During the recent COVID-19 pandemic, the health system worldwide has been heavily burdened by the sudden influx of patients. The increasing number of hospital admissions due to this pandemic’s vast prevalence has caused overburdening staff and shortages of medical resources 10–12. Under such circumstances, predicting the patient’s outcome can help the resource allocation and planning in the hospital, which is especially important when there is a massive influx due to a pandemic12.
ML has excellent potential in healthcare data analysis, as it allows us to analyze a large amount of data and, more importantly, to identify hidden trends and unknown interactions among different variables concerning the outcome 13. Several studies have used machine learning approaches for COVID-19 diagnosis and prognosis 14. However, the correlations between the models’ features are low 12. Olmedo et al. developed an ML-based algorithm to predict the mortality of COVID-19 patients in Spain during hospitalization. In their algorithm, LDH, CRP, neutrophils, UREA and age were the five most essential features 14. In an Italian ML-based algorithm, age, Hs-TnI, BUN, Alb, and BNP were the five most important features to predict in-hospital mortality in COVID-19 patients 13. In research from the Mayo Clinic, age was the most important feature in predicting 7-day, 14-day and 30-day mortality 15. A meta-analysis also emphasized the importance of age on mortality in patients with COVID-19 16. While age was reported as an essential feature in the literature, it was not included in our algorithm.
It could be inferred that a customized predictive algorithm that could correctly predict in-hospital mortality in COVID-19 patients would be more effective for the specific population.
The clinical spectrum of COVID-19 is vast 3. Physicians are often unable to accurately predict the prognosis of COVID-19 patients upon admission until later stages of the disease 4. In such circumstances, routine laboratory test results would be more objective than the clinical evaluation from the physicians.
We developed an ML-based algorithm for predicting in-hospital mortality in COVID-19 patients, namely MOP@COVID. The MOP@COVID used routine laboratory test results and could be assessed online.
Compared to the existing algorithms, in the MOP@COVID, CKMB had the highest weight, followed by APTT, PT, CRP, LY#, NT-proBNP, PCT, GLU, ALB and EO#.
Studies have shown that significantly higher cardiac markers, including CKMB, NT-proBNP, BNP, and troponin, were found in those who died compared with survivors of COVID-19 17. SARS-CoV-2 not only can exacerbate pre-existing cardiovascular disease but could also be responsible for new adverse cardiac events 18. As reflected by elevated cardiac troponin levels, acute myocardial injury was found in 7–36% of COVID-19 patients and correlated with increased in-hospital mortality 18. Individual cases of fulminant heart failure with severe systolic dysfunction were also reported 19,20. SARS-CoV-2 is proven to invade cardiomyocytes in vitro in a cathepsin directly- and ACE2-dependent way 21. In autopsy studies, high cardiac viral loads were found in 41% of subjects 22 and interstitial infiltrates of mononuclear immune cells, macrophages with viral particles, and endotheliitis were also reported 18,20,23,24. Both CKMB and NT-proBNP were included in the MOP@COVID features. All the cardiac biomarkers, including TNT, CKMB, MYO, and NT-proBNP, were higher in the MO group. It could be indicated that the cardiac event was an essential adverse event in our cohort. Cardiac biomarkers are necessary for risk assessment in COVID-19 patients.
Prolonged PT and APTT were associated with DIC and in-hospital mortality in COVID-19 patients 25,26. Through pathological examination of patients who died of COVID-19, researchers found that the virus can lead to disorders of the coagulation system, resulting in a hypercoagulable state and microthrombosis 27,28. Viral infections may induce severe complications, such as acute respiratory distress syndrome and multi-organ dysfunction syndrome, which are two conditions frequently associated with hypercoagulation and disseminated intravascular coagulation 29. D-dimer greater than 1 µg/mL was found to be associated with fatal outcomes of COVID-19 in an early study in China 30. High D-dimer levels were associated with an increased risk of thromboembolic events in patients with COVID-19 25,31. The MO group had a significantly higher D-dimer level in our data set. Although D-dimer was not included in the MOP@COVID, it could help to explain why PT and APTT were prolonged in patients with severe and critical COVID-19 27. PT and APTT were the first and third important features in the MOP@COVID. It could be inferred that coagulation disorder was an important adverse event in our cohort. Coagulation tests should be a concern in COVID-19 patients.
The plasmatic levels of PCT are usually undetectable in physiological conditions. It increases considerably in bacterial and fungal infections compared to viral infections, which makes it a differential diagnosis marker between them 32,33. Studies have shown that secondary bacterial and fungal infections, expressed by a high level of PCT, are associated with severe and potentially fatal forms of COVID-19, justifying the necessity of rational use of antibiotics in COVID-19 32. CRP is well-established as a marker of systemic inflammation and severe infection. CRP is strongly associated with VTE, AKI, critical illness, and mortality in COVID-19 34,35. PCT and CRP were included in the MOP@COVID. It could be indicated that infection and inflammation were important risk factors for in-hospital mortality in COVID-19 patients.
Studies have reported that lymphocyte and eosinophil levels were significantly lower in COVID-19 patients with critical diseases than those with moderate and severe diseases 36,37. A study in Wuhan, Hubei Province, China, showed that decreased lymphocytes were related to a higher risk of disease severity in male COVID-19 patients with hypertension 38. However, another study from Hubei province showed that a poorer prognosis was associated with elevated basophil or lymphocyte counts 39. In our data set, the MO group had increased levels of WBC and neutrophil, lymphocyte, monocyte, eosinophil, and basophil. This reemphasized the necessity to develop a predictive algorithm for the local population.
Glucose at admission ≥ 165 mg/dL was reported as being associated with an increased risk of in-hospital mortality and length of hospitalization in diabetic patients with COVID-19 40. In our data set, the MO group had higher glucose levels, while glucose was important for predicting in-hospital mortality in the MOP@COVID.
As we excluded all the cases that had missing values, the algorithm could indicate the true importance of the laboratory tests. In the training set, the MOP@COVID’s accuracy, F1 score, recall rate, sensitivity, specificity, NPV, and PPV were all above 0.98. Although the sensitivity was relatively lower (0.793) in the validation set, the specificity was relatively high (0.967). To ensure reliability, reproducibility, and robustness, MOP@COVID was further validated by an external validation cohort. The MOP@COVID yielded a sensitivity of 0.818, a specificity of 0.987, an accuracy of 0.973, and an AUC of 0.958 in the external validation cohort. It could be inferred that the MOP@COVID is effective in predicting in-hospital mortality in COVID-19 patients. With the webpage of MOP@COVID, it is accessible and feasible for clinical doctors and decision-makers.
For COVID-19 patients, predicting the risk of in-hospital mortality could support decisions on the level of healthcare required and more aggressive therapeutic interventions. Clinical doctors and healthcare providers can use the predictive results of our webpage to tailor management strategies for patients with COVID-19. The webpage tool allows them to easily type in the patient’s laboratory results and use the predictive results to develop the hospitalization rules for COVID-19 patients.
The MOP@COVID used routine laboratory test results at hospital admission and could predict the risk of in-hospital mortality. With the webpage tool, MOP@COVID could provide helpful information to clinical doctors and healthcare providers.