Artificial Intelligence for COVID-19 Severity Assessment: A Systematic Review and Meta-Analysis
Salma Muteb Almatrafi, Norah Mohammed Alkhulaif, Abdulrahman Abdullah Altuwaim, Itidal Mohammed Aljohani, Almaha Hamdan Alanazi, Abdulaziz Saeed Alserhani, Abdulmajeed Zaher Alzaher, Meshal Mansour Almuhanna, Naif Khalid Alhumaydani, Layan Talal Alraddadi, Rawan Maatouk Kheimi, Wajd Almehmadi, Abdullah BasnawiThe accurate assessment of coronavirus disease 2019 (COVID-19) severity remains a cornerstone for optimized resource allocation and clinical treatment planning. This systematic review and meta-analysis aimed to evaluate and compare the diagnostic performance of artificial intelligence (AI) models utilizing chest X-ray (CXR) versus lung ultrasound (LUS) modalities for COVID-19 severity stratification. Following the PRISMA 2020 guidelines, we conducted a comprehensive literature search across PubMed, Scopus, Web of Science, and Google Scholar from 2020 through April 2025. Inclusion criteria specifically targeted studies employing AI for severity assessment, while excluding secondary research, case reports, and non-English publications. Our analysis of ten selected studies revealed a progressive evolution in model performance for both binary and multi-class classification tasks. Detailed meta-regression indicated that transformer-based architectures and domain-specific pre-training contributed to higher sensitivity levels, particularly in early-stage stratification. Although CXR was the more prevalent modality in the literature, LUS-based AI models exhibited comparable diagnostic efficacy, offering a portable and radiation-free alternative that enhances clinical workflows in resource-constrained environments and point-of-care settings. Furthermore, the results indicate that the integration of domain knowledge and the application of rigorous external validation significantly enhance model generalizability. The analysis underscores a persistent performance gap in cross-institutional validation, suggesting a need for more diverse training cohorts. We conclude that while AI-driven CXR and LUS tools show high potential for severity assessment, the path to clinical deployment necessitates standardized external validation and the fusion of multi-modal clinical data to ensure robust predictive accuracy in diverse healthcare settings.