Rapid Triage of Children with Suspected COVID-19 Using Laboratory-Based Machine-Learning Algorithms

Area of Science:

• Computational biology
• Infectious disease diagnostics
• Pediatric medicine

Background:

• Early detection and isolation of SARS-CoV-2 are crucial for limiting its spread.
• Machine learning (ML) offers a promising approach for rapid disease diagnosis.
• Identifying pediatric COVID-19 cases efficiently is essential for timely intervention.

Purpose of the Study:

• To evaluate the effectiveness of common ML algorithms for rapid triage of pediatric COVID-19.
• To utilize easily accessible and inexpensive laboratory parameters for ML model development.
• To compare the performance of six distinct ML algorithms in diagnosing COVID-19 in children.

Main Methods:

• A cross-sectional study involving 566 children with respiratory diseases (280 SARS-CoV-2 positive, 286 controls).
• Six ML algorithms (Random Forest, SVM, LDA, ANN, KNN, Decision Tree) were trained and validated using blood laboratory data.
• Model performance was assessed using stratified cross-validation and an independent test set, evaluating accuracy, sensitivity, specificity, and F1 score.

Main Results:

• Random Forest (RF) and Support Vector Machine (SVM) models achieved the highest diagnostic accuracy (85% and 82.1%, respectively).
• Models exhibited higher sensitivity than specificity and better negative predictive value than positive predictive value.
• The RF model demonstrated a superior F1 score (85.2%) compared to the SVM model (82.3%).

Conclusions:

• ML algorithms, particularly RF and SVM, show significant potential for the rapid triage of pediatric COVID-19 cases using routine lab data.
• These models can assist healthcare providers in early identification, potentially preceding definitive PCR or antigen testing.
• Implementing ML tools could enhance diagnostic efficiency in healthcare facilities without additional costs.