Predicting the Classification of Home Oxygen Therapy for Post-COVID-19 Rehabilitation Patients Using a Neural Network

Abstract

Objective: We evaluated the accuracy of a neural network to classify and predict the possibility of home oxygen therapy at the time of discharge from hospital based on patient information post-coronavirus disease (COVID-19) at admission.

Methods: Patients who survived acute treatment with COVID-19 and were admitted to the Amagasaki Medical Co-operative Hospital during August 2020-December 2021 were included. However, only rehabilitation patients (n = 88) who were discharged after a rehabilitation period of at least 2 weeks and not via home or institution were included. The neural network model implemented in R for Windows (4.1.2) was trained using data on patient age, gender, and number of days between a positive polymerase chain reaction test and hospitalization, length of hospital stay, oxygen flow rate required at hospitalization, and ability to perform activities of daily living. The number of training trials was 100. We used the area under the curve (AUC), accuracy, sensitivity, and specificity as evaluation indicators for the classification model.

Results: The model of states at rest had as AUC of 0.82, sensitivity of 75.0%, specificity of 88.9%, and model accuracy of 86.4%. The model of states on exertion had an ACU of 0.82, sensitivity of 83.3%, specificity of 81.3%, and model accuracy of 81.8%.

Conclusion: The accuracy of this study's neural network model is comparable to that of previous studies recommended by Japanese Guidelines for the Physical Therapy and is expected to be used in clinical practice. In future, it could be used as a more accurate clinical support tool by increasing the sample size and applying cross-validation.

Keywords: COVID-19; Classification; Home oxygen therapy; Machine learning; Neural network.

Fig. 1.

The neural network model x_i (i = 1, 2, 3, …, 7): input node. h_j1 (j = 1, 2, 3, 4, 5), h_k1 (k = 1, 2, 3): each intermediate node in the hidden layer. y: output node in the output layer. ω_ij, ω_jk, ω_ky: weights between each node of each layer. b_j, b_k, b_y: bias to each node of each layer

Fig. 2.

Schematic diagram of subject data flow and machine learning through the study

Fig. 3.

ROC curve of the ESO₂ model The classification performance of home oxygen therapy at rest at the time of discharge. ROC, receiver operating characteristic

Fig. 4.

ROC curve of the EWO₂ model The classification performance of home oxygen therapy during exertion at the time of discharge. ROC, receiver operating characteristic