Machine learning model based on clinical features to diagnose unexplained dizziness with patent foramen ovale

Abstract

Objectives: Until recently, unexplained dizziness has received little attention. Clinicians have troubles to offer for the prevention and treatment of unexplained dizziness. This study aims to develop a machine learning model for diagnosing unexplained dizziness with patent foramen ovale and investigate key clinical features of this disorder. Methods: A cross-sectional, observational study of adults referred for unexplained dizziness enrolled from January 2019 to June 2023 at our tertiary hospital was undertaken. Subjects were split into dizziness with PFO group and non-PFO group. Demographic and clinical characteristics were collected. We randomly sampled data into training and validation sets by 7:3. LASSO regression was used to select the saliently important features in the training sets. Seven machine learning algorithms were used to develop a machine learning model for diagnosing unexplained dizziness with PFO. The area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, and F1-score were measured to evaluate the performance of models. The SHapley Additive exPlanations (SHAP) value was used to visualize the importance of key factors. Results:233 unexplained dizziness patients met the criteria for PFO,160 unexplained dizziness patients without PFO. Dizziness triggers and aggravating factors, recurrent episodes, chest tightness, amaurosis, palpitation, chronic persistent and age were selected according to LASSO regression in the training sets. Repeated cross-validation showed that the LGB model had the highest AUROC of 0.871(95%CI 0.86,0.882), AUPRC of 0.886(95%CI 0.878,0.894) among seven models. The model had high comprehensive performance and strong predictive power. SHAP analysis demonstrated the visual interpretation of the optimal model and revealed how the interaction of age and other features in the model affected the diagnosis of PFO. Conclusions: The LGB diagnosing model could effectively and reliably diagnose unexplained dizziness with PFO. SHAP analysis could improve the visual interpretation of machine learning models and help physicians better understand how the features affect outcomes.