Development of a Portable Tool to Identify Patients With Atrial Fibrillation Using Clinical Notes From the Electronic...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Data Analysis

Background:

• Electronic medical records contain valuable unstructured clinical information.
• Identifying specific patient cohorts, such as those with atrial fibrillation (AF), from free text is challenging.
• Existing methods may not fully leverage the rich data within clinical narratives.

Purpose of the Study:

• To develop and validate a natural language processing (NLP) algorithm for identifying patients with atrial Fibrillation (AF) using only free-text data from electronic medical records.
• To assess the performance of the NLP algorithm across different clinical sites.

Main Methods:

• Developed and trained 54 NLP algorithms using a training dataset of patients with AF billing codes.
• Optimized models based on F-score, varying parameters like feature count and stop words.
• Utilized logistic regression with term frequency-inverse document frequency (TF-IDF) for feature extraction.
• Validated the best-performing algorithm on two independent datasets (internal and external sites).

Main Results:

• The optimal algorithm, a logistic regression model, achieved high performance in the training set (F-score >0.93).
• The algorithm demonstrated strong performance on validation sets, with F-scores exceeding 0.90 at one site and 0.80 at another.
• Sensitivity and specificity were consistently high, indicating robust AF identification capabilities.

Conclusions:

• A novel NLP algorithm can effectively identify patients with atrial fibrillation using only unstructured text data.
• This approach significantly improves the utilization of clinical narratives for high-throughput and precise patient cohort identification.
• The algorithm shows promise for enhancing clinical research and patient management.