ABSTRACT Developing predictive models for COVID‐19 diagnosis remains challenging due to the inherent uncertainty and qualitative nature of clinical assessments. While conventional machine learning techniques such as logistic regression, support vector machines, ensemble methods, and neural networks are widely used, their ability to effectively represent ambiguity and linguistic information is limited. This study introduces an uncertainty‐aware computational framework based on asymmetric Z‐number fuzzy logistic regression, specifically designed for diagnostic settings characterized by limited and imprecise information. In particular, the proposed approach focuses on low‐dimensional scenarios in which the available predictor space is highly constrained (e.g., a single binary RT‐PCR result), while the diagnostic outcome is expressed through linguistic and uncertain assessments. Model parameters are estimated using two approaches: (i) a matrix‐based least squares estimation (LSE) method derived from the extension principle and asymmetric fuzzy number representation, and (ii) a neural network‐based estimation approach incorporating an exponential likelihood function. The framework is evaluated using real‐world diagnostic data with eight linguistic outcome levels, as well as simulated Z‐number data. Model performance is assessed using accuracy, Cohen's Kappa, and Mean Degree of Membership (MDM). The results indicate that, under data‐constrained and uncertainty‐rich conditions, the LSE‐based approach yields more stable and consistent estimates within the proposed modeling framework. For instance, in the real‐data setting, the LSE approach achieved an accuracy of 0.8653, a Kappa value of 0.65, and an MDM of 0.7122 for the restriction component, compared to 0.7346, 0.5181, and 0.6339, respectively, for the ANN‐based approach. Importantly, the proposed framework is intended as a methodological contribution to uncertainty‐aware diagnostic modeling rather than a high‐dimensional predictive benchmarking study. The model outputs reflect diagnostic confidence associated with linguistic assessments rather than direct measures of physiological disease severity. These findings demonstrate the applicability of asymmetric Z‐number‐based models for supporting decision‐making in early‐stage or resource‐constrained diagnostic environments, where ambiguity and expert judgment play a central role.
Ganjgah et al. (Mon,) studied this question.