Low-cost algorithms for clinical notes phenotype classification to enhance epidemiological surveillance: A case study

Area of Science:

• Public Health
• Infectious Disease Epidemiology
• Health Informatics

Background:

• Emerging pandemics necessitate robust epidemic intelligence systems.
• Event-based surveillance using electronic health records (EHRs) is crucial for early detection.
• Challenges include limited data, evolving symptoms, and non-standardized coding.

Purpose of the Study:

• To enhance epidemic intelligence through event-based surveillance in an emerging pandemic context.
• To classify EHRs for predicting COVID-19-related categories with limited disease knowledge and data.
• To develop rapid, cost-effective natural language processing (NLP) methods for resource-limited settings.

Main Methods:

• Utilized NLP techniques to develop rapid and cost-effective classification models.
• Annotated a corpus for training and testing models, including logistic regression and fine-tuned transformers.
• Evaluated model performance using F1 scores and Pearson correlation with official case counts.

Main Results:

• Spanish-adapted transformer models (BETO Clínico, RoBERTa Clínico) achieved high performance (F1=88.13%, 87.01%).
• A simple logistic regression (LR) model performed competitively (F1=85.09%), outperforming more complex models.
• LR and BETO Clínico showed strong correlation with official COVID-19 data, identifying uncoded cases.

Conclusions:

• Simple, resource-efficient NLP methods can yield results comparable to complex approaches.
• BETO Clínico and LR models demonstrate strong correlation with official surveillance data.
• These findings support cost-effective strategies for epidemic response, especially in resource-limited settings.