ABSTRACT
The severity of COVID-19 varies dramatically, ranging from asymptomatic infection to severe respiratory failure and death. Currently, few prognostic markers for disease outcomes exist, impairing patient triaging and treatment. Here, we train feed-forward neural networks on electronic health records of 819 confirmed SARS-CoV-2 positive patients admitted to a two-site NHS Trust hospital in London, England. To allow early risk assessment, the models ingest data collected in the emergency department (ED) to predict subsequent admission to intensive care, need for mechanical ventilation and in-hospital mortality. We apply univariate selection and recursive feature elimination to find the minimal subset of clinical variables needed for accurate prediction. Our models achieve AUC-ROC scores of 0.78 to 0.87, outperforming standard clinical risk scores. This accuracy is reached with as few as 13% of clinical variables routinely collected within the ED, which increases the practical applicability of such algorithms. Hence, state-of-the-art neural networks can predict severe COVID-19 accurately and early from a small subset of clinical variables. © 2021, ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering.
ABSTRACT
Introduction COVID-19 mortality rates are high, particularly in patients requiring invasive ventilatory support, developing a cytokine storm, or experiencing thromboembolic disease. Our goal was to determine if traffic-light driven, personalised care was associated with improved survival in acute hospital settings. Methods Outcomes were evaluated during two implementation phases of a real-time clinical decision support tool that had been developed as part of a Trust's COVID-19 response, using a reporting and bioinformatics team to support Clinical and Operational teams. Following optimisation, the tool defined patients' clinical status in terms of risk of preventable complications based on blood test results (Ddimer, C reactive protein and ferritin). Feedback to wardbased clinicians enabled rapid modification of care pathways, in the first phase following a daily review, and in the second phase, in real-time (dashboard updated every 10 minutes). Results 1039 COVID-19 positive patients were admitted by 21/05/2020. Focusing on the first 939 completed encounters to death or home discharge (median age 69ys;60% [563/939] male), 568/939 (60.4%) received thromboembolism risk flags, and 212/939 (22.5%) cytokine storm flags. The maximum thromboembolism flag discriminated completed encounter mortality between no flag (9.97% [37/371]);medium-risk (28.5% [68/239]);high-risk (51.2% [105/205]);and suspected thromboembolism (52.4% [65/124]), Kruskal Wallis p<0.0001. 173 of 535 consecutive COVID-19 positive patients whose hospital encounter completed before real-time introduction died (32.3% [95% confidence intervals 28.0, 36.0]), compared to 46 of 200 (23.0% [95% CI 17.1, 28.9]) admitted after implementation of real-time traffic light flags (p=0.013). The realtime cohort were older (median age 72ys compared to 67ys, p=0.037), and were more likely to flag at risk of thromboembolism on admission. However, adjusted for age/sex, the probability of death was 0.33 (95% confidence intervals 0.30, 0.37) before real-time implementation, and 0.22 (0.17, 0.27) after real-time implementation (p<0.001). In subgroup analyses, older patients, males, and patients with hypertension (p£0.01) and/or diabetes (p=0.05) derived the greatest benefit from admission under the real-time traffic light system. Conclusion Personalised early interventions were associated with a reduction in mortality. We suggest benefit predominantly resulted from early triggers to review/enhance anticoagulation management, without exposing lower-risk patients to potential risks of full anticoagulation therapy.