EHR-BERT: A BERT-based model for effective anomaly detection in electronic health records

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Clinical Data Analysis

Background:

• Physicians rely on electronic health records (EHRs) for patient care decisions.
• Health information technology (HIT) systems must be reliable to ensure patient safety.
• Existing deep learning methods for EHR anomaly detection have limitations including high false negatives and computational costs.

Purpose of the Study:

• To propose a novel framework, EHR-Bidirectional Encoder Representations from Transformers (BERT), for detecting anomalies in EHRs.
• To address the limitations of current deep learning methods in EHR anomaly detection.
• To enhance the reliability of EHR data for improved patient safety.

Main Methods:

• The EHR-BERT framework utilizes the Sequential Masked Token Prediction (SMTP) method.
• EHRs are treated as natural language sentences, with tokens masked during training and prediction.
• The model learns bidirectional EHR sequence patterns to identify anomalies based on deviations from normal models.

Main Results:

• EHR-BERT demonstrates significant improvements over existing models on large EHR datasets.
• The framework substantially reduces false positives and enhances anomaly detection rates.
• Improved performance is attributed to minimized information loss and maximized data utilization.

Conclusions:

• EHR-BERT shows great potential in reducing medical errors linked to anomalous clinical events.
• The framework enhances patient safety and the quality of healthcare services.
• EHR-BERT offers a promising solution for ensuring the reliability and quality of health data.