Tree-based classification model for Long-COVID infection prediction with age stratification using data from the...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health Research

Background:

• Post-acute sequelae of SARS-CoV-2 infection (PASC), or Long COVID, presents a complex diagnostic challenge due to its heterogeneous symptoms.
• Electronic Health Records (EHRs) contain valuable data but often suffer from sparsity and require sophisticated analysis for conditions like PASC.

Purpose of the Study:

• To propose and validate a domain knowledge-driven classification model for diagnosing PASC using EHR data.
• To develop a robust and interpretable model that incorporates clinically relevant features and demographic stratification.

Main Methods:

• Developed an XGBoost tree-based classification model incorporating features indicative of PASC and COVID-19 severity, guided by literature review.
• Utilized data from the Long COVID Computation Challenge (L3C), a subset of the National COVID Cohort Collaborative (N3C).
• Fine-tuned and calibrated models for optimal Area Under the Receiver Operating characteristic curve (AUROC) and F1 score, addressing class imbalance in N3C data.

Main Results:

• Achieved an average 5-fold cross-validated AUROC of 0.844 and F1 score of 0.539 across age-stratified populations in the training data.
• Attained an overall AUROC of 0.814 and F1 score of 0.545 on an independent testing dataset.
• Demonstrated strong performance and generalizability across different datasets and age groups.

Conclusions:

• The study validates the effectiveness of age-stratified, tree-based classification models for PASC diagnosis.
• Knowledge-driven feature engineering and demographic stratification are crucial for diagnosing complex conditions like PASC from sparse EHR data.
• The model's interpretability and robustness offer potential for clinical translation in managing Long COVID.