A team of researchers have developed a machine learning (ML)-based algorithm to analyse electronic health record (EHR) data and reliably predict near-term mortality in COVID-19 patients.
COVID-19 mortality
Since its emergence, COVID-19 has rampaged across the world, causing a rise in hospitalisations and intensive care unit admissions. The varied impact this virus has had on individuals has astounded the scientific and research community. The importance of effective prognostication has become clear. Having a clear prognosis can enable clinicians, who are already working with limited resources, to develop an appropriate plan of care based on individuals’ varying risks of deterioration. Most importantly, timely and targeted delivery of palliative care services for patients is essential. Nevertheless, determining this is difficult. This is due to the largely unpredictable disease trajectory of COVID-19 and how quickly patients’ conditions can deteriorate.
Since the start of the pandemic, it has become clear that age and pre-existing health conditions were baseline predictors of mortality. However, individuals without pre-existing health conditions have also needed hospitalisation, required ICU care or died (26%, 23% and 5%, respectively). Therefore, this suggests using these baseline risk factors to assess mortality risk may have limited clinical utility in this context.
ML-based approach
In this study, published in BMJ Supportive and Palliative Care, researchers developed a novel supervised ML-based prediction tool to help clinical teams identify COVID-19 patients at higher risk of near-term in-hospital mortality. Determining this could assist palliative care clinicians in deciding when to have difficult conversations with families regarding prognosis and care for these patients. Specifically, the team used inpatient time-series data from the EHR system and then applied a random forest (RF) approach. Their cohort comprised of 567 patients with COVID-19 at a large acute care healthcare system.
The model yielded a sensitivity of 87.8%, specificity of 60.6%, accuracy of 65.5% and area under the receiver operating characteristic curve of 85.5%. This model provided adequate discrimination without the need for manual pre-processing of data. Unlike using static variables, this model translated the variability in patients’ conditions into mortality risk predictions.
The team’s ML-based approach can be used to analyse EHR data and reliably predict near-term mortality. Using such a model within hospitals during this period could help improve care. In addition, this could aid with clinical decisions with prognosis in critically-ill patients with COVID-19.
Image credit: By wirestock – www.freepik.com