VentNet: A Real-Time Multimodal Data Integration Model for Prediction of Respiratory Failure in Patients with COVID-19

Abstract The COVID-19 pandemic has led to massive challenges for health care systems and for global economics. The surge in cases which occurred abruptly strained the existing resources to care for the volume of patients, leading to a shortage of supply in many medications, personnel and equipment. The mechanical ventilator became a particular problem as one newly published study reported that 18 out of 24 patients with COVID-19 in the study (75%) required mechanical ventilation. During the early months of the pandemic many providers decided to intubate early on the assumption that patients would eventually need mechanical ventilation so as to avoid ‘crash intubation’ and potential contamination. A recent observational study of intensive care unit patients with COVID-19 suffering from acute hypoxemic respiratory failure revealed that early invasive mechanical ventilation was associated with an increased risk of day-60 mortality. One central problem in this context was caregivers’ inability to predict which patients may need mechanical ventilation since existing methods using clinical parameters are often subjective and inconsistent across different institutions. We have thus applied machine learning algorithms to commonly available data in electronic health records (EHR) to develop and validate a predictive model for 24-hours ahead prediction of respiratory failure. This novel predictive model has demonstrated AUCs in the range of 0.90-0.94 in our internal and external COVID-19 datasets. That is, we have a robust ability now to predict which patients may need mechanical ventilation and which will not. We are now planning to deploy clinically and to improve iteratively on our model by adding other data streams such as imaging to not only improve our predictive ability but also to make the predictions more ‘actionable’, so that clinicians can pursue timely interventions rather than just being told a prognosis. We are further addressing the many barriers to implementation by addressing ‘clinician buy-in’ which involves making the underlying reasoning of our algorithms more transparent, making the predictions seamlessly integrated into clinical workflow, and finding actionable parameters that will allow both predictions and therapeutic interventions. Such an algorithm will enhance the ability of clinicians to estimate the risk for respiratory failure, and ideally, to anticipate and respond to patient needs in a timely fashion. Moreover, given a long enough prediction horizon (48-72 hours) such systems can facilitate triage and optimization of related resources (ventilators and personnel) within a given hospital and across healthcare systems. Finally, while the COVID-19 pandemic highlighted the need for optimizing the timing of mechanical ventilation, the techniques developed under this proposal are broadly applicable to other causes of respiratory failure and to other types of organ support technologies.