Parsimonious and explainable biomarker-based severity score for hospitalised patients with COVID-19-related respiratory infections: development, validation and XAI benchmarking

BACKGROUND: Severity scoring systems are increasingly important tools for stratifying hospitalised patients, guiding treatment decisions, and enabling analyses that capture illness severity. However, many existing models are complex, lack transparency, or depend on clinical observations and physiological measures that can vary between clinicians or be influenced by factors unrelated to the disease. In respiratory infection-related diseases, beyond traditional scores (e.g. CURB-65, PSI, NEWS2), there is interest in biomarker-driven tools that provide quantitative, reproducible measures that directly reflect underlying pathophysiology, enabling earlier and more accurate risk stratification, dynamic monitoring and prognostic tracking. OBJECTIVES: To develop and validate an interpretable, parsimonious severity score based on routine biomarkers and age, designed to classify illness severity in patients with COVID-19-related respiratory infection, and to benchmark its performance against an explainable machine learning model. METHODS: Two parallel biomarker scoring methods, interquartile range (IQR) based and standard deviation (SD) based, were developed to standardise 22 routine biomarkers. Records were categorised by disease severity based on the patient's location - critical care units (CCU) or general wards. Features were selected using logistic regression. The resulting biomarker severity score (BioSeSco), comprising six features (age, albumin, interleukin-6, urea, creatinine and white cell count), was validated on an unseen patient-disjoint holdout cohort, and benchmarked against an XGBoost model using SHAP explainability. RESULTS: The BioSeSco demonstrated excellent discriminative performance, achieving an AUC of 0.94 on the unseen holdout cohort, and strongly correlated with SHAP sum values of XGBoost. Stratification analysis showed increasing CCU admission rates across score deciles, supporting its use as a measure of continuous disease severity. CONCLUSION: We present an interpretable severity score developed in COVID-19 patients. The score demonstrated performance comparable to that of more complex machine learning approaches, while offering transparency and ease of implementation. It enables patient stratification and provides a transparent model for evaluating the performance of more complex AI models, with potential for future integration into clinical decision support.