Using Machine Learning Methods to Predict Hospitalization Based on Brixia Score and Patient Clinical Data (from the COVID-19 Pandemic)

<i>Background and Objectives:</i> The use of a standard chest X-ray has become a routine diagnostic method in daily clinical practice for the evaluation of a wide range of lung diseases. During the COVID-19 pandemic, significant challenges occurred in achieving accurate diagnostics and selecting appropriate therapies for patients with different symptoms of diseases. The aim was to cross-correlate radiological findings and clinical data and to develop models to predict hospitalization status, while evaluating the prognostic importance of the different variables. <i>Materials and Methods:</i> A set of variables including Brixia score, and clinical data: gender, age, hypertension, and diabetes was used to explore their association with patient hospitalization. Four different machine learning (ML) methods (<i>Decision Tree</i>-DT, <i>Logistic Regression</i>-LR, <i>Random Forest</i>-RF and <i>Support Vector Machine</i>-SVM) were used for hospitalization outcome prediction. <i>Results:</i> SVM appeared to be with the highest AUC (0.851), with low sensitivity, while DT was the most balanced in the context of AUC, accuracy, sensitivity, and specificity. Brixia score appeared to be the most important predictor for hospitalization within the group of predictors (gender, age, hypertension and diabetes). <i>Conclusions:</i> All four ML models that used in this study provided "good" prediction capabilities (AUC > 0.8), with the exception of SVM that had low sensitivity, emphasizing Brixia score as the strongest predictor of hospitalization. Application of ML methods have considerable potential in various aspects of medical clinical practice and future studies could potentially indicate the importance of applying the ML model in more precise diagnosis, therapy and prognosis of the patient's clinical condition.