An interpretable Day-1 machine-learning model for COVID-19 mortality and severity prognosis: development in a Pakistani hospital cohort and external evaluation in Chinese and Italian populations

Early risk stratification of COVID-19 patients at admission informs triage and resource allocation, particularly in low- and middle-income settings where intensive-care capacity is constrained. Most published models are not interpretable, rarely externally validated, and developed almost exclusively in high-income cohorts. We developed an interpretable random-forest classifier for three-class COVID-19 severity (Mild: not ventilated, survived; Severe: ventilated, survived; Fatal: died) using only data available within 24 hours of admission in 321 hospitalised PCR-confirmed COVID-19 patients in Pakistan. From 2,278 raw variables we retained 20 predictors by clinical screening and permutation-importance selection, with leakage-controlled multiple imputation by chained equations and SHapley Additive exPlanations (SHAP) for interpretation. A 15-feature variant was evaluated in two independent public cohorts: the Chinese iCTCF cohort (n = 894) and an Italian complete-blood-count cohort (n = 1,218). The development model achieved a macro F1 of 0.55, identifying the Fatal class most reliably (F1 0.67, recall 0.73) and Mild cases least well. SHAP identified a parsimonious set of admission-time predictors (oxygen saturation, lactate dehydrogenase, urea, C-reactive protein, age and radiographic pneumonia), including a raised urea-to-creatinine ratio in fatal cases most consistent with dehydration (median urea 80 vs 50 mg/dL, fatal vs mild; p < 0.001). Externally, the model retained useful discrimination in the Chinese cohort (mortality AUC 0.82 in 719 patients with recorded outcomes; severe-or-worse AUC 0.75; macro F1 0.50) but only modest discrimination in the Italian cohort (AUC 0.62, 95% CI 0.56-0.68), which we treat as a feature-availability stress test rather than a second validation. Cross-cohort SHAP rankings correlated strongly in China (Spearman ρ = 0.91) and moderately in Italy (ρ = 0.76), where most laboratory predictors were imputed and carried no information beyond the complete blood count; discrimination tracked how many predictors were actually measured. The admission-time signature converges with established markers of hypoperfusion and inflammation.