificity across all racial and ethnic groups. This groundbreaking software represents a significant advancement in medical diagnosis, offering a more equitable approach to addressing healthcare disparities and improving AD detection.
Potential Applications:
• Early detection of undiagnosed Alzheimer’s disease
• Personalized medicine and treatment plans
• Public health policy and research into prevalence of undiagnosed Alzheimer’s disease
• Epidemiology studies
Advantages:
• Reduction in healthcare disparities
• Higher sensitivity and specificity for prediction of diagnosis
• Enhanced fairness metrics and equitable healthcare practices
• Seamless integration into existing HER systems
• Scalable to analyze large datasets
• Improved early detection and treatment options
Development-To-Date:
The present code was first created in 2024 and has U.S. Provisional Patent Application No. 63/737,184 entitled SYSTEMS AND METHODS FOR FAIR PREDICTION OF UNDIAGNOSED ALZHEIMER'S DISEASE.
Reference:
UCLA Case No. 2025-113
Lead Inventor:
Timothy Chang