Study population and setting

The data were collected retrospectively from the Integrated Medical Database of National Taiwan University Hospital (NTUH-iMD) from January 2009 to December 2019. National Taiwan University Hospital—a tertiary medical center—typically saw approximately 100,000 emergency department visits per year. Eligible patients included the following: (1) adult patients who experienced non-traumatic OHCA, (2) treated in the emergency department, and (3) successfully resuscitated with ROSC. Overall, 544 patients underwent brain CT imaging within 12 h after ROSC and were enrolled in the study. The Institutional Review Board of National Taiwan University Hospital approved this study, along with a waiver of informed consent from the patients’ relatives or physicians, on October 6, 2020 (IRB No. 202004037RINA, Study title: Prognosis and treatment evaluation of post-cardiac arrest patients—a multimodal, autonomic, neuroprognostic model). All procedures performed in this study adhered to the ethical standards set by the responsible committee on human experimentation (institutional or regional) and the Declaration of Helsinki in 1975.

Patient data acquisition and outcome measurement

All medical history and details of cardiac arrest events were coded in accordance with the Utstein style and extracted from the electronic ambulance and medical records. This included patient characteristics, interventions provided, and outcomes. To predict the outcome in the early post-cardiac arrest period, variables such as age, sex, pre-existing comorbidities, initial rhythm, resuscitation events, non-contrast brain CT findings, hemodynamic parameters, and laboratory results following ROSC were included for further analysis. Brain CT adhered to the post-cardiac arrest care protocol of the medical center during the study period. Briefly, non-contrast brain CT scans were performed when vital signs were relatively stable after ROSC. Patients were sent to the CT examination room after providing informed consent for CT scan studies. The average time to undergo CT after resuscitation was \(103\pm 77\) min (median [interquartile range, IQR] = \(88~[64-115]\)).

The outcome endpoints included favorable neurological outcomes at hospital discharge, which were defined as cerebral performance category (CPC) scores of 1 or 2 and survival to hospital discharge. The CPC score is a validated 5-point scale indicating neurological disability (CPC 1: good cerebral performance; CPC 2: moderate cerebral disability; CPC 3: severe cerebral disability; CPC 4: coma/vegetative state; and CPC 5: brain death). Patients with a CPC score of 1 or 2 generally exhibited adequate cerebral function to live independently. Withdrawal of life sustaining therapy before hospital discharge adhered to the protocol of the medical center. Briefly, at least two attending physicians evaluated the neurological status by examining pupillary light reflexes, spontaneous respiratory drives, and 24-lead electroencephalograms within 7 days after cardiac arrest. Initial brain CT images were not routinely used but could be referenced by physicians during the evaluation process. Decisions regarding the withdrawal of life sustaining therapy were made after discussing with the family of the patients with cardiac arrest in a comatose state.

Brain CT image analysis and GWR acquisitionGWR calculation formulas

This study focused on two GWR calculation formulas: GWR at the basal ganglia level (GWR_b) and a simplified version, GWR_s. GWR_b was determined by summing the HU values of the caudate nuclei (CN) and putamen (PU), then dividing by the sum of HU values of the posterior limb of the internal capsule (PIC) and corpus callosum (CC) in Eq. 1. The simplified GWR (GWR_s) was calculated as the HU value of the PU divided by the HU value of the PIC, as shown in Eq. 2. The simplified GWR was proposed and compared in the study owing to the well-localized characteristics of these two areas in the brain CT images.

$$\begin }\_}&= \frac}+}}}+}} \end$$

(1)

$$\begin }\_}&= \frac}}}} \end$$

(2)

Manual method

The GWR was manually measured to compare its efficacy in predicting outcomes with that of the automatically computed GWR. Following the methodologies of previous studies [22,23,24], eight regions of interest (ROIs) at the basal ganglia level, including the CC, CN, PU, and PIC, were selected for annotation, (Fig. 1a). Two emergency or critical care physicians who were blinded to the outcomes annotated each circular region. The process involved one doctor providing their assessment initially, and the other doctor subsequently reviewing and verifying the findings. If disagreement persists, a third physician, also blinded to the outcomes, would make the final decision. The area for each circular region was approximately 10 \(\mathrm\). The HU for each circular region was calculated by averaging all pixels within the region. The physicians were blinded to the clinical data, survival, and neurological outcomes of the patients before and during the manual annotation of brain CT images.

Fig. 1

Visualization of a manual segmentations, b automated segmentations, and c automated segmentations in 3D view. CC, corpus callosum; CN, caudate nuclei; PU, putamen; PIC, posterior limb of the internal capsule

Automated method

An automated method was developed to calculate the GWR of head CT scans. Our method comprised four steps: image registration, K-means segmentation, segmentation refinement, and GWR calculation (Fig. 2). For image registration, a nonlinear algorithm from ANTsPy (Version 0.2.9) aligned the CT scan (\(I_\)) with the Eve template (\(I'_\)) [25, 26]. The objective was to overlay \(M'_\) onto the registered CT (\(I'_\)) and subsequently derive all ROI segmentations. However, registration accuracy could be influenced by various factors, potentially affecting the precision of ROI segmentations. Subsequently, K-means segmentation and ROI segmentation refinement were implemented to adjust the \(M'_\). First, we extracted the brain from \(I'_\) and then utilized K-means clustering to obtain the gray (\(M'_\)) and the white matter mask (\(M'_\)). Following that, techniques including filtering, closing, and opening were employed to refine \(M'_\) and \(M'_\), resulting in the creation of the refined Eve ROIs mask (\(M'_\)). Finally, the eight ROIs (four on each side) were inversely transformed to \(M_\), and GWR_b and GWR_s were computed according to Eqs. 1 and 2, respectively. Additional file 1 shows further details about the automated method. Compared to the manual method, the automated approach generated a 3D volume for each ROI (Fig. 1b, c). Conversely, the manual method produced a 2D circular region for each ROI (Fig. 1a). Furthermore, Fig. 1b, c delineates the CC into its genu, body, and splenium parts; however, for GWR calculation, all three parts were combined as the entire CC. The automated method offered an advantage in that GWR was calculated using a fixed algorithm, reducing subjectivity among various physicians and ensuring reproducibility. Additionally, the GWR derived from 3D ROI segmentations includes a greater number of voxels, thus providing a more comprehensive characterization of the ROI.

Fig. 2

Workflow of the automated method

Excluded CTs management

The symmetry of brain hemispheres could be influenced by head malpositioning during CT examination or pre-existing localized brain lesions. However, symmetry loss in the automated GWR calculation indicated a lack of symmetry on the brain CT images, potentially leading to GWR miscalculation. If physicians identified a loss of symmetry during the initial evaluation, the case was excluded and defined as a manually excluded CT. The remaining cases were included in the automated calculation. The reasons for manual exclusion included loss of symmetry (N = 2), structural change (including intracranial hemorrhage or brain tumor, N = 11), severe brain atrophy resulting in symmetry loss (N = 4), and marked signal interference or incomplete imaging (N = 5). Figure 3 shows representative examples. After manual exclusion, CTs were further excluded during the automated process by identifying low registration accuracy and any missing ROI segmentations. Subsequently, the automated method was employed for the remaining cases following manual exclusion. However, improper head orientation may cause poor registration with the Eve template, resulting in incorrect segmentation and GWR evaluation (N = 12). Moreover, due to the proximity of the CC and CN to the ventricle, the automated segmentation method may not capture them accurately. In such cases, the CC or CN might be missing, rendering the evaluation of the GWR impossible (N = 5).

Fig. 3

Representative examples of manually (upper row) and automatically (lower row) excluded CTs. In the upper row, manual exclusion criteria were applied, including a loss of symmetry, b structural change (e.g., intracranial hemorrhage), c severe signal interference, and d chronic large brain lesion with atrophy. In the lower row, automated exclusion of CTs resulted from inaccurate registration, leading to e loss of CC segmentation, f loss of CN segmentation, g loss of both CC and CN segmentations, and h incorrect segmentation. The regions enclosed by dotted lines indicate missing ROI segmentations. CC, corpus callosum; CN, caudate nucleus

Statistical analysis

Continuous variables, presented as the mean with standard deviation, were compared using Student’s t test, while categorical variables expressed as median with the interquartile range were compared employing the Chi-square test. Pearson’s correlation coefficient (PCC) was utilized to evaluate the correlation between manual and automated calculations of intensities across different brain gray and white matter regions. Receiver operating characteristic (ROC) curves were constructed for manual and automated GWR. The performance of different GWR calculation methods was compared by evaluating the difference in the area under the curve (AUCs) of the ROC curves for each method. Variables with a significance level of \(p<0.1\) were included in the multiple logistic regression model. Multiple logistic regression analyses were performed to identify independent variables associated with predicting favorable neurological outcomes. A \(p<0.05\) was considered significant. Statistical analyses were performed using R software 4.2.0.