Use of artificial intelligence to assess genetic predisposition to develop critical COVID-19 disease: a comparative study of machine learning models

• Salomón Martín Pérez ^[1] ; Flora Sanchez Jimenez ^[2] ; Marta Jímenez Barragan ^[2] ; Catalina Sanchez Mora ^[2] ; Juan M. Borreguero Leon ^[2] ; Arrobas Velilla Teresa ^[2] ; Antonio León Justel ^[2] ; Sandra Fuentes Cantero ^[3] ; Agustín Valido Morales ^[4] ; Juan A. Delgado Torralbo ^[4]
  1. [1] Service of Clinical Biochemistry Virgen Macarena University Hospital Seville, Fedriani nº3, Seville, Spain, E-mail: salomon.martin.perez@gmail.com. https://orcid.org/ 0000-0002-2086-3597
  2. [2] Service of Clinical Biochemistry Virgen Macarena University Hospital Seville, Seville, Spain
  3. [3] Department of Clinical Laboratory Chemistry Rio Tinto General Hospital Huelva, Huelva, Spain. https://orcid.org/0000-0003- 4125-8625
  4. [4] Unit of Pulmonology, Virgen Macarena University Hospital Seville, Seville, Spain

  Mostrar afiliaciones +
• Localización: Advances in Laboratory Medicine / Avances en Medicina de Laboratorio, ISSN-e 2628-491X, Vol. 6, Nº. 2, 2025, págs. 181-189
• Idioma: inglés
• Enlaces
  • Texto Completo Ejemplar
• Resumen

  • Objectives Early prediction of critical COVID-19 disease is crucial for an optimal clinical management. The objective of this study was to optimize predictive models for critical COVID-19 disease. Clinical data, laboratory data and genetic polymorphisms were integrated into AI models to compare the performance of different machine learning algorithms. Methods Data from 155 inpatients were analyzed, 23 of whom developed critical disease. A univariate analysis was performed to assess potential correlations between seven SNPs, nine clinical variables and 10 laboratory parameters at admission. Results Of the 7 SNPs, only three SNPs demonstrated a significant association with critical disase, namely: rs77534576, rs10774671 and rs10490770. The ensemble models exhibited the best performance: Random Forest (AUC=0.989), XGBoost (AUC=0.954) and AdaBoost (AUC=0.927). Variable importance varied across models, with age, C-reactive protein, heart diseases and the three SNPs being the most influential features. The predictive power of models improved with the integration of the three SNPs, as compared to previous studies where genetic data were not included. Internal validation confirmed the superiority and stability of the ensemble models. Conclusions Machine learning models may help predict progression into critical COVID-19-disease. The predictive power of models improves when SNPs associated with COVID-19 severity are integrated with laboratory and clinical data. Prior to implementation in clinical practice, larger studies in different populations are needed to validate and support the generalization of these results.