Explainable Artificial Intelligence-Based Prediction of Poor Neurological Outcomes by Cranial Computed Tomography Immediately After Resuscitation

Study design and population

This single-center, retrospective, observational study was conducted using data from the medical records of patients with out-of-hospital cardiopulmonary arrest who were transported to the Advanced Critical Care and Critical Care Center, Nara Medical University. This study was approved by the Institutional Review Board of Nara Medical University (No. 3131). As this is an observational study, the need for written informed consent was waived by the Ethics Committee of Nara Medical University.This research report follows TRIPOD guidelines¹⁷All methods in our research were performed in accordance with the tenets of the Declaration of Helsinki.

Consecutive out-of-hospital cardiopulmonary arrest patients who were hospitalized between April 2015 and March 2021 and treated with ROSC or extracorporeal circulation were included in this study. Patients <18 years of age; Patients with a history of head trauma, stroke, and intracranial disease, excluding lacunar infarction; Patients transferred from other hospitals; Patients (based on CT scan duration of previous study)²); patients with insufficient CT coverage. A patient who died during aortic rupture surgery immediately after cardiopulmonary resuscitation. Also excluded were patients who discontinued life-sustaining treatment after ROSC at the request of the patient or family. We do not discontinue life-sustaining treatment.

Post-resuscitation care

According to our standardized protocol, patients were managed with sedation, analgesia and ventilation according to resuscitation guidelines¹⁸An exclusion criterion for target temperature management (TTM) was shock (systolic blood pressure < 90 mmHg) despite use of vasopressors. Core body temperature was maintained at 33 °C for 24 h using the Arctic Sun® Temperature Management System (Bard, BD, Covington, GA, USA) and then rewarmed at a rate of 0.25 °C/h to 37 °C. maintained at another 24 hours. If TTM failed, other treatments were given as well.

Participant characteristics data collection

Data collected retrospectively from electronic medical records included age, sex, witnesses, bystanders, initial rhythm, cause of cardiac arrest, time from cardiac arrest to recirculation, time from recirculation to CT scan, and CPC after a month. Resuscitation of inpatients and outpatient follow-up of discharged patients.

Original neurological outcome assessment

Current recommendations define poor neurological outcome as CPC¹⁹ 3 to 5^20,21Based on the minimum acceptable time up to 1 month after resuscitation, patients with a favorable neurological outcome on head CT immediately after resuscitation and who died within 1 month were classified as CPC 5, This indicates a poor neurological outcome. Accurate labels are essential for training machine learning models on different data features. Since this study aims to predict neurological outcome from head CT image data, neurological outcome information should be reflected in the training data (i.e., head CT). Predictive criteria for CPC 1 or 2 are that the patient is awake, able to communicate, able to perform commanded movements without impairment, and no evidence of limb paralysis. To minimize the risk of misclassification in this study, patients classified as CPC 1 or 2 after CT imaging who died within 1 month of admission but not due to intracranial disease were Still classified as CPC 1 or 2. Although favorable neurological outcomes were presumed to be obtained after performing CT, it is not appropriate to a priori exclude them from studies aimed at predicting neurological outcome on the basis of head CT imaging. I thought no.

CT protocol and conditions for image acquisition

All CT images were acquired on a 64-row helical CT system (Optima CT660; GE Healthcare, Chicago, IL). Here are the scan settings: 120 kVp. automatic current; rotation time, 0.5 seconds; helical pitch, 0.531; noise index, 3.0. and image noise, SD10. Shortening the examination time is a priority for his CT scan after resuscitation. Therefore, in this study, reconstructed images with an orbital outer diameter baseline were used to reduce variability in imaging conditions. His CT images used for machine learning were of Monroe foramen and pineal level slices used in GWR-based studies. Images were acquired as portable network graphics of size 1 × 256 × 256 with a window level of 40 Hounsfield units (HU) and a width of 80 HU.

machine learning model

The prepared image dataset was stratified and split into training and validation datasets in a commonly used 8:2 ratio. The training dataset was then stratified to split the training and test datasets in a ratio of 8:2 and used to build the model. Validation data that were not used for training were used for model validation. VGG19^{twenty two} A machine learning model was used. It is a 19-layer convolutional neural network with transfer learning to apply parameters obtained by training on 1 million images (Figure 4). Transfer learning is a method of transferring learning on large amounts of high-quality data to create highly accurate models on small datasets. Although models with better predictive accuracy are available, his VGG19 was used in the present study because of the high accuracy achieved by this relatively simple model in previous studies on postresuscitation head CT. rice field.⁸.

Image data acquired at a size of 256 × 256 are center-cropped, resized to 224 × 224, and then normalized. Due to the small number of data, transformations were used to fit the data for image augmentation. Image augmentation is the creation of new training samples from existing images by slightly modifying the original images. In this study, we induced random variations in the training data by adjusting sharpness, rotation, and erasure. Because the dataset consisted of class-biased imbalanced data, adjustments were made using weights. Grid search was used for hyperparameter tuning. To avoid overfitting the model on the current dataset, the number of epochs was determined by ‘earlystopping’, which stops training when the accuracy of the validation data decreases.

The areas of focus were further explored using the Grad-CAM method to generate a ‘visual explanation’ of the model’s class decisions.¹³The Grad-CAM technique uses gradients that flow into a final convolutional layer to generate coarse localization maps that highlight important regions in the image and predict classes.

Measurement of GWR based on CT scan

GWR was measured for all patients in the study using the previously described method^5,6,7,23Briefly, head CT scans were retrospectively reviewed twice by emergency physicians blinded to patient outcome. We measured the average HU of circular ROIs (10.0–15 mm²) on both sides of the basal ganglia, centroid semiovale, and high cortical level. The caudate nucleus (CN), putamen (PU), posterior internal capsule (PLIC), and corpus callosum (CC) were measured at the basal ganglia level, and the medial cortex (MC) and medial white matter (MW) were measured. Measured at the centrohemiovoid level (MC1 and MW1) and the high cortical level (MC2 and MW2), respectively. The relationship between two measures was assessed using Spearman’s correlation coefficient and the mean of both measures was used for subsequent assessment. GWR was calculated according to the previously reported formula as: GWR-BG = (CN + PU)/(PLIC + CC), GWR-CE = (MC1 + MC2)/(MW1 + MW2) , and GWR-AV = ( GWR-BG + GWR-CE)/2. In this study, we used a GWR cutoff value of 1.2, as stated in our previous study.^Five.

statistical analysis

Continuous variables are expressed as median (interquartile range) and categorical variables are expressed as number of patients (percentage).Mann Whitney cormorant and Fisher’s exact test were used to compare continuous and categorical variables, respectively.Statistical significance is P.< 0.05. The neurological outcome prediction performance of the method was evaluated by plotting ROC curves and comparing AUCs. Since the dataset in this study consisted of imbalanced data with approximately 8:2 ratio labels, PR curves were drawn and AUCs were compared. ROC curves are frequently used to compare model performance because they are not affected by class proportions in the data. However, our study cohort was considered to have class imbalance (minority CPC 1/2), so class bias had to be considered. A PR curve is a good representation of method performance when the proportions of classes in the test data are close to the proportions expected when the model is actually applied.^{twenty four}All analyzes were performed using Python 3.8.5 (Python Software Foundation, Beaverton, OR).