Machine Learning-Based Prediction Study of Hematoma Enlargement in Patients with Cerebral Hemorrhage

• Published in: Journal of sensors Vol. 2022; pp. 1 - 7
• Main Authors: Tingting, Zhao, Kailun, Chen, Yiqing, Lu
• Format: Journal Article
• Language: English
• Published: New York Hindawi 29.09.2022; John Wiley & Sons, Inc
• Subjects: Algorithms; Artificial intelligence; Artificial neural networks; Breast cancer; Classification; Computed tomography; Computer science; Deep learning; Diabetes; Diabetic retinopathy; Disease; Enlargement; Hematoma; Hemorrhage; Image segmentation; Machine learning; Medical imaging; Multilayers; Neural networks; Performance evaluation; Performance prediction; Physicians; Sensitivity; Signs; Skin cancer; Tomography
• ISSN: 1687-725X; 1687-7268; 1687-7268
• DOI: 10.1155/2022/4470134

The enlarged hematoma is closely related to the poor neurological prognosis of patients with cerebral hemorrhage (intracerebral hemorrhage, ICH). Therefore, it is of great clinical significance to accurately predict whether ICH patients. In this study, we explored the predictive ability of a two-model machine learning, machine learning (ML) method to predict hematoma expansion. This method features nonenhanced CT (noncontrast computed tomography, NCCT) in ICH patients. The image information is combined with multiple clinical data for the prediction of hematoma enlargement. We retrospectively collected 140 ICH patients (including 58 patients with hematoma enlargement) from our hospital in 2021, and obtained a total of 5616 NCCT hematoma images (including 2635 images of hematoma enlargement) and 10 clinical data for each patient. The dual-model ML method used in this study contains 2 steps. The first step is to use a single-model predictor based on deep convolutional neural network (DCNN), which uses only the patients. The baseline NCCT images were performed for the prediction of hematoma enlargement. To select an appropriate DCNN model, we simultaneously compared the prediction performance of the three DCNN models, including ResNet34 (residual neural network with 34 layers), VGGNet (visual geometry group network), and GoogLeNet (Google inception network). In this step, we also explored whether the method of hematoma segmentation could improve the prediction outcome. The second step is to use the dual-model predictor based on multilayer perception (MLP), where the results of the single-model predictor in step 1 are combined with multiple clinical data of the patient to predict the final result. The sensitivity, specificity, positive predictive value, and negative predictive value were calculated for each model, and were predicted using the subject operating characteristic curve (the receiver operating characteristic, ROC) and area under curve (AUC) to evaluate prediction performance. The experimental results show that the ML method proposed in this study can comprehensively analyze the patient NCCT image information and clinical data, which can achieve 86.5% accuracy and have relatively equal sensitivity and specificity. Therefore, this ML method can be used as a predictive tool to effectively identify people at high risk of hematoma expansion. This study can make an effective prediction of hematoma expansion in patients with clinical cerebral hemorrhage, which can better treat patients, improve the doctor-patient relationship, reduce complications, etc.