[Retracted] Prediction and Risk Assessment Models for Subarachnoid Hemorrhage: A Systematic Review on Case Studies

Table 1

Comparative analysis of the prediction methods used for detection of SAH.

Method	Dataset	Advantages	Limitations
Applied random kernels to extract the features from physiological time series features [76].; Classifiers such as kernel SVM and partial least square were applied for prediction.; The model was evaluated on the 488 consecutive data from a tertiary care hospital.	Dataset is collected from Columbia University Medical Center.	The evaluation shows that random kernel and kernel SVM has considerable performance in the prediction.	The feature relations are not effectively analyzed due to applied kernel function and SVM has lower performance in handling imbalance dataset.; SVM requires more data instances to develop the optimal hyperplane.
-nearest neighbor (KNN) was applied to impute the missing data [77].; Classifiers such as deep learning, gradient boosting machine, gradient boosting machine, and generalized linear modeling were used.; The method was evaluated in 300 patient datasets.	Dataset is collected from the National Institute of Health Stroke Scale.	The analysis shows that the developed model has considerable performance in the prediction.; The model has considerable performance in a large dataset.	KNN method computes the distance between the new data instance and the existing data instance. The distance calculation for high dimensions creates the overfitting problem in training. The overfitting problem has affected the performance of the model.
Deep learning and Grapcut-based segmentation were carried out for segmentation of SAH [78].; The deep learning method is applied for feature extraction, and the softmax method is applied for classification.; The Benchmark dataset was used to evaluate the performance.	Physionet benchmark dataset	The deep learning and Grabcut method have higher efficiency in the segmentation.; The model has higher performance in the benchmark dataset.	The convolution and upsampling of data in the deep learning model creates an overfitting problem. The model has an overfitting problem that affects the performance.; The model has lower efficiency in imbalance dataset due to model requires more data instance for the class to train.
Correlation with the clinical and radiologic findings are analyzed in the model [79].; A linear mixture model is used to evaluate the performance of the method.; The dataset with 64 patient data was used to evaluate the performance.	Sahlgrenska University Hospital, Gothenburg, Sweden	This model has considerable performance in the analysis.	The model has a lower efficiency in analyzing the nonlinear relation in the features.
Convolutional neural network (CNN) model is applied for the segmentation of SAH [80].; Voxel-wise segmentation was used for the segmentation.; This model is evaluated on two datasets, and performance is analyzed.	Collaborative European neurotrauma effectiveness research in TBI study	The CNN has the higher efficiency in the segmentation.; The model has the higher performance in the analysis.	The CNN model convolution and pooling process create more data for the training, and this creates an overfitting problem. The model has an overfitting problem that affects the performance of the model.
A multilevel linear regression model is applied for the relational analysis of heart rate variability (HRV) and SAH [81].; EEG data of SAH patients were collected to evaluate the performance of the model.	Columbia University Medical Center Institutional Review Board	The model has a higher performance in the feature analysis.	Relevant features were required to be extracted to analyze the HRV, and effective classifier is required to analyze the relation between HRV and SAH.
The elastic net logistic regression model is applied for the prediction of SAH [82].; EEG and clinical factors data were used for the prediction.; This model is evaluated in a large dataset to analyze the performance.	Nonelective cEEG at Yale University/Yale New Haven Hospital, Brigham and Women’s Hospital, or Emory University Hospital	The model has considerable performance in the prediction.; This method effectively analyses the EEG data.	The feature relations in EEG and clinical factors are not effectively analyzed.
Random forest with conditional inferences trees were optimized to predict the SAH [83].; The feature importance is analyzed in this model.; The dataset of 630 SAH patients was used to evaluate the performance of the model.	World Federation of Neurosurgical Societies (WFNS)	The random forest-based model has considerable performance in the dataset.; The feature importance helps to improve efficiency.	The random forest model is ineffective when a number of trees is more and has overfitting when a number of trees is less.
The faster recurrent convolution neural network (RCNN) model is applied for cerebral aneurysm detection in CT images [84].; The collected dataset is used to evaluate the performance of the model.	Dataset collected from three medical centers.	The deep learning method has a higher performance in detection.	The faster RCNN model generates more data instance to train the network that creates an overfitting problem. Overfitting problem affects the performance of the model.; The parameter selection process requires for optimal performance of a network.
Logistic regression model for prediction of SAH and evaluated based on Glasgow Outcome Scale (GOS) [85].; The collected SAH patient dataset was used to evaluate the performance.	Massachusetts General Hospital aSAH database	The evaluation shows that model has the considerable performance.; The model has higher efficiency in the feature analysis.	The nonlinear relation between the features of SAH is required for analysis for effective performance. The model has lower efficiency in feature relation analysis.
Method: a multivariate logistic regression model; Dataset: in-hospital; Mortality in patients with sub arachnoid hemorrhage [86].; Variables relating to their demographic characteristics (age and sex), comorbidities (assessed by Charlson index), SAH (delayed hospital arrival and symptomatology at onset), and severity	Dataset collected from the University Hospital Complex of A Coruña (Spain).	A multivariate logistic regression model was developed to predict the likelihood of in-hospital mortality, adjusting it exclusively for variables present on admission. A predictive equation of in-hospital mortality was then computed based on model’s coefficients, along with a point-based risk-scoring system.	Although a multicollinearity analysis was performed, it cannot be ruled out that this issue could have influenced the associated effect sizes and maybe the associations themselves. Finally, this study only represents the clinical experience at our hospital, and so, our results must be validated externally with an independent cohort.
The acute infarction area of diffusion-weighted imaging (DWI) and hypoperfusion of perfusion-weighted imaging (PWI) was labeled manually. Two forms of datasets (volume of interest [VOI] data sets and slice data sets) were analyzed, respectively [87].	Nanjing First Hospital and the Affiliated Jiangning Hospital of Nanjing Medical University	The developed DMTC model in an independent external validation set, and comparing it with the training set, both the VOI data set and the slice data set had good performance in predicting HT, which showed good generalization ability.	First, the sample size is relatively small. Finally, as a result of the sample size, patients receiving bridging therapy were also enrolled, Health Quality Ontario demonstrated that EVT did not show an increased incidence of clinically relevant HT in comparison with IVT.
Multilayer perceptron (MLP), Naïve Bayes, and SVM methods were applied for the prediction process [88].	CT scan images of Charité Universitaetsmedizin Berlin were used to test the performance.	The developed method has considerable performance in prediction process.	The developed model has the imbalance data problem.
Logistic regression model is applied for the prediction process [89].	Teaching hospital in Barcelona (Spain)	The developed method has considerable performance in the risk factor analysis.	The developed method has lower performance in assumption of linearity between dependent and independent variables.