Leveraging Clinical Record Geolocation for Improved Alzheimer's Disease Diagnosis Using DMV Framework.

Area of Science:

• Computational linguistics
• Medical informatics
• Public health

Background:

• Early Alzheimer's disease (AD) detection is crucial but hindered by costly clinical and neuroimaging methods.
• Natural language processing (NLP) of clinical records offers a scalable detection alternative.
• Integrating geolocation may capture environmental risk signals for AD.

Purpose of the Study:

• To develop and evaluate a framework (DMV) for early AD detection using NLP and geolocation.
• To assess the performance of different language models (Llama3-70B, GPT-4o, GPT-5) combined with a Random Forest regressor.
• To determine the impact of incorporating spatial context on AD risk estimation accuracy.

Main Methods:

• The DMV framework framed early AD detection as a regression task predicting a continuous risk score.
• Evaluated embeddings from Llama3-70B, GPT-4o, and GPT-5, combined with a Random Forest regressor.
• Utilized a large CDC-derived dataset (≈284k records) with 10-fold cross-validation, assessing MSE, MAE, and R².

Main Results:

• Geolocation consistently improved model performance across all evaluated methods.
• The Random Forest baseline showed a 48.6% decrease in MSE with geolocation.
• GPT-5 embeddings combined with geolocation achieved the best performance (MSE=14.03, MAE=2.37, R²=0.978), with statistically significant error reduction (p < 0.001).

Conclusions:

• Combining advanced text embeddings with patient geolocation significantly enhances Alzheimer's disease risk prediction.
• Incorporating spatial context alongside clinical text aids in accounting for environmental and regional risk factors.
• This approach offers a scalable, data-driven workflow for improved early AD detection.