Limited ‘heft’ of weight-based outcomes in predicting influenza A virus disease severity in ferrets

Studies evaluating viral pathogenicity in small mammalian models often quantify disease severity using the magnitudes of temperature rise and weight loss post-challenge. However, no rigorous assessment on the transformation of serially collected data into features suitable for predictive models has been conducted. Using data aggregated from ferrets inoculated with a diverse panel of influenza A viruses (IAV) spanning a broad range of clinical outcomes, we assessed statistical correlations and predictive performance of temperature and weight loss, summarized by conventional and novel approaches. Conventional summary metrics (peak values or area under the curve) were weak and inconsistent correlates of overall disease severity and viral titers. Novel dynamic weight metrics capturing onset, duration, slope, and volatility over 14 days showed lower coefficients of variation than conventional summary approaches. However, inclusion of novel metrics did not meaningfully improve the predictive performance of machine learning models for disease severity outcomes in IAV-inoculated ferrets. Mixed-effects models indicated that weight loss post-IAV infection is driven by time and viral burden, with temperature contributing little additional information. Collectively, these findings support that derived metrics are at least comparable, if not enhanced, to conventional summaries for data science analyses of serially generated clinical data from in vivo pathogen studies. However, because pathogen disease severity in mammals is multifactorial, models that rely solely on weight and temperature metrics without additional quantitative measures of clinical perturbation within-host are unlikely to achieve strong predictive performance.Author summary: Viral pathogens are often studied in small mammals, where daily body temperature and body weight are routinely tracked to gauge disease severity. However, machine-learning efforts using these in vivo records have struggled to predict influenza morbidity reliably, raising a key question: is this due to inherent limitations in temperature and weight measurements, or in how these serial measurements are converted into model-ready features? We addressed this gap by analyzing two standard morbidity readouts (temperature rise and weight loss) using daily records from more than 800 ferrets infected with over 100 diverse influenza A virus strains (human, avian, and swine origin), forming one of the largest datasets of its kind. We systematically tested common summary metrics, identified which are most variable across infections, and introduced new derived metrics designed to better capture informative patterns in longitudinal trajectories. Combining statistical evaluation with machine-learning experiments, we show that careful feature construction from serially collected in vivo data can meaningfully affect predictive performance and interpretation. Because many pathogens and mammalian models generate similar serially-collected records, our methodological approach and results are broadly transferable, especially for researchers aiming to repurpose serial in vivo measurements for data-science and predictive modeling.