Temporal Patterns of Vital Signs and Their Association With 28-Day Survival in Patients With COVID-19-Associated Acute Respiratory Distress Syndrome (ARDS) Treated With High-Flow Nasal Oxygen: A Machine Learning Model.

BACKGROUND: High-flow nasal oxygen (HFNO) is widely used in COVID-19-associated acute respiratory distress syndrome (ARDS) to delay or avoid invasive mechanical ventilation (MV). However, prolonged HFNO in the presence of worsening respiratory physiology may contribute to patient self-inflicted lung injury, and the optimal timing for escalation to MV remains uncertain. This study evaluated whether temporal patterns in vital signs and the duration of HFNO exposure were associated with 28-day mortality. METHODS: A retrospective observational study was conducted, including 98 adults with confirmed COVID-19 pneumonia admitted to the ICU on HFNO who subsequently required MV. Longitudinal vital signs were recorded three times daily for up to 28 days. Data were transformed into a patient-day panel to capture end-of-day clinical decision points. Feature engineering incorporated daily summaries, short-term trends, and persistence of abnormalities. Four modeling approaches-elastic-net logistic regression, discrete-time survival modeling, gradient boosting, and shallow decision trees-were trained using patient-grouped cross-validation. Model performance was assessed using precision-recall metrics and calibration. RESULTS: The cohort had a mean age of 61.8±13.9 years and a mean BMI of 28.5±6 kg/m²; overall mortality was approximately 80%, with 79 patients dying during hospitalization and 19 patients surviving to 28 days. Non-survivors exhibited persistently elevated respiratory rates, progressive declines in oxygen saturation, and greater variability in cardiovascular and respiratory vital signs. Divergence in respiratory rate between survivors and non-survivors became pronounced around day 12. BMI appeared to influence respiratory rate trajectories. Across all models, prolonged HFNO duration in the presence of worsening respiratory physiology was associated with increased mortality risk. Gradient boosting achieved the highest discrimination, while logistic regression and discrete-time survival models demonstrated superior calibration and interpretability. CONCLUSION: Worsening respiratory trajectories-rather than isolated vital-sign abnormalities-were strongly associated with mortality among patients treated with HFNO prior to MV. The interaction between prolonged HFNO exposure and physiological deterioration suggests a time-sensitive escalation window during which intubation may improve outcomes. These findings are exploratory and require validation in larger, prospective cohorts to define actionable thresholds for the optimal timing of MV in COVID-19 ARDS.