Abstract

Abstract Background/Aims Since early in the COVID-19 pandemic, there has been interest in the concept that some morbidity and mortality may be due to excessive inflammation. Several definitions of COVID-19 hyperinflammation (COV-HI) have been proposed, including Manson criteria (C-reactive protein, CRP ≥150mg/L or doubling above 50mg/L in 24 hours and/or ferritin 1500ug/L); and Webb criteria (includes CRP ≥150mg/L or ferritin ≥750ug/L). A consistent finding has been worse outcomes. Little is known regarding the underlying pathologies separating these patients from others. Aim To investigate whether machine learning using standard laboratory features can identify a distinguishing ‘COV-HI signature’. Methods A database of daily clinical and laboratory features was collected from 611 patients admitted to hospital with confirmed COVID-19 during the first wave of community-acquired infection at University College London Hospitals, Sheffield Teaching Hospitals, Newcastle upon Tyne Hospitals and Royal Wolverhampton. All data prior to mechanical ventilation were interrogated. Patients were categorised as COV-HI based on Webb thresholds (CRP >150 mg/L or ferritin ≥750ug/L). Laboratory features (peak or nadir depending on recognised predictors of illness severity) included: minimum lymphocyte count 10^9/L; minimum monocyte count 10^9/L; minimum haemoglobin g/dL, minimum albumin g/L; maximum neutrophils count 10^9/L; maximum alanine aminotransferase IU/L; maximum platelet count 10^9/L and maximum creatinine μmol/L. The data were analysed using unsupervised clustering and supervised machine learning models (logistic regression, support vector machine, decision trees, random forest classification (RF), naïve bayes and k-nearest neighbours) using Orange 3.29.1 software. Results 463 patients had sufficient data to determine Cov-HI status: 361 met COV-Hi definition at any point pre- ventilation during admission (median age 71, 65.93% male), 102 patients did not (median age 73, 51% male). In keeping with our previous work, COV-HI was associated with increased mortality (p < 0.0001), Odds ratio 3.506 (CI 1.871-6.916), relative risk 2.708 (CI 1.600-4.734). Multiple logistic regression revealed no significant relationship between sex and mortality (p 0.6673, Male Odds ratio 0.90). Unsupervised hierarchical clustering using laboratory features identified two clusters: Cluster-1, comprising 72.8% patients without hyperinflammation; and Cluster-2, comprising 73.3% COV-HI. Supervised machine learning models were tested using the same features. All models predicted COV-HI with good accuracy; the RF model performed best (area under the curve 0.818, classification accuracy 0.803, F1 0.79, Precision 0.85, Recall 0.74) and identified maximum neutrophil count and minimum albumin level as the most important features contributing to the classification. Conclusion Patients with hyperinflammation defined by CRP and ferritin thresholds share other global derangements in laboratory markers, suggesting shared pathology. Outcomes are less good in this patient group. The importance of neutrophils to the models is consistent with an association with COVID-19 disease severity and the possible contribution of neutrophil extracellular traps to pathology, while albumin is a known marker of inflammation. Disclosure M. Hutchinson: None. R. Tattersall: None. E.C. Jury: None. E. Hawkins: None. J. Manson: None.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call