Predicting clinical outcomes at hospital admission of patients with COVID-19 pneumonia using artificial intelligence: a secondary analysis of a randomized clinical trial

Abstract

Background: Predicting clinical improvement after hospital admission in patients with COVID-19 is crucial for effective resource allocation. Machine-learning tools can help identify patients likely to show clinical improvement based on real-world data. This study used two approaches-least absolute shrinkage and selection operator (LASSO) and CombiROC-to identify predictive variables at hospital admission for detecting clinical improvement after 7 days.

Methods: A secondary analysis was conducted on the modified intention-to-treat placebo group from a previous randomized clinical trial (RCT, NCT04561219) of patients with COVID-19. The analysis assessed clinical, laboratory, and blood markers at admission to predict clinical improvement, defined as a two-point increase on the World Health Organization clinical progression scale after 7 days. LASSO and CombiROC were used to select optimal predictive variables. The Youden criteria identified the best threshold for different variable combinations, which were then compared based on the highest area under the curve (AUC) and accuracy. AUCs were compared using DeLong's algorithm.

Results: Overall, 203 patients were included in the analysis, and they were divided into two groups; clinical improvement (n = 154) and no clinical improvement (n = 49). The median age was 55 years (interquartile range, 46-66 years); 65% were male. LASSO identified three predictive variables (SaO₂, hematocrit, and interleukin [IL]-13) with high sensitivity of 98% (95% confidence interval [CI], 92-100%) but low specificity of 26% (95% CI, 10-48%) for clinical improvement. CombiROC selected a broader set of variables (T cell-attracting chemokine, hemoglobin, hepatocyte growth factor, hematocrit, IL-3, PDGF-BB, RANTES, and SaO₂), achieving balanced sensitivity of 82% (95% CI, 69-91%) and specificity of 74% (95% CI, 49-91%). LASSO and CombiROC showed comparable accuracy (82 and 80%, respectively) and similar area under the ROC curves (LASSO: AUC, 0.704; 95% CI, 0.571-0.837; CombiROC: AUC, 0.823; 95% CI, 0.708-0.937; p = 0.185).

Conclusion: For patients hospitalized with COVID-19 pneumonia, predictive variables identified by LASSO and CombiROC analyses demonstrated similar accuracy and AUCs in predicting clinical improvement. LASSO, with fewer variables (SaO₂, hematocrit, and IL-13) showed high sensitivity but low specificity, whereas CombiROC's broader selection of variables provided balanced sensitivity and specificity for predicting clinical improvement.

Clinical trial registration: Brazilian Registry of Clinical Trials (REBEC) number RBR-88bs9x and ClinicalTrials.gov number NCT04561219.

Keywords: COVID-19; CombiROC; LASSO; biomarkers; clinical improvement; machine learning.

Figure 1

Flowchart of the study. CTACK, T cell–attracting chemokine; GM-CSF, granulocyte-macrophage colony-stimulating factor; Hb, hemoglobin; HGF, hepatocyte growth factor; HR, heart rate; Htc, hematocrit; IFN, interferon; IL, interleukin; LDH, lactate dehydrogenase; MCP, monocyte chemotactic protein; MIG, monokine induced by IFN-γ; MIP, macrophage inflammatory protein; mITT, modified intention-to-treat; PDGF, platelet-derived growth factor; RCT, randomized clinical trial; RR, respiratory rate; SCF, stem cell factor.

Figure 2

(A) ROC curve of variables selected by LASSO; (B) confusion matrix of variables selected by LASSO. AUC, area under the curve; CI, confidence interval.

Figure 3

(A) ROC curve of variables selected by CombiROC; (B) confusion matrix of variables selected by CombiROC. AUC, area under the curve; CI, confidence interval.

Figure 4

ROC curves of the LASSO model and the top 5 obtained after CombiROC analysis. AUC, area under the curve; CI, confidence interval.