Time-specific comparison of gray-white matter ratio models for predicting outcomes after cardiac arrest: basal ganglia as the leading model

Abstract

BACKGROUND To identify optimal time-specific gray-white matter ratio (GWR) models for predicting poor neurological outcomes after cardiac arrest (CA).

METHODS We retrospectively analyzed a prospective single-center cohort of comatose adults who survived CA. Brain computed tomography (CT) scans were performed at three predefined phases following the return of spontaneous circulation (ROSC): the early phase (<12 h), intermediate phase (12-72 h), and recovery phase (>72 h to 1 week). Patients with ≥1 scan were included. Five GWR models (basal ganglia [BG], cortical [CO], average [AV], simplified [SI], and cerebellar [CE]) and a combined model (BG+CE) were evaluated for their ability to predict poor neurological outcomes (cerebral performance category of 3-5) at discharge.

RESULTS In all, 173 patients were included, 40 (23.1%) of whom achieved favorable neurological outcomes at discharge. The five GWR values were consistently lower in the poor-outcome group across multiple phases. In the early phase, the area under the curves (AUCs) for GWR-BG, GWR-AV, GWR-CE, GWR-SI, and GWR-CO were 0.83 (95% CI 0.75-0.91), 0.78 (95% CI 0.68-0.87), 0.73 (95% CI 0.61-0.84), 0.72 (95% CI 0.62-0.82), and 0.64 (95% CI 0.51-0.74), respectively. In the intermediate phase, the AUCs increased to 0.87 (95% CI 0.77-0.95), 0.88 (95% CI 0.78-0.96), 0.84 (95% CI 0.73-0.93), 0.81 (95% CI 0.70-0.92), and 0.79 (95% CI 0.66-0.88), respectively, while in the recovery phase, they decreased to 0.76 (95% CI 0.63-0.88), 0.75 (95% CI 0.61-0.86), 0.74 (95% CI 0.60-0.87), 0.64 (95% CI 0.49-0.77), and 0.67 (95% CI 0.52-0.80), respectively. GWR-BG remained comparable to the other individual models. The combined model achieved AUCs of 0.87, 0.90, and 0.80 across the three phases. In the early phase, the combined model significantly outperformed the other four individual models and exhibited numerically higher performance than GWR-BG.

CONCLUSION GWR-BG was superior or comparable to the other GWR models across all three post-ROSC phases; GWR-CE was validated as a predictor of poor neurological outcomes, and the combination of GWR-BG and GWR-CE further enhanced early-phase prediction.

INTRODUCTION

Cardiac arrest (CA) remains a major global health challenge, as most survivors remain comatose after the return of spontaneous circulation (ROSC).^[1,2] Accurate prediction is essential to avoid inappropriate withdrawal of life-sustaining treatment (WLST) in potentially recoverable patients.^[3] Current guidelines recommend a multimodal neuroprognostication approach,^[4,5] in which the gray-white matter ratio (GWR) on noncontrast-enhanced brain computed tomography (CT) is a widely used imaging marker of cerebral edema. A lower GWR indicates more severe edema and is strongly associated with poor neurological outcomes.^[6,7]

However, substantial methodological heterogeneity exists in GWR measurements across studies. GWR is typically measured in the basal ganglia (GWR-BG) because of its vulnerability to hypoxic-ischemic injury.^[8,9] Alternatively, the GWR can be measured in the cerebrum to reflect diffuse cortical damage (GWR-CO)^[10,11] or as an average of both (GWR-AV).^[7,12] A simplified putamen-to-posterior-limb internal capsule ratio (GWR-SI) has also been validated.^[13,14] Moreover, the predictive value of GWR in the cerebellum (GWR-CE) remains largely unevaluated in CA survivors, despite its good predictive value for brain death.^[15-17] These methods have never been directly compared; therefore, the optimal GWR model has not been defined. Furthermore, it is currently unclear how the predictive performance of these models evolves throughout the entire dynamic edema process.

Accordingly, this study aimed to compare five GWR methods across three post-ROSC phases within the first week after ROSC. We hypothesized that GWR-BG would be superior or comparable to the other methods across all phases and that the combination of GWR-CE with GWR-BG would further improve the predictive performance. Through the identification of optimal time-specific GWR models, we sought to address methodological heterogeneity and provide clear guidance for clinical neuroprognostication.

METHODS

Study design

This study was a retrospective analysis of a prospective, single-center post-cardiac arrest syndrome (PCAS) cohort at Peking University Third Hospital, initiated in January 2014. In the overall cohort, brain CT scans were scheduled for three predefined phases on the basis of the 2008 ILCOR/AHA consensus:^[18] early phase (<12 h post-ROSC, scanned as soon as possible), intermediate phase (12-72 h, but preferably at 48-72 h), and recovery phase (>72 h to 1 week; preferably around 1 week). We included comatose adults who survived CA (aged ≥18 years), who were admitted between January 2014 and January 2025 and who underwent ≥1 brain CT scan within the first week after ROSC. The exclusion criteria were as follows: (1) pre-arrest cerebral performance category (CPC) 3-4; (2) acute intracranial pathology as the cause of CA; (3) poor-quality scans that would have prevented reliable GWR measurement; and (4) missing key CT slices or regions of interest (ROIs) required for GWR analysis.

Target temperature management (TTM) was initiated within 60 min of admission using an intravascular cooling system, with target temperatures adapted to contemporary guidelines. Decisions regarding the WLST were made by a multidisciplinary team according to a delayed multimodal assessment. The detailed WLST protocol is provided in the supplementary material. The CPC score was administered at hospital discharge. CPC scores of 1-2 were used to define “good” outcomes, while CPC scores of 3-5 were used to define “poor” outcomes.

GWR measurement

Non-contrast brain CT scans were acquired using a SOMATOM Definition Flash scanner (Siemens Healthcare, Germany) according to standard clinical protocols. In accordance with established methodologies,^[15,19,20] circular ROIs (10-15 mm²) were placed independently on axial CT images at the five anatomical levels by two experienced radiologists who were blinded to the outcomes. Hounsfield unit (HU) measurements were conducted bilaterally at each level (Figure 1). (1) Basal ganglia level included corpus callosum (CC), caudate nucleus (CN), putamen (PU), posterior limb of the internal capsule (PLIC) and thalamus (TH). (2) Centrum semiovale and high convexity levels included medial cortex (MC1, MC2) and medial white matter (MW1, MW2). (3) Cerebellar level included cerebellar gray matter (CBC) and cerebellar white matter (CBW). (4) Pons level included brainstem central region (P) and bilateral anterior brainstem regions (LP, RP); mean of the three regions used.

Figure 1. Locations of regions of interest (ROIs) at the selected brain regions. Diagram showing the five standard axial brain CT levels where circular ROIs (10-15 mm²) were placed. A: basal ganglia level; B: centrum semiovale level; C: high convexity level; D: cerebellar level; E: pons level.

The following GWR values were calculated:

GWR-BG (basal ganglia) = (CN + PU)/(PLIC + CC)

GWR-SI (simplified) = PU/PLIC

GWR-CO (cortical) = (MC1 + MC2)/(MW1 + MW2)

GWR-AV (average) = (GWR-BG + GWR-CO)/2

GWR-CE (cerebellar) = CBC/CBW

Statistical analysis

Statistical analyses were conducted using SPSS 27.0 (IBM, USA) and MedCalc 19.1.7 (Belgium). Continuous variables are expressed as means±standard deviation (SD) or medians (interquartile range, IQR) and were compared using the independent t test or the Mann-Whitney U test with the Shapiro-Wilk test. Categorical data are expressed as numbers (percentages) and were compared using the Chi-square test or Fisher’s exact test. Inter-rater reliability was assessed using two-way mixed-effects intra-class correlation coefficients (ICCs). Predictive performance was assessed by receiver operating characteristic (ROC) curve analysis; DeLong’s test was used to compare the area under the curve (AUC) with P-values adjusted via the Holm-Bonferroni method. Bootstrap internal validation (1,000 resamples, percentile method) was used to calculate 95% confidence intervals (CIs) for the AUC estimates. Cut-off values were determined according to Youden’s index, and sensitivity was assessed at 100% specificity. GWR-BG and GWR-CE were combined using logistic regression to construct a composite model for ROC analysis. Missing data on time to ROSC (n=21, 12.1%) were managed via complete-case analysis. Two-tailed P-values <0.05 indicated statistical significance.

RESULTS

Patient demographics

Among the 533 screened patients, 309 were excluded because they had no ROSC (n=213) or had received no brain CT within 1 week (n=96). Fifty-one patients were further excluded because they had neurogenic CA (n=9), missing CT levels/ROIs (n=30), or their scans were of poor image quality (n=12), and thus 173 patients were included in the final analysis (Figure 2). In all, 78 (45.1%) patients survived to hospital discharge, and 40 (23.1%) had good neurological outcomes. In-hospital mortality was 54.9% (95/173), with only 7 deaths (7.4% of non-survivors) following a formal WLST decision and a median time from ROSC to death of 111.1 hours (IQR 47.8-495.8). Among survivors, the median hospitalization duration was 18.0 d (IQR 11.8-26.0). Patients with good outcomes were younger and had a lower prevalence of hypertension, higher TTM utilization, and more frequent cardiac etiology and shockable initial rhythm (all P<0.05; Table 1).

Figure 2. Flowchart of the study. ROSC: return of spontaneous circulation; CT: computed tomography.

Table 1.

Baseline characteristics

Baseline characteristics	Good outcome (n=40)	Poor outcome (n=133)	χ2/Z-value	P-value
Sex, male, n (%)	24 (60.0)	91 (68.4)	0.979	0.323
Age, years, median [IQR]	58.0 [36.0-68.0]	72.0 [55.5-81.0]	2.917	0.004*
Comorbidities, n (%)
Hypertension	14 (35.0)	77 (57.9)	6.465	0.011*
Diabetes	9 (22.5)	42 (31.6)	1.219	0.270
Myocardial Infarction	4 (10.0)	14 (10.5)		1.000
PCI	2 (5.0)	9 (6.8)		1.000
CABG	0 (0)	2 (1.5)		1.000
Heart Failure	5 (12.5)	9 (6.8)		0.318
Malignant tumor	2 (5.0)	6 (4.5)		1.000
Etiology of cardiac arrest, n (%)
Cardiac	35 (87.5)	86 (64.7)	7.777	0.020*
Respiratory	4 (10.0)	32 (24.1)
Other non-cardiac	1 (2.5)	15 (11.3)
Out-of-hospital CA, n (%)	19 (47.5)	62 (46.6)	0.141	0.707
Witness in OHCA	19 (100.0)	61 (98.4)		1.000
Bystander CPR in OHCA	14 (73.7)	40 (64.5)	0.550	0.458
Time from CA to ROSC, min, median [IQR] a	16.0 [6.0-34.0]	16.0 [10.0-37.5]	0.852	0.394
Shockable rhythm, n (%)	22 (55.0)	26 (19.5)	19.278	<0.001*
TTM, n (%)	16 (40.0)	17 (12.8)	14.758	<0.001*
ECPR, n (%)	5 (12.5)	14 (10.5)		0.774

^a: time from CA to ROSC data were available for 35 patients with good outcomes and 117 with poor outcomes; 21 missing cases were excluded (complete-case analysis). Categorical variables were compared using the Chi-square or Fisher’s exact test (expected counts <5; no χ² reported). *P<0.05. IQR: interquartile range; CA: cardiac arrest; OHCA: out-of-hospital cardiac arrest; CPR: cardiopulmonary resuscitation; ROSC: return of spontaneous circulation; PCI: percutaneous coronary intervention; CABG: coronary artery bypass grafting; TTM: targeted temperature management; ECPR: extracorporeal cardiopulmonary resuscitation.

Brain attenuation, inter-rater reliability, and GWR values

Gray matter attenuation in the basal ganglia (CN, PU) and high-convexity medial cortex (MC2) was consistently lower in the poor-outcome group across all phases (P<0.05), as was cerebellar gray matter (CBC) attenuation in the intermediate phase (P=0.035) (supplementary Table 1). The ICCs for the average measures were 0.809 for GWR-BG, 0.712 for GWR-SI, 0.696 for GWR-AV, 0.687 for GWR-CE, and 0.574 for GWR-CO. All five GWR models showed significantly lower values in patients with poor outcomes across all phases (Table 2).

Table 2.

GWR comparison between good and poor outcome groups across three phases within the first week post-resuscitation

GWR model	Good outcome	Poor outcome	P-value
Early phase	(n=25)	(n=101)
GWR-BG	1.25 (1.23-1.28)	1.19 (1.14-1.23)	<0.001*
GWR-AV	1.22 (1.20-1.25)	1.19 (1.14-1.21)	<0.001*
GWR-CE	1.36 ± 0.08	1.29 ± 0.08	<0.001*
GWR-SI	1.28 (1.24-1.30)	1.23 (1.16-1.27)	<0.001*
GWR-CO	1.19 (1.16-1.23)	1.17 (1.13-1.20)	0.034*
Intermediate phase	(n=25)	(n=52)
GWR-BG	1.25 (1.23-1.28)	1.16 (1.07-1.20)	<0.001*
GWR-AV	1.23 (1.21-1.25)	1.15 (1.11-1.19)	<0.001*
GWR-CE	1.36 ± 0.07	1.27 ± 0.07	<0.001*
GWR-SI	1.26 (1.23-1.29)	1.19 (1.09-1.23)	<0.001*
GWR-CO	1.20 (1.19-1.23)	1.16 (1.11-1.19)	<0.001*
Recovery phase	(n=24)	(n=36)
GWR-BG	1.25 (1.20-1.27)	1.21 (1.12-1.24)	<0.001*
GWR-AV	1.22 (1.19-1.23)	1.18 (1.08-1.21)	<0.001*
GWR-CE	1.38 ± 0.07	1.31 ± 0.06	<0.001*
GWR-SI	1.27 (1.22-1.29)	1.23 (1.18-1.29)	0.043*
GWR-CO	1.18 (1.16-1.20)	1.16 (1.09-1.18)	0.027*

Data are presented as medians (interquartile ranges) or means ± standard deviations. *P<0.05. GWR: gray-white matter ratio; GWR-BG: basal ganglia GWR; GWR-AV: average GWR; GWR-CE: cerebellar GWR; GWR-SI: simplified GWR; GWR-CO: cortical GWR.

Time-dependent predictive performance of the five GWR models for poor neurological outcomes

Table 3 summarizes the time-dependent performance of the five GWR models. In the early phase, the AUCs of GWR-BG, GWR-AV, GWR-CE, GWR-SI, and GWR-CO were 0.83 (95% CI 0.75-0.91), 0.78 (95% CI 0.68-0.87), 0.73 (95% CI 0.61-0.84), 0.72 (95% CI 0.62-0.82), and 0.64 (95% CI 0.51-0.74), respectively. DeLong’s tests revealed that the predictive performance of GWR-BG was comparable to that of GWR-AV and GWR-CE (Figure 3) but significantly better than that of GWR-SI (P=0.005) and GWR-CO (P=0.001). In the intermediate phase, the AUCs of GWR-BG, GWR-AV, GWR-CE, GWR-SI, and GWR-CO increased to 0.87, 0.88, 0.84, 0.81, and 0.79, respectively. In the recovery phase, the AUCs of GWR-BG, GWR-AV, GWR-CE, GWR-SI, and GWR-CO decreased to 0.76, 0.75, 0.74, 0.64, and 0.67, respectively. In the intermediate phase and recovery phase, GWR-BG remained statistically comparable to the alternative models (all P>0.05; Figure 3). The Youden index-derived cut-offs, sensitivity, and specificity, in addition to the cut-offs and sensitivity at 100% specificity, are detailed in supplementary Table 2.

Table 3.

Time-dependent predictive performance of five GWR models and the combined model for poor neurological outcome

GWR models	Early phase AUC (95% CI)	Intermediate phase AUC (95% CI)	Recovery phase AUC (95% CI)
GWR-BG	0.83 (0.75-0.91)	0.87 (0.77-0.95)	0.76 (0.63-0.88)
GWR-AV	0.78 (0.68-0.87)	0.88 (0.78-0.96)	0.75 (0.61-0.86)
GWR-CE	0.73 (0.61-0.84)	0.84 (0.73-0.93)	0.74 (0.60-0.87)
GWR-SI	0.72 (0.62-0.82)	0.81 (0.70-0.92)	0.64 (0.49-0.77)
GWR-CO	0.64 (0.51-0.74)	0.79 (0.66-0.88)	0.67 (0.52-0.80)
GWR-BG & GWR-CE	0.87 (0.78-0.94)	0.90 (0.79-0.97)	0.80 (0.67-0.91)

95% CIs were calculated using Bootstrap internal validation with 1,000 resamples. AUC: area under the curve; CI: confidence interval; GWR: gray-white matter ratio; GWR-BG: basal ganglia GWR; GWR-AV: average GWR; GWR-CE: cerebellar GWR; GWR-SI: simplified GWR; GWR-CO: cortical GWR.

Figure 3. Comparison of GWR models across three phases by DeLong’s test. Comparison of prognostic performance by heatmaps during the early (A), intermediate (B), and recovery (C) phases. Each cell shows the exact P-value; color intensity corresponds to −log₁₀(P-value). Asterisks (*) denote comparisons significant after Holm-Bonferroni correction. GWR: gray-white matter ratio; GWR-BG: basal ganglia GWR; GWR-AV: average GWR; GWR-CE: cerebellar GWR; GWR-SI: simplified GWR; GWR-CO: cortical GWR.

Performance of the combined GWR-BG and GWR-CE model

The combined model had AUCs of 0.87, 0.90, and 0.80 for the early, intermediate, and recovery phases, respectively. In the early phase, the combined model significantly outperformed the four individual models (all P<0.05) and numerically exceeded the GWR-BG (P=0.14). In the latter two phases, the AUCs were numerically greater than those of all five individual models (Table 3, Figure 3).

Subgroup analysis by arrest location

Subgroup analyses revealed that the AUCs of all the GWR models were highly consistent across OHCA patients and IHCA patients in all three phases, with no significant differences or interaction effects (all DeLong’s P>0.05 and interaction P>0.05; supplementary Tables 3 and 4).

DISCUSSION

Our findings indicated that GWR-BG was superior or comparable to the other GWR methods across all three post-ROSC phases. This study demonstrated GWR-CE as a potential predictor of poor neurological outcomes, and the combination of GWR-BG and GWR-CE further improved the predictive performance in the early phase.

Although a reduced GWR strongly predicts poor neurological outcomes, its measurement methods vary widely across studies. GWR-AV, the average of GWR-BG and GWR-CO, was proposed as a comprehensive measure and was considered by several early studies to achieve good predictive performance among GWR methods.^[19,20] Moreover, our data revealed that the predictive efficacy of GWR-BG was comparable to that of GWR-AV across all three phases within the first week after resuscitation, which suggests that the prognostic value of GWR-AV was largely driven by GWR-BG. This is further supported by both our results and those of previous reports that demonstrated the limited predictive performance of GWR-CO.^[19,21] The poor performance of GWR-CO may be attributed to the thin cortical ribbon and its susceptibility to partial volume effects, which obscure the gray-white matter boundary and introduce considerable subjective variability in ROI placement.^[22] In contrast, the basal ganglia are not only sensitive to hypoxicischemic injury but are also anatomically more distinct and less prone to partial volume artifacts; therefore, GWR-BG is a simpler and more reliable prognostic tool.^[21,23,24]

The cerebellum, with its highly vulnerable Purkinje cells, is among the regions that are most susceptible to hypoxic-ischemic brain injury following CA.^[25] The loss of Purkinje cells may present as post-hypoxic myoclonus, which is a strong predictor of poor outcomes.^[4,25] This study evaluated the ability of GWR-CE to predict outcomes in comatose patients who experienced CA, while prior investigations have focused mainly on its use in predicting brain death.^[15,17] We found that GWR-CE was significantly lower in the poor-outcome group and that its performance is comparable to that of GWR-BG and GWR-AV. Notably, the combination of GWR-CE with GWR-BG significantly improved prognostic accuracy, particularly in the early phase. Postmortem studies have indicated that the cerebellum is highly susceptible to hypoxic-ischemic neuronal death, second only to the hippocampus, and that 9.1% of patients exhibit predominant cerebellar injury.^[26,27] The incorporation of cerebellar vulnerability into imaging assessments may provide complementary information to deep gray matter evaluation and may help refine early prognostication.

All the GWR models consistently achieved higher AUCs in the intermediate phase (12-72 h). This temporal pattern reflects the pathophysiology of cerebral edema after CA. In the early phase (<12 h), loss of gray-white matter differentiation may not be fully evident,^[28] which limits the prognostic sensitivity of early CT. This finding is consistent with a previous study that shows that delayed imaging provides higher accuracy.^[29] Cytotoxic and vasogenic edema typically peaks between 48 and 72 h,^[30] allowing maximal discrimination of the GWR. Beyond 72 h, however, tissue recovery or irreversible infarction may reduce gray-white matter contrast, leading to decreased predictive performance.^[31]

Limitations

Several limitations warrant consideration. First, this was a single-center study. Second, cohort differences across time windows limited the longitudinal analysis to cross-sectional comparisons. Third, outcome assessment using CPC at discharge may be suboptimal compared with the 3- or 6-month follow-up. Finally, although only 7 deaths (7.4% of non-survivors) followed formal WLST, a self-fulfilling prophecy cannot be entirely excluded.

CONCLUSION

GWR-BG was superior or comparable to the other GWR models across all three post-ROSC phases; GWR-CE was validated as a predictor of poor neurological outcomes, and combination of GWR-BG and GWR-CE further enhanced early-phase prediction.

Funding: This study was supported by grants from the National Natural Science Foundation of China (62273009) and Peking University Third Hospital (BYSYDL2025023).

Ethical approval: The study was approved by the Ethics Committee of Peking University Third Hospital (PUTH2022-160-02), and informed consent was obtained from all participants or their next of kin.

Conflicts of interest: The authors declare that they have no known competing financial interests or personal relationships that could have influenced the work reported in this paper.

Contributors: HKC and LF contributed equally to this work and are the co-first authors. HKC contributed to manuscript preparation, manuscript review, data analysis, and investigation. LF contributed to manuscript preparation, manuscript review and data analysis. ZD contributed to manuscript review and investigation. XTL contributed to manuscript review and investigation. YW contributed to manuscript review, validation, and methodology. XZZ contributed to manuscript review, project administration, and methodology. LFD contributed to manuscript review, supervision, funding acquisition, and conceptualization. All authors have read and approved the final manuscript. LFD is the guarantor.

All the supplementary files in this paper are available at http://wjem.com.cn.