Inter-rater agreement between humans and computer in quantitative assessment of computed tomography after cardiac arrest

Analysis of Cerebral CT Based on Supervised Machine Learning as a Predictor of Outcome After Out-of-Hospital Cardiac Arrest

BACKGROUND AND OBJECTIVES

In light of limited intensive care capacities and a lack of accurate prognostic tools to advise caregivers and family members responsibly, this study aims to determine whether automated cerebral CT (CCT) analysis allows prognostication after out-of-hospital cardiac arrest.

METHODS

In this monocentric, retrospective cohort study, a supervised machine learning classifier based on an elastic net regularized logistic regression model for gray matter alterations on nonenhanced CCT obtained after cardiac arrest was trained using 10-fold cross-validation and tested on a hold-out sample (random split 75%/25%) for outcome prediction. Following the literature, a favorable outcome was defined as a cerebral performance category of 1-2 and a poor outcome of 3-5. The diagnostic accuracy was compared with established and guideline-recommended prognostic measures within the sample, that is, gray matter-white matter ratio (GWR), neuron-specific enolase (NSE), and neurofilament light chain (NfL) in serum.

RESULTS

Of 279 adult patients, 132 who underwent CCT within 14 days of cardiac arrest with good imaging quality were identified. Our approach discriminated between favorable and poor outcomes with an area under the curve (AUC) of 0.73 (95% CI 0.59-0.82). Thus, the prognostic power outperformed the GWR (AUC 0.66, 95% CI 0.56-0.76). The biomarkers NfL, measured at days 1 and 2, and NSE, measured at day 2, exceeded the reliability of the imaging markers derived from CT (AUC NfL day 1: 0.87, 95% CI 0.75-0.99; AUC NfL day 2: 0.90, 95% CI 0.79-1.00; AUC NSE day: 2 0.78, 95% CI 0.62-0.94).

DISCUSSION

Our data show that machine learning-assisted gray matter analysis of CCT images offers prognostic information after out-of-hospital cardiac arrest. Thus, CCT gray matter analysis could become a reliable and time-independent addition to the standard workup with serum biomarkers sampled at predefined time points. Prospective studies are warranted to replicate these findings.

Regional Brain Net Water Uptake in Computed Tomography after Cardiac Arrest - A Novel Biomarker for Neuroprognostication

BACKGROUND

Selective water uptake by neurons and glial cells and subsequent brain tissue oedema are key pathophysiological processes of hypoxic-ischemic encephalopathy (HIE) after cardiac arrest (CA). Although brain computed tomography (CT) is widely used to assess the severity of HIE, changes of brain radiodensity over time have not been investigated. These could be used to quantify regional brain net water uptake (NWU), a potential prognostic biomarker.

METHODS

We conducted an observational prognostic accuracy study including a derivation (single center cardiac arrest registry) and a validation (international multicenter TTM2 trial) cohort. Early (<6 h) and follow-up (>24 h) head CTs of CA patients were used to determine regional NWU for grey and white matter regions after co-registration with a brain atlas. Neurological outcome was dichotomized as good versus poor using the Cerebral Performance Category Scale (CPC) in the derivation cohort and Modified Rankin Scale (mRS) in the validation cohort.

RESULTS

We included 115 patients (81 derivation, 34 validation) with out-of-hospital (OHCA) and in-hospital cardiac arrest (IHCA). Regional brain water content remained unchanged in patients with good outcome. In patients with poor neurological outcome, we found considerable regional water uptake with the strongest effect in the basal ganglia. NWU >8% in the putamen and caudate nucleus predicted poor outcome with 100% specificity (95%-CI: 86-100%) and 43% (moderate) sensitivity (95%-CI: 31-56%).

CONCLUSION

This pilot study indicates that NWU derived from serial head CTs is a promising novel biomarker for outcome prediction after CA. NWU >8% in basal ganglia grey matter regions predicted poor outcome while absence of NWU indicated good outcome. NWU and follow-up CTs should be investigated in larger, prospective trials with standardized CT acquisition protocols.

Standardised and automated assessment of head computed tomography reliably predicts poor functional outcome after cardiac arrest: a prospective multicentre study

PURPOSE

Application of standardised and automated assessments of head computed tomography (CT) for neuroprognostication after out-of-hospital cardiac arrest.

METHODS

Prospective, international, multicentre, observational study within the Targeted Hypothermia versus Targeted Normothermia after out-of-hospital cardiac arrest (TTM2) trial. Routine CTs from adult unconscious patients obtained > 48 h ≤ 7 days post-arrest were assessed qualitatively and quantitatively by seven international raters blinded to clinical information using a pre-published protocol. Grey-white-matter ratio (GWR) was calculated from four (GWR-4) and eight (GWR-8) regions of interest manually placed at the basal ganglia level. Additionally, GWR was obtained using an automated atlas-based approach. Prognostic accuracies for prediction of poor functional outcome (modified Rankin Scale 4-6) for the qualitative assessment and for the pre-defined GWR cutoff < 1.10 were calculated.

RESULTS

140 unconscious patients were included; median age was 68 years (interquartile range [IQR] 59-76), 76% were male, and 75% had poor outcome. Standardised qualitative assessment and all GWR models predicted poor outcome with 100% specificity (95% confidence interval [CI] 90-100). Sensitivity in median was 37% for the standardised qualitative assessment, 39% for GWR-8, 30% for GWR-4 and 41% for automated GWR. GWR-8 was superior to GWR-4 regarding prognostic accuracies, intra- and interrater agreement. Overall prognostic accuracy for automated GWR (area under the curve [AUC] 0.84, 95% CI 0.77-0.91) did not significantly differ from manually obtained GWR.

CONCLUSION

Standardised qualitative and quantitative assessments of CT are reliable and feasible methods to predict poor functional outcome after cardiac arrest. Automated GWR has the potential to make CT quantification for neuroprognostication accessible to all centres treating cardiac arrest patients.

Outcome prediction of cardiac arrest with automatically computed gray-white matter ratio on computed tomography images

BACKGROUND

This study aimed to develop an automated method to measure the gray-white matter ratio (GWR) from brain computed tomography (CT) scans of patients with out-of-hospital cardiac arrest (OHCA) and assess its significance in predicting early-stage neurological outcomes.

METHODS

Patients with OHCA who underwent brain CT imaging within 12 h of return of spontaneous circulation were enrolled in this retrospective study. The primary outcome endpoint measure was a favorable neurological outcome, defined as cerebral performance category 1 or 2 at hospital discharge. We proposed an automated method comprising image registration, K-means segmentation, segmentation refinement, and GWR calculation to measure the GWR for each CT scan. The K-means segmentation and segmentation refinement was employed to refine the segmentations within regions of interest (ROIs), consequently enhancing GWR calculation accuracy through more precise segmentations.

RESULTS

Overall, 443 patients were divided into derivation N=265, 60% and validation N=178, 40% sets, based on age and sex. The ROI Hounsfield unit values derived from the automated method showed a strong correlation with those obtained from the manual method. Regarding outcome prediction, the automated method significantly outperformed the manual method in GWR calculation (AUC 0.79 vs. 0.70) across the entire dataset. The automated method also demonstrated superior performance across sensitivity, specificity, and positive and negative predictive values using the cutoff value determined from the derivation set. Moreover, GWR was an independent predictor of outcomes in logistic regression analysis. Incorporating the GWR with other clinical and resuscitation variables significantly enhanced the performance of prediction models compared to those without the GWR.

CONCLUSIONS

Automated measurement of the GWR from non-contrast brain CT images offers valuable insights for predicting neurological outcomes during the early post-cardiac arrest period.

Deep learning-enabled detection of hypoxic-ischemic encephalopathy after cardiac arrest in CT scans: a comparative study of 2D and 3D approaches

Objective

To establish a deep learning model for the detection of hypoxic-ischemic encephalopathy (HIE) features on CT scans and to compare various networks to determine the best input data format.

Methods

168 head CT scans of patients after cardiac arrest were retrospectively identified and classified into two categories: 88 (52.4%) with radiological evidence of severe HIE and 80 (47.6%) without signs of HIE. These images were randomly divided into a training and a test set, and five deep learning models based on based on Densely Connected Convolutional Networks (DenseNet121) were trained and validated using different image input formats (2D and 3D images).

Results

All optimized stacked 2D and 3D networks could detect signs of HIE. The networks based on the data as 2D image data stacks provided the best results (S100: AUC: 94%, ACC: 79%, S50: AUC: 93%, ACC: 79%). We provide visual explainability data for the decision making of our AI model using Gradient-weighted Class Activation Mapping.

Conclusion

Our proof-of-concept deep learning model can accurately identify signs of HIE on CT images. Comparing different 2D- and 3D-based approaches, most promising results were achieved by 2D image stack models. After further clinical validation, a deep learning model of HIE detection based on CT images could be implemented in clinical routine and thus aid clinicians in characterizing imaging data and predicting outcome.

Brain-derived extracellular vesicles as serologic markers of brain injury following cardiac arrest: A pilot feasibility study

AIM

Assessment of neurologic injury within the immediate hours following out-of-hospital cardiac arrest (OHCA) resuscitation remains a major clinical challenge. Extracellular vesicles (EVs), small bodies derived from cytosolic contents during injury, may provide the opportunity for "liquid biopsy" within hours following resuscitation, as they contain proteins and RNA linked to cell type of origin. We evaluated whether micro-RNA (miRNA) from serologic EVs were associated with post-arrest neurologic outcome.

METHODS

We obtained serial blood samples in an OHCA cohort. Using novel microfluidic techniques to isolate EVs based on EV surface marker GluR2 (present on excitatory neuronal dendrites enriched in hippocampal tissue), we employed reverse transcription quantitative polymerase chain reaction (RT-qPCR) methods to measure a panel of miRNAs and tested association with dichotomized modified Rankin Score (mRS) at discharge.

RESULTS

EVs were assessed in 27 post-arrest patients between 7/3/2019 and 7/21/2022; 9 patients experienced good outcomes. Several miRNA species including miR-124 were statistically associated with mRS at discharge when measured within 6 hours of resuscitation (AUC = 0.84 for miR-124, p < 0.05). In a Kendall ranked correlation analysis, miRNA associations with outcome were not strongly correlated with standard serologic marker measurements, or amongst themselves, suggesting that miRNA provide distinct information from common protein biomarkers.

CONCLUSIONS

This study explores the associations between miRNAs from neuron-derived EVs (NDEs) and circulating protein biomarkers within 6 hours with neurologic outcome, suggesting a panel of very early biomarker may be useful during clinical care. Future work will be required to test larger cohorts with a broader panel of miRNA species.

Use of Prognostication Instruments in Prognostication Procedures of Postanoxic Coma Patients over Time: A Retrospective Study

BACKGROUND

Many survivors of cardiovascular arrest remain in a postanoxic coma. The neurologist's task is to provide the most accurate assessment of the patient's neurologic prognosis through a multimodal approach of clinical and technical tests. The aim of this study is to analyze differences and developments in the concept of neurological prognosis assessment and in-hospital outcome of patients over a five year-period.

METHODS

This retrospective observational study included 227 patients with postanoxic coma treated in the medical intensive care unit of the University Hospital, Mannheim from January 2016 to May 2021. We retrospectively analyzed patient characteristics, post-cardiac arrest care, and the use of clinical and technical tests for neurological prognosis assessment and patient outcome.

RESULTS

Over the observation period, 215 patients received a completed neurological prognosis assessment. Regarding the multimodal prognostic assessment, patients with poor prognosis (54%) received significantly fewer diagnostic modalities than patients with very likely poor (20.5%), indeterminate (24.2%), or good prognosis (1.4%; p = 0.001). The update of the DGN guidelines in 2017 had no effect on the number of performed prognostic parameters per patient. The finding of bilaterally absent pupillary light reflexes or severe anoxic injury on CT contributed most to a poor prognosis category (OR 8.38, 95%CI 4.01-7.51 and 12.93, 95%CI 5.55-30.13, respectively), whereas a malignant EEG pattern and NSE > 90 µg/L at 72 h resulted in the lowest OR (5.11, 95%CI 2.32-11.25, and 5.89, 95%CI 3.14-11.06, respectively) for a poor prognosis category. Assessment of baseline NSE significantly increased over the years (OR 1.76, 95%CI 1.4-2.22, p < 0.001), and assessment of follow-up NSE at 72 h trended to increase (OR 1.19, 95%CI 0.99-1.43, p = 0.06). In-hospital mortality was high (82.8%), remained unchanged over the observation period, and corresponded to the number of patients in whom life-sustaining measures were discontinued.

CONCLUSIONS

Among comatose survivors of cardiac arrest, the prognosis remains poor. Prognostication of a poor outcome led nearly exclusively to withdrawal of care. Prognostic modalities varied considerably with regard to their contribution to a poor prognosis category. Increasing enforcement of a standardized prognosis assessment and standardized evaluation of diagnostic modalities are needed to avoid false-positive prognostication of poor outcomes.