GACEMV: An ensemble learning framework for constructing COVID-19 diagnosis and prognosis models

Abstract

The accurate prediction of the severity and prognosis of COVID-19 patients is essential in order to select the ideal treatment plan, which can help reduce mortality rates associated with the virus. Artificial intelligence (AI)-based disease prediction models show potential in diagnosing and prognosis COVID-19 patients. However, their black-box nature, overfitting, and low accuracy limit their application. To address these challenges, this study proposes a GACEMV ensemble learning framework for building practical and high-precision COVID-19 patient diagnosis models. The GACEMV utilizes genetic algorithm to optimize the hyperparameters of base learners in order to enhance their predictive ability. Additionally, a comprehensive evaluation method is adopted to select the optimal combination of base learners, reducing performance differences and avoiding interference from noisy models. The Ensemble learning model constructed by GACEMV achieved good accuracy in predicting the severity and prognosis of COVID-19 patients through external data sets. Further ablation experiments confirm the necessity of genetic algorithm and comprehensive evaluation method for GACEMV. Moreover, the Ensemble learning model shows outstanding performance in identifying suspected COVID-19 patients in a specific data set, validating its identification ability. Furthermore, the SHAP method identifies a set of biomarkers associated with the severity and prognosis of COVID-19 patients, which are consistent with previous reports.