Diagnosis of patients with blood cell count for COVID-19: An explainable artificial intelligence approach

Resumo

Objective: Present an explainable artificial intelligence (AI) approach for COVID-19 diagnosis with blood cell count. Methods: Five AI algorithms were evaluated: Logistic Regression, Random Forest, Support Vector Machine, Gradient Boosting and eXtreme Gradient Boosting. A Bayesian optimization with 5-Fold cross-validation was used to hyper-parameters tuning. The model selection evaluated three results: cross validation performance, test set prediction performance and a backtest: performance on identifying patients negative for COVID-19, but positive for others respiratory pathologies. Shapley Additive explanations (SHAP) was used to explain the chosen model. Results: A Random Forest model was obtained with 77.7% F1-Score (IC95%:57.1;92.3), 85.9% AUC (IC95%:73.7;95.9), 74.4% Sensitivity (IC95%:50.0;92.1) and 97.5% Specificity (IC95%:93.6;100.0). The main features were leukocytes, platelets and eosinophils. Conclusion: The research highlights the importance of model interpretability, demonstrating blood cell count as a possibility for COVID-19 diagnosis. The methodological structure developed, using TRIPOD’s guidelines, can be extrapolated to other pathologies.