Background: As a result of the COVID-19 pandemic, various clinical intervention methods and AI methods have been employed in the detection, diagnosis, and prognosis of COVID-19 cases. However, limited instances of applying AI to the prognosis of COVID-19 cases in Africa have been reported in the literature. Thus, case studies on the application of machine learning to guide for decision-making on the treatment of COVID-19 cases in Africa are essential. Methods: We applied three machine learning (ML) algorithms: Deep Multi-layer Perceptron (Deep MLP), Extreme Boosted Trees (XGBoost) and Support Vector Machines (SVM) for predicting the outcome of intensive care patients of COVID-19 with comorbidities in a South African hospital. We compared the performance and interpretability of the three ML models when cross-validation (CV) and principal component analysis (PCA) were applied for the prognosis of COVID-19 mortality risk. Results: We found that Deep MLP had the best overall performance when CV and SMOTE were applied without PCA (F1=0.92; AUC = 0.94), followed by SVM (F1=0.83; AUC=0.82). We found that the performance of both SVM and MLP can be enhanced through CV without PCA. XGBoost (F1= 0.81; AUC = 0.79) performed best when none of CV, PCA or SMOTE was applied. XGBoost is not affected by CV and performs worse with PCA. From the model predictions, we identified Length of stay in the hospital, Duration in ICU, Time to ICU from Admission, Days discharged or death, D-dimer (blood clotting factor), and blood pH as the six most critical variables for the prediction of mortality or survival of the COVID-19 patients. We also found other variables: Age at admission, Pf Ratio (PaO2/FiO2 ratio), TropT, Ferritin, ventilation, CRP, and Symptom of Acute respiratory distress syndrome (ARDS) associated with the severity and fatality of COVID-19 cases. Conclusions: This study demonstrates how ML can be applied to identify variables that have prognostic value in the treatment and management of critically ill COVID-19 patients. The findings also reveal the effect of CV and PCA when predicting clinical outcomes of COVID-19 cases
Background: In a normal regression analysis for determinants of TB outcomes, assumptions that the sample is homogenous is made. This model does not account for the overall effect of unobserved or unmeasured covariates. This study aims to quantify the amount of heterogeneity that exists at community level, and to ascertain the determinants of TB mortality across all the catchment areas in Lesotho. Methods: This was a retrospective record review of patients on TB treatment registered between January 2015 to December 2020 at 12 health care facilities in the district of Butha Buthe, Lesotho. Data collected from patient medical and statistical analysis was performed using R and INLA statistical software. Descriptive statistics were presented using frequency tables. Differences between binary outcomes were analysed using Person's X2 test. Mixed effect model with five Bayesian regression models of varying distributions were used to assess heterogeneity at facility level. Kaplan-Meier curves were used to demonstrate time-to-death events. Results: The total number of patients included in the analysis were 1729 of which 70% were males. And half of them were employed (54.2%). Being over 60 years (HR: 0.02, Cl: 0.01-0.04) and having a community health worker as a treatment contact person (HR: 0.36, Cl: 0.19-0.71) decreased the risk of dying. Miners had 1.73 times increased risk of dying (HR: 1.73, Cl: 1.07-2.78). The frailty variance was observed to be very minimal (<0.001), but significant indicating heterogeneity between catchment areas. Although similar hazard ratios and confidence intervals of covariates are seen between Gamma and Gaussian frailty log-logistic models, the credibility intervals for the Gamma model are consistently narrower. Conclusion: The results from both Gamma and Gaussian demonstrate that heterogeneity affected significance of the determinants for TB mortality. The results showed community level to significantly affect the risk of dying indicating differences between catchment areas.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
customersupport@researchsolutions.com
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.