# Artificial Intelligence-Enabled Comprehensive Electronic Health Record Phenotyping at a Large Scale

## TL;DR

This paper uses AI to accurately identify ten diseases in a large patient dataset using electronic health records, achieving high accuracy with balanced data and cross-validation.

## Contribution

A novel AI-enabled EHR phenotyping framework using balanced data and cross-validation for ten diseases in a large multi-site dataset.

## Key findings

- AI-enabled EHR phenotyping achieved AUCs above 0.95 and AUPRCs above 0.83 across ten diseases.
- Manual annotation and balanced dataset design improved model training and accuracy.
- Detailed error analysis provided insights into false positives and negatives for each disease.

## Abstract

The Electronic health record (EHR) contains rich and ever-growing information, especially for the gerontologic population with multiple comorbidities. With the advent of powerful artificial intelligence (AI) tools, we can perform accurate EHR phenotyping, which is the foundation for downstream analyses. Here, we performed EHR phenotyping of ten diseases in a large multi-site clinical dataset of 145,787 unique patients, including epilepsy (and subtypes), ischemic stroke, subarachnoid hemorrhage, subdural hematoma, Alzheimer’s diseases and related dementias, Parkinson’s disease, cardiac arrest, traumatic brain injury, brain tumor, and congestive heart failure. We used AI-enabled natural language processing that extracts the presence of keywords from unstructured clinical notes while considering negations, as well as structured diagnosis codes (ICD) and medications. We used logistic regression, random forest, and XGBoost. The dataset was constructed from two sites by sampling an equal number of participants in people with or without ICD or medication in the EHR, which forms four groups: ICD+Med+, ICD+Med-, ICD-Med+, and ICD-Med-. We manually annotated each case as positive or negative by reading the clinical notes. In this way, the dataset has a roughly balanced positive/negative ratio to ease training. We then did leave-one-site-out cross-validation. The areas under the receiver operator curve are higher than 0.95 across the diseases. The areas under the precision-recall curve are higher than 0.83 across the diseases. For each disease, we conducted detailed error analyses for both false positives and negatives. The results lead to accurate phenotyping and insights from patient phenotypes at both population and individual levels.