Researches Of Machine Learning For Diagnosis And Prediction Of Prognosis Or Surgical Complications In LVO

Background : Large vessel occlusion(LVO)has a higher rate of disability and death,With a heavy burden on patients,their families and society.The most effective treatment for LVO is to recanalize the occluded vessel as soon as possible.Nevertheless,the treatment window for LVO is limit and a substantial number of patients are delayed or do not receive effective treatment within the time window for a variety of reasons.Despite the attainment of reperfusion,a significant majority of patients continue to experience a poor prognosis.Nonetheless,the efficacy of endovascular therapy is highly time-dependent.Therefore,a rapid diagnosis of LVO is crucial for improving patient prognosis.In patients with LVO treated with mechanical thrombectomy,the prediction of functional outcomes,including prognosis,the occurrence of futile recanalization(FR),and the presence of symptomatic intracranial haemorrhage(SICH),will provide individualized stroke management,treatment modalities,and help clinicians make treatment decisions.Objective: Using original CTA images and clinical data from LVO patients,this study aims to develop models for LVO diagnosis,prognosis prediction,FR prediction,and SICH prediction.Using an approach based on machine learning,the models aim to improve the efficiency of LVO diagnosis and aid clinicians in screening patients who may benefit from mechanical thrombectomy and guiding subsequent treatment.Methods:1.The pre-treatment CTA images of 8650 ischemic stroke patients were obtained retrospectively from Rhode Island Hospital,Brown University,USA,and used artificial intelligence to automatically screen out the original images of thin-layer transverse CTA sequences to construct the diagnostic model of LVO.Simultaneously,an "end-to-end" artificial intelligence stroke diagnosis and archive platform,DIANA,was developed and connected to the hospital PACS system to validate the LVO diagnosis model.The study also contrasted the speed of the traditional LVO diagnostic process to the feasibility of developing an AI-assisted clinical decision-making platform.Sensitivity,specificity,negative predictive values,and AUC curves were used to evaluate model performance.2.We screened the preoperative CTA imaging data and clinical data of 323 LVO patients with mechanically thrombectomy from patients collected in 1 and constructed a 3D CNN model to predict the short-term prognosis(m RS≤2 at 30 days after surgery)of endovascular treatment in LVO patients.Utilizing sensitivity,specificity,and AUC curves,model performance was evaluated.In addition,we retrospectively collected clinical data from a total of 771 LVO patients from Hunan Provincial People’s Hospital,Chenzhou First People’s Hospital and Xiangya Hospital of Central South University,respectively.The variables were screened by LASSO regression,a logistic regression model was constructed for the purpose of short-term prognosis prediction.In addition,clinical data collected from 323 patients at Rhode Island Hospital were utilized for external validation,and model performance was evaluated using AUC curves,calibration curves,and DCA curves.3.We collected 465 anterior circulation LVO patients with reperfusion(m TICI≥2b)from Hunan Provincial People’s Hospital,Chenzhou First People’s Hospital.And we used a total of 37 clinical variables to develop the FR prediction model.“Select best” was used to select 10 clinical variables with the highest correlation,and automatic machine learning method was used to construct and screen out the optimal FR prediction model.The model was then externally validated using 80 LVO patients at Xiangya Hospital,Central South University.The outcome of the model was whether the patients suffered FR(defined as LVO patients with revascularization after endovascular treatment and had an m RS>2 at 90 days postoperatively).We also used 27 preoperatively variables to construct the FR prediction model,and compare the performance of the two models.In this study,model performance was evaluated by AUC curves,calibration curves,and DCA curves.4.We retrospectively collected clinical data from a total of 664 LVO patients received endovascular treatment in Hunan Provincial People’s Hospital and Chenzhou First People’s Hospital.Using the method described in 3,the optimal SICH prediction model was constructed and screened(defined as intracranial haemorrhage within 48 hours after endovascular treatment,detected by imaging,consistent with the Heidelberg haemorrhage staging,and where the patient suffered worsening clinical symptoms).109 LVO patients received endovascular treatment at Xiangya Hospital,were then used as externally validate the model.In this study,model efficacy was evaluated using AUC curves,calibration curves,and DCA curves.Results:1.We effectively developed a LVO diagnose model with AUC of0.74,sensitivity of 61%,specificity of 74%,and negative predictive value of 93%.Meanwhile,the DIANA system was deployed on the computer network at Rhode Island Hospital and initially tested behind the scenes(clinicians were not informed of the predictions).The system predicted the head and neck thin-layer CTA images of 411 patients with AUC of0.428,sensitivity of 81%,specificity of 14%,and negative predictive value of 88%.Comparing the average latency of the DIANA system to that of the traditional LVO diagnostic procedure,we discovered that the DIANA system was faster(27.57 vs 41.11 min,p=0.0027).2.This study constructed a 3D CNN model for predicting the short-term prognosis of LVO patients undergoing mechanical thrombectomy based on preoperative thin-layer CTA images by deep learning algorithms.Then,9 clinical variables were screened by LASSO regression,and a logistic regression model of short-term prognosis was constructed.This model had AUC area of 0.797 on the training set and0.844 on the external validation set.High consistency existed between the probability of prediction and the probability of occurrence of the outcome event in the calibration curve model for this model.When p=0.04-0.91 on the DCA curve of the external validation set,using this model to predict the short-term prognosis has net benefit.3.We successfully constructed fourteen models for the prediction of FR,with the model based on the XGBoost algorithm performing the best on the internal test set.In the internal test set and external validation set,the AUC of the model is 0.891 and 0.826,respectively.The two models’ calibration curve sufficiently exhibits the high consistency.The DCA curves also indicate that using the model to predict FR has net benefit at prediction probabilities p=0.11-0.86 and p=0.49-0.97 for two different models,respectively.Comparing the two models’ performance on the external validation set,we found no significant difference in their predictive ability(p=0.3395).4.In this study,11 SICH prediction models were developed.After screening,the Neural Net MXNet performed the best,with AUC values of0.81 in the internal test set and 0.71 in the external validation set.The calibration curve of this model has certain fluctuations.In the range of prediction probability p=0.16-0.63,the DCA curve demonstrated that using this model to predict SICH has net benefit.Conclusions:1.We developed machine learning models which can be used to achieve diagnosis of LVO,the prognosis prediction of endovascular treatment,FR prediction,and SICH prediction.2.Concurrently,the "end-to-end" AI LVO diagnosis platform could enhance the efficiency of LVO diagnosis and is crucial for optimizing the stroke green channel diagnosis procedure.3.The model with preoperative variables can predict FR of LVO patients,and help neurointerventionists screen the potential beneficiaries of endovascular treatment.