A Natural Language Processing Model for COVID-19 Detection Based on Dutch General Practice Electronic Health Records...

Area of Science:

• Computational linguistics
• Medical informatics
• Epidemiology

Background:

• Natural language processing (NLP) models like BERT show potential for improving disease identification from electronic health records (EHRs).
• The COVID-19 pandemic highlighted challenges in disease surveillance, particularly with unstructured EHR data.
• General practitioners' EHRs in the Netherlands contain valuable information but are difficult to analyze due to free-text entries.

Purpose of the Study:

• To develop and validate a BERT model for detecting COVID-19 consultations within Dutch general practice EHRs.
• To assess the model's accuracy and reliability in identifying COVID-19 cases.

Main Methods:

• A BERT model was pre-trained on Dutch language data and fine-tuned using a dataset of COVID-19 and non-COVID-19 general practitioner consultations.
• Model performance was evaluated on an independent test set and validated through external EHR data, PCR test results, and hospitalization data.
• The study utilized 300,359 general practitioner consultations for model development.

Main Results:

• The developed BERT model achieved high accuracy (0.97) and F1-score (0.90) for COVID-19 consultation detection.
• External validation demonstrated comparable high performance, while PCR validation showed high recall but lower precision and specificity.
• The model significantly correlated with COVID-19 hospitalizations and could predict cases weeks before the first confirmed case in the Netherlands.

Conclusions:

• The validated BERT model accurately identifies COVID-19 cases in general practice EHRs, even preceding confirmed diagnoses.
• This demonstrates the potential of NLP models for early disease outbreak detection and effective disease surveillance.
• The study provides a blueprint for using similar models for early recognition of various illnesses.