| Background and Purpose:Acute Ischemic Stroke（AIS）is a serious health hazard due to its high morbidity,disability,mortality and recurrence rate.Intravenous thrombolytic with Recombinant Tissue Plasminogen Activator（rt-PA）within 4.5 hours after onset is the most widely accepted treatment for the disease.There is a strict time window for this treatment because patients given intravenous thrombolytic rt-PA after the time window are at significantly increased risk of intracranial hemorrhage.However,in the real world,some patients who are suitable for rt-PA intravenous thrombolytic fail to receive the treatment due to uncertain onset（such as awakened stroke or patients who cannot provide an exact medical history）,resulting in poor prognosis.For AIS patients with unknown onset time,whether the onset time of stroke is within 4.5 hours can be roughly determined by reading the mismatch information in Diffusion Weighted Imaging（DWI）and fluid attenuation inversion recovery,but it is difficult to accurately distinguish them by conventional manual interpretation.Machine learning is the study of how computers acquire new knowledge by imitating the way humans learn,and then reorganized the knowledge to improve their performance.With the continuous improvement of computer technology,machine learning has been widely used in medical image assisted diagnosis.At present,a number of studies have reported machine learning methods combining multimodal images to identify AIS patients within 4.5 hours,but there are relatively few reports on machine learning methods that only use DWI images to assess the onset time of AIS patients.Based on the DWI images of AIS patients with clear onset time,this study trained machine learning models that could be used to identify AIS patients with onset time within4.5 hours,in order to assist clinicians to effectively evaluate whether AIS patients with unknown onset time are suitable for rt-PA intravenous thrombolytic.Methods:This study was a retrospective study.A total of 227 AIS patients admitted to China-Japan Friendship Hospital of Jilin University from January 2021 to July 2021 with definite onset time and receiving DWI examination within 24 hours after onset time were included.According to the time from onset to DWI examination,patients are divided into two categories: onset ≤ 4.5 hours（70 cases）and onset > 4.5 hours（157 cases）.The two groups of patients are divided into a training sets（158 cases,including 49 cases ≤ 4.5 hours,109 cases > 4.5 hours）and a test sets（69 cases,including 21 cases ≤ 4.5 hours,48 cases > 4.5 hours）by completely randomized method in a rate of 7:3.DWI images of patients were collected and regions of interest were divided on DWI using ITK-SNAP image labeling software.The Pyradiomics software package based on Python was used to extract multiple image features from the region of interest images.After Z-score regularization processing for each feature value,features with high correlation degree were excluded through Spearman correlation test.Least absolute shrinkage and selection operator model combined with 10-fold cross-validation was used to screen out the optimal characteristics with significant differences between the two groups of patients.Based on the selected optimal features,in the training set,seven machine learning algorithms such as Support Vector Machine,K-adjacency,Decision Tree,Extra Tree,Random Forest Tree,Extreme Gradient Boosting（XGBoost）,and Light Gradient Boosting（Light GBM）are applied to train the models.In the test set,the area under the receiver operating characteristic curve was used to evaluate the performance of classifiers.The best model of each algorithm was selected and compared after 100 times of random verification.Results:1.107 features were extracted from the DWI image using Pyadiomics,and 22 optimal features were screened out by the Spearman correlation test and the least absolute shrinkage and selection operator model,including 4 first-order features,6morphological features,and 12 texture features.2.Among the 7 models,the XGBoost model had the best recognition effect on AIS patients within 4.5 hours,with the area under the receiver operating characteristic curve being 0.817,and the accuracy,sensitivity and specificity were0.739,0.733 and 0.814.Conclusion:1.DWI imaging is meaningful in assessing the onset time of AIS patients.2.Machine learning models based on DWI imaging outperform traditional manual methods at identifying AIS patients within 4.5 hours.3.The XGBoost model based on DWI imaging can effectively identify AIS patients within 4.5 hours,which has certain clinical transformation of value for assisting clinician to assess the onset time of AIS patients. |
