Forecasting readmission in COVID-19 patients utilizing blood biomarkers and machine learning in the Hospital-at-Home program

Abstract

Objectives: During the coronavirus disease 2019 (COVID-19) pandemic, the Hospital-at-Home (HaH) program played a key role in expanding healthcare capacity and managing COVID-19 pneumonia. This study aims to evaluate the factors contributing to readmission from HaH to conventional hospitalization and to apply classification algorithms that support discharge decisions from conventional hospitalization to HaH.

Methods: Blood biomarkers (IL-6, Hs-TnT, CRP, ferritin, and D-dimer) were collected from 871 patients transferred to HaH after conventional hospitalization for COVID-19 at the Hospital Universitari Germans Trias i Pujol. Of these, 840 patients completed their recovery without any complications, while 31 of them required readmission. Statistical tests were conducted to assess differences in blood biomarkers between the first day of conventional hospitalization and the first day of HaH, as well as between patients who successfully completed HaH and those who were readmitted. Various classification algorithms (bagged trees, KNN, LDA, logistic regression, Naïve Bayes, and the support vector machine [SVM]) were implemented to predict readmission, with performance evaluated using accuracy, sensitivity, specificity, F1 score, and the Matthews Correlation Coefficient (MCC).

Results: Significant differences were observed in IL-6, Hs-TnT, CRP (p < 0.001), and ferritin (p < 0.01) between the first day of conventional hospitalization and the first day of HaH for patients who were not readmitted. However, no significant differences were found in patients who were readmitted. At HaH, readmitted patients exhibited higher CRP and Hs-TnT values. Among the classification algorithms, the SVM showed the best performance, achieving 85% sensitivity, 87% specificity, 86% accuracy, 84% F1 score, and 71% MCC.

Conclusion: Hs-TnT was a key predictor of readmission for COVID-19 patients discharged to HaH. Classification algorithms can aid clinicians in making informed decisions regarding patient transfers from conventional hospitalization to HaH.

Keywords: COVID-19; Hospital-at-Home program; Hs-TnT; biomarkers; machine learning.

Figure 1

Distribution of COVID-19 patients admitted to the Hospital Universitari Germans Trias i Pujol.

Figure 2

Boxplot representation of the variables IL6, Hs-TnT, CRP, Ferritin and DDIMER the first day of conventional hospitalization and the first day of hospital at home for (a) D1 group and (b) D2 group.

Figure 3

Boxplot representation of the variables IL-6, Hs-TnT, CRP, ferritin, and D-dimer at HaH for non-readmitted (D1) and readmitted (D2) patients.

Figure 4

Results of data augmentation for patients returning to conventional hospitalization from HaH. Red dots represent original data values, while blue circles indicate the synthetic data created.

Figure 5

Mean and standard deviation obtained from the 5-fold cross-validation process for the algorithms with an accuracy above 80% (bagged trees, KNN, and the SVM).

Figure 6

Classification metrics for the blind test data using the SVM-trained model, along with the confusion matrix showing accuracy, sensitivity, specificity, F1 score, and MCC obtained.

Figure 7

Diagram illustrating the integration of machine learning model outcomes into clinical decision-making.