Public Health.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Gerontology

Background:

• Early identification of Alzheimer's Disease and Related Dementias (ADRD) is crucial for effective patient management.
• Traditional methods like manual chart review are time-consuming and not scalable for large research studies.
• The Chronic Conditions Data Warehouse (CCW) e-phenotype, relying on ICD-10 codes, is a common but potentially limited approach.

Purpose of the Study:

• To evaluate novel methods for ADRD e-phenotyping using unstructured clinical notes.
• To compare the performance of models based on ICD codes versus those utilizing clinical notes.
• To investigate performance differences across genders in ADRD e-phenotyping models.

Main Methods:

• Analysis of unstructured clinical notes from 1,576 patients with and without ADRD.
• Development of various e-phenotyping models, with the CCW serving as a reference standard.
• Validation against a gold standard derived from manual chart review for 200 patients, using approximate randomization for statistical comparisons.

Main Results:

• The ICD-10 code approach (F1=0.83) showed a numerically higher overall F1-score than the Random Forest (RF) model (F1=0.81), but this difference was not statistically significant.
• The ICD approach exhibited significant gender-based bias, while the RF model did not.
• All models demonstrated numerically higher F1-scores on female patients' notes, irrespective of the balanced study corpus.

Conclusions:

• No single e-phenotyping model consistently outperformed others across all performance metrics (AUC, accuracy, specificity, precision, recall/sensitivity).
• The choice of e-phenotyping model should align with specific research needs, prioritizing metrics like recall or specificity.
• Demographic disaggregation, particularly by gender, is essential for a comprehensive understanding of model performance and potential biases in ADRD e-phenotyping.