Objective:

Background:

Dementia diagnosis is often delayed in primary care, leading to poor outcomes. While complex AI models achieve high accuracy, they require hundreds of features from specialist workups, making them impractical for frontline screening. The Medicare AWV, conducted over 9 million times annually, represents a unique opportunity for scalable, early detection by leveraging its limited, routinely collected data.

Design/Methods:

We developed an AI model to operate on a streamlined set of approximately 50 features consistently documented during AWVs, including demographics, medical history, and cognitive screening results. The model was validated on a publicly available dataset of 5,000 patient records (ages 50-89) from the National Alzheimer’s Coordinating Center (NACC). This large, multi-center dataset served as a robust proxy for real-world clinical data.

Results:

Despite the constrained feature set, the model performed well. For dementia (DE) classification, it achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.98 and 0.92 accuracy. Furthermore, it demonstrated a high sensitivity of 0.92 and specificity of 0.93 for dementia classification, balancing case identification with a low false-positive rate. For normal cognition (NC), it achieved an AUROC of 0.95. Performance for identifying MCI was moderate (AUROC: 0.86), reflecting the clinical difficulty of this transitional state.

Conclusions:

This AI framework demonstrates excellent screening accuracy using only limited AWV data, offering a practical and scalable tool to enhance early dementia detection in primary care. By helping to reduce diagnostic delays and facilitate timely specialist referrals, this approach can significantly improve patient care pathways. Ongoing prospective validation on local clinical data is underway to confirm these findings and move towards clinical implementation for identifying at-risk individuals sooner.