Identification and Validation of Hematoma Volume Cutoffs in Spontaneous, Supratentorial Deep Intracerebral Hemorrhage

Abstract

Background and Purpose:

Clinical trials in spontaneous intracerebral hemorrhage (ICH) have used volume cutoffs as inclusion criteria to select populations in which the effects of interventions are likely to be greatest. However, optimal volume cutoffs for predicting poor outcome in deep locations (thalamus versus basal ganglia) are unknown.

Methods:

We conducted a two-phase study to determine ICH volume cutoffs for poor outcome (modified Rankin Scale 4-6) in the thalamus and basal ganglia. Cutoffs with optimal sensitivity and specificity for poor outcome were identified in the Ethnic/Racial Variations of Intracerebral Hemorrhage (ERICH) study (derivation cohort) using receiver-operating curves (ROC) analysis. The cutoffs were then validated in the Antihypertensive Treatment of Acute Cerebral Hemorrhage-2 (ATACH-2) trial by comparing the c-statistic of regression models for outcome (including dichotomized volume) in the validation cohort.

Results:

Of 3000 patients enrolled in ERICH, 1564 (52%) had deep ICH, of whom 1305 (84%) had complete neuroimaging and outcome data (660 thalamic and 645 basal ganglia hemorrhages). ROC analysis identified 8 ml in thalamic (area under the curve [AUC] 0.79, sensitivity 73%, specificity 78%) and 18 ml in basal ganglia ICH (AUC 0.79, sensitivity 70%, specificity 83%) as optimal cutoffs for predicting poor outcome. The validation cohort included 834 (84%) patients with deep ICH and complete neuroimaging data enrolled in ATACH-2 (353 thalamic and 431 basal ganglia hemorrhages). In thalamic ICH, the c-statistic of the multivariable outcome model including dichotomized ICH volume was 0.80 (95% CI 0.75-0.85) in the validation cohort. For basal ganglia ICH, the c-statistic was 0.81 (95% CI 0.76-0.85) in the validation cohort.

Conclusions:

Optimal hematoma volume cutoffs for predicting poor outcome in deep ICH vary by the specific deep brain nucleus involved. Utilization of location-specific volume cutoffs may improve clinical trial design by targeting deep ICH patients that will obtain maximal benefit from candidate therapies.

INTRODUCTION

Spontaneous intracerebral hemorrhage (ICH) remains the most devastating type of stroke with no therapies of proven benefit.¹ Clinical trials for this disease have begun using volume cutoffs and location (lobar versus deep) as eligibility criteria to target populations in which the effects of candidate therapies are likely to be greatest,^2-4 a strategy that has been successful in trials for ischemic stroke.^5-7 Due to its severity and relatively homogenous pathophysiology (cerebral small vessel disease secondary to long-standing hypertension), deep ICH has become a population of interest for recent trials.^8,9 However, current selection criteria utilized in these studies do not take into account known differences in radiographic and clinical features among deep structures (thalamus and basal ganglia) that may be important for proof-of-concept trials.^10,11

It is well known clinically that small hemorrhages in deep brain nuclei can have devastating effects,¹² but this observation has not been well-studied or quantified for use in trial eligibility criteria. Prior small single center studies found that hematoma volume cutoffs for poor outcome were smaller in thalamic than basal ganglia ICH.¹³ However, these studies were prone to overfitting given their single-center design and were likely underpowered to determine specific volume cutoffs associated with poor outcome among deep locations.

We therefore present a two-phase analysis in which we develop two tools for predicting poor outcome in deep ICH: first, we derive and validate a prognostic model using predictors identified in multivariable analysis; second, we identify and validate location-specific ICH volume cutoffs in the thalamus and basal ganglia. We used two well-phenotyped studies of ICH: the multicenter observational Ethnic/Racial Variations of Intracerebral Hemorrhage (ERICH) study¹⁴ to develop a prognostic model and identify volume cutoffs, followed by validation of these prediction tools in the Antihypertensive Treatment of Acute Cerebral Hemorrhage-2 (ATACH-2) randomized clinical trial.²

METHODS

Data Availability

Anonymized data from ERICH and ATACH-2 are available through the National Institute of Neurological Disorders and Stroke (NINDS) Archive of Clinical Research Datasets and can be accessed at https://www.ninds.nih.gov.

Study design and Inclusion Criteria

We conducted a two-phase study to derive and validate a prognostic model and hematoma volume cutoffs for poor outcome in the thalamus and basal ganglia. Patients enrolled in the ERICH and ATACH-2 studies with spontaneous, supratentorial deep ICH and available neuroimaging and 3-month outcome data were included. We first derived a prediction model and identified volume cutoffs using data from the ERICH study, a U.S.-based, prospective, multi-center, case-control study of 3,000 ICH cases, including 1,000 patients of each race-ethnicity (non-Hispanic whites, non-Hispanic blacks and Hispanics).¹⁴ We then validated these prediction tools in the ATACH-2 study, a randomized, international, multicenter trial of intensive blood pressure control.² We used a clinical trial population as our validation cohort since it is the population in which the proposed volume cutoffs are likely to be used. Informed consent for participation in each individual study was obtained from participants or their legally designated surrogates. Institutional review boards or ethical committees reviewed and approved all study protocols. Data were accessed via tailored application submitted to the study steering committee application (ERICH) or the NINDS Archive for Clinical Research Datasets (ATACH-2).

Neuroimaging

Neuroimaging cores specific to the individual studies and blinded to clinical status determined hemorrhage location and measured ICH and IVH volumes using semi-automated methods. If multiple hemorrhages were present, the location of the first symptomatic hemorrhage was used.

Clinical Outcomes

The primary outcome was poor outcome at 3-months, defined as a modified Rankin Scale (mRS) of greater than or equal to 4. The mRS was analyzed as a dichotomous variable as this has been the standard in ICH clinical trials. The mRS was obtained via in-person clinical evaluation (ATACH-2) or structured telephone interviews (ERICH) at 3 months.

Statistical Methods

Descriptive statistics are presented as counts (percentages [%]) for categorical variables, means (standard deviation [SD]) for continuous normally distributed variables, and medians (interquartile range [IQR]) for non-normal continuous variables. Unadjusted differences in baseline and imaging characteristics by location (thalamus versus basal ganglia) were evaluated using the Fisher exact test (2‐tailed), Kruskal–Wallis test, or unpaired t test, as appropriate.

Regression Analyses.

We performed univariable and multivariable logistic regression analysis to identify covariates associated with a poor outcome (mRS ≥ 4) at 3 months in the derivation cohort. Multivariable model building proceeded as follows: first, covariates with p<0.1 in univariable analyses were included; second, universal confounders (age and sex) were force entered; third, covariates with p>0.1 were backward eliminated; fourth, collinear covariates, as expressed by a variance inflation factor >5, were identified and one covariate was removed from the model. Specific deep location (thalamus versus basal ganglia) was included as a covariate.

Prognostic Model.

We used the covariates identified in multivariable regression to construct a model for predicting poor outcome at 3 months in the derivation cohort. We evaluated the discrimination ability of this prognostic model in the validation cohort using the c-statistic and by comparing the observed versus expected risk of poor outcome by quintile of predicted probability.

ROC Analyses.

Receiver operating characteristic (ROC) curves were used to determine ICH volume cutoffs and associated sensitivity and specificity in predicting 3-month poor outcome in thalamic and basal ganglia locations. Volume cutoffs were chosen to optimize the sensitivity and specificity for poor outcome in univariable analysis. The identified cutoffs were then validated by comparing the area under the curve (AUC) from the ROC analysis and c-statistic of multivariable models for outcome including age, sex, admission GCS, presence of IVH, and ICH volume dichotomized at the chosen cutoff in the derivation and validation cohorts. We then performed three sensitivity analyses: first, we determined volume cutoffs for poor outcome in thalamic ICH with IVH to investigate the interaction between IVH and ICH volume on poor outcome; second, we determined optimal cutoffs using ICH volume at 24 hours; third, we determined volume cutoffs after excluding patients who underwent withdrawal of care in the derivation cohort.

For all statistical analyses, a two-sided p-value of 0.05 was set as the significance threshold and 95% confidence intervals were reported for all odds ratios. R (version 3.5.1) was used for all analyses.

RESULTS

Of the 3,000 patients enrolled in ERICH, 1,564 (52%) had supratentorial deep ICH. Of these, 1,305 (84%) patients had complete neuroimaging and outcome data (660 thalamic and 645 basal ganglia [537 putamen and 108 caudate]) and were included in the derivation cohort. Of the 1,000 patients enrolled in ATACH-2, 870 (87%) had supratentorial, deep ICH, and 836 (96%) of these had complete neuroimaging and outcome data (355 thalamic and 481 basal ganglia) and were included in the validation cohort (Figure 1). In both studies, patients with thalamic ICH were older and more likely to have atrial fibrillation and a history of ischemic stroke (Table 1).

Figure 1. Participant Enrollment and Eligibility Criteria.

Abbreviations: ICH = intracerebral hemorrhage. Flowchart of patient inclusion and exclusion criteria. Patients with supratentorial, spontaneous deep ICH and complete neuroimaging and outcome data were included.

Table 1.

Baseline Demographic and Clinical Characteristics, by Location.

Variables	Derivation Cohort ERICH (n=1305)			Validation Cohort ATACH-2 (n=834)
Deep Location (No. of individuals)	Thalamus (n=660)	Basal Ganglia (n=645)	p	Thalamus (n=353)	Basal Ganglia (n=481)	p
Age (years), mean (SD)	61 (13)	57 (13)	<0.001	65 (12)	59 (13)	<0.001
Sex (male)	395 (65)	405 (65)	0.82	195 (60)	279 (65)	0.25
Black	247 (40)	216 (35)	0.05	33 (10)	45 (10)	0.22
White	153 (25)	171 (28)	0.34	91 (28)	91 (21)	0.03
Hispanic	212 (35)	233 (38)	0.31	18 (6)	36 (8)	0.22
Hypertension	553 (91)	522 (85)	0.001	248 (80)	340 (81)	0.75
Diabetes	161 (26)	184 (30)	0.20	61 (19)	64 (15)	0.17
Hyperlipidemia	238 (40)	252 (41)	0.58	84 (28)	85 (21)	0.05
Congestive heart failure	43 (7)	31 (5)	0.33	14 (4)	6 (1)	0.02
Atrial fibrillation	71 (12)	43 (7)	0.006	15 (5)	8 (2)	0.04
Prior ischemic stroke	81 (13)	57 (9)	0.03	69 (22)	51 (12)	0.001
Smoker*	235 (38)	247 (39)	0.85	157 (49)	181 (42)	0.08
Cocaine use	50 (8)	40 (7)	0.29	7 (2)	10 (2)	1.00
On antihypertensive medication	346 (57)	308 (50)	0.02	159 (50)	191 (45)	0.24
Admission GCS, median (IQR)*	15 (4)	14 (5)	0.02	15 (1)	15 (2)	<0.001
Admission systolic BP	196 (48)	192 (54)	0.20	173 (24)	179 (26)	0.006
Admission diastolic BP	113 (62)	114 (76)	0.74	111 (20)	114 (20)	0.05
Admission INR*	1.2 (0.7)	1.1 (0.4)	0.10	1.0 (0.1)	1.0 (0.1)	0.03
Admission ICH volume, mL, median (IQR)	7 (10)	14 (25)	<0.001	7 (8)	12 (15)	<0.001
IVH present on admission	403 (66)	212 (34)	<0.001	156 (48)	51 (12)	<0.001
3-Month Poor Outcome (mRS 4-6)	344 (52)	306 (47)	0.07	163 (46)	162 (34)	<0.001
3-Month Mortality	120 (18)	108 (16)	0.53	26 (7)	24 (5)	0.20

All values displayed as count (%) unless otherwise specified.

Abbreviations: ICH = intracerebral hemorrhage; IVH = intraventricular hemorrhage; GCS = Glasgow Coma Scale; IQR = interquartile range; BP = blood pressure; mRS = modified Rankin Scale

Radiographic Characteristics by Deep Location

Patients with thalamic ICH had smaller median ICH volumes and the median volumes for both thalamic and basal ganglia ICH were consistent across studies (ERICH, 7 mL [IQR 10 mL] in thalamic versus 14 mL [IQR 25 mL] in basal ganglia ICH; p < 0.001, ATACH-2, 7 mL [IQR 8 mL] versus 12 mL [IQR 15]; p <0.001). In both studies, patients with thalamic ICH were more likely to have intraventricular hemorrhage (IVH) (p <0.001).

Specific Deep Location and Functional Outcome

In the derivation cohort, about half (48%) of all patients with supratentorial, deep ICH had a poor outcome (mRS ≥ 4) at three months. The median ICH volume was higher among those with a poor versus good outcome in both thalamic (13 mL [IQR 13 mL] versus 4 mL [IQR 5 mL]; p <0.001) and basal ganglia (28 mL [IQR 34 mL] versus 8 mL [IQR 12 mL]; p <0.001) locations. In univariable analysis, thalamic location was not associated with poor outcome (OR 1.23, 95% CI 0.99–1.53; p=0.06). However, in the multivariable model adjusted for age, sex, admission GCS, ICH volume, and IVH, thalamic location was independently associated with poor outcome at 3 months (OR 2.04, 95% CI 1.47–2.83; p=<0.001) (Table 2).

Table 2.

Predictors of Poor Outcome in the Derivation Cohort

Predictor	Univariable		Multivariable
	OR (95% CI)	p	OR (95% CI)	p
Age (years)	1.04 (1.03-1.05)	<0.001	1.07 (1.05-1.08)	<0.001
Female Sex	1.54 (1.23-1.93)	0.007	1.58 (1.16-2.13)	0.002
GCS	0.75 (0.72-0.78)	<0.001	0.82 (0.78-0.86)	<0.001
ICH Volume (mL)	1.08 (1.06-1.10)	<0.001	1.09 (1.08-1.10)	<0.001
IVH	3.40 (2.69-4.30)	<0.001	1.52 (1.11-2.07)	0.009
Thalamic location	1.23 (0.99-1.53)	0.06	2.04 (1.47-2.83)	<0.001

Abbreviations: ICH = intracerebral hemorrhage; IVH = intraventricular hemorrhage; GCS = Glasgow Coma Scale; OR = odds ratio; CI = confidence interval

Prognostic Model

Multivariable logistic regression analyses identified age, sex, ICH volume, GCS, presence of IVH, and thalamic location as strong independent predictors of outcome after ICH. We used the estimated regression coefficients of these variables and the mathematical relationship between risk and odds (risk = odds / [ 1 + odds]) to construct the following formula for the predicted probability of 3-month poor outcome in deep ICH:

$${\mathit{Risk}}_{(3\mathrm{month}\mathrm{mRS}4-6)}=\frac{{\mathrm{e}}^{-3.1+0.10[\text{ICH volume}]+0.06[\text{age}]+0.48[\text{female}]-0.21[\text{GCS}]+0.44[\text{IVH}]+0.67[\text{thalamic}]}}{1+{\mathrm{e}}^{-3.1+0.10[\text{ICH volume}]+0.06[\text{age}]+0.48[\text{female}]-0.21[\text{GCS}]+0.44[\text{IVH}]+0.67[\text{thalamic}]}},$$

where ICH volume is measured in mL and age in years, GCS is the total GCS score, and female sex, IVH, and thalamic location are given a value of 1 if present and 0 if absent. The prognostic model predicted poor outcome with an AUC of 0.86 in the derivation cohort and an AUC of 0.81 in the validation cohort. The proportion of patients with a poor outcome in the derivation and validation cohorts by quintile of predicted probability is shown in Table 3.

Table 3.

Observed versus Expected Risk of Poor Outcome by Quintile of Predicted Probability

Quintile of Predicted Probability of Poor Outcome	Poor Outcome in the Derivation Cohort (n=1305)		Poor Outcome in the Validation Cohort (n=834)
	Expected	Observed	Expected	Observed
0.01-0.17	12%	12% (31/261)	11%	9% (16/156)
0.18-0.33	29%	22% (57/261)	25%	21% (40/192)
0.34-0.57	55%	42% (109/261)	44%	45% (95/213)
0.58-0.85	84%	77% (203/261)	70%	61% (104/170)
0.85-1.0	99%	93% (243/261)	92%	92% (69/75)
Model Discrimination
AUC	0.86		0.81

AUC = area under the curve

ICH Volume Cutoffs for Poor Outcome by Deep Location

The area under the curve (AUC) for ICH volume in thalamic ICH was 0.79 in the derivation cohort and 0.75 in the validation cohort. In basal ganglia ICH, the AUC for ICH volume was 0.79 in the derivation cohort and 0.77 in the validation cohort. Univariable ROC analysis identified 8 mL in thalamic and 18 mL in basal ganglia ICH as having optimal sensitivity and specificity for predicting poor outcome in the derivation cohort (Table 4). The 8 mL cutoff had a positive predictive value (PPV) of 76% in thalamic locations, and the 18 mL cutoff had a PPV of 77% in basal ganglia locations. In thalamic ICH, the multivariable model for poor outcome including ICH volume dichotomized at 8 mL, age, sex, admission GCS and IVH had a c-statistic of 0.86 (95% CI 0.84–0.89) in the derivation cohort and 0.80 (95% CI 0.75–0.85) in the validation cohort (Figure 2A). In basal ganglia ICH, the multivariable model had a c-statistic of 0.85 (95% CI 0.82–0.88) in the derivation cohort and 0.80 (95% CI 0.76–0.85) in the validation cohort (Figure 2B). In a secondary analysis in patients with thalamic ICH and IVH in the derivation cohort, we found the volume cutoff for poor outcome was slightly higher from that in all thalamic ICH (8.3 mL vs 8.0 mL). In exploratory analyses using 24 hour ICH volume in the derivation cohort, the optimal cutoff for poor outcome in thalamic ICH increased from 8 mL to 9 mL and the cutoff in basal ganglia ICH increased from 18 mL to 29 mL. Exclusion of patients who underwent withdrawal of care in the derivation cohort did not change the optimal volume cutoffs for poor outcome in thalamic (7.9 mL in population without withdrawal of care versus 8.0 mL overall) or basal ganglia ICH (18.4 mL in population without withdrawal of care vs 18.0 mL overall).

Table 4.

ICH Volume Cutoffs for Predicting Poor Outcome in the Derivation Cohort

Location	Cutoff Volume (mL)	AUC	Sensitivity	Specificity	PPV	NPV
Thalamus	8	0.79	72%	78%	76%	75%
Basal Ganglia	18	0.79	70%	83%	77%	77%

Abbreviations: ICH = intracerebral hemorrhage; AUC = area under the curve; CI = confidence interval; PPV = positive predictive value; NPV = negative predictive value

Figure 2. Multivariable ROC Curves for Predicting Poor Outcome Using Volume Cutoffs, by Location.

A. Multivariable ROC curves for 3-month poor outcome (mRS 4-6) in thalamic ICH. Black line shows the area under the curve (AUC) of c=0.87 for a multivariable model including ICH volume dichotomized at the 8 mL cutoff in the derivation cohort (ERICH). Grey line shows an AUC of c=0.80 of the same model in the validation cohort (ATACH-2). B. Multivariable ROC curves for 3-month poor outcome (mRS 4-6) in basal ganglia ICH. Black line shows the area under the curve (AUC) of c=0.86 for a multivariable model including ICH volume dichotomized at the 8 mL cutoff in the derivation cohort (ERICH). Grey line shows an AUC of c=0.81 of the same model in the validation cohort (ATACH-2).

DISCUSSION

We report a large, 2-stage (derivation and validation) study in well-phenotyped observational and clinical trial data to develop two outcome prediction tools for use in deep ICH. First, we developed a model for prognostication in clinical trials using independent predictors, including ICH volume and thalamic location, identified in multivariable analysis. Second, we identified 8 mL in thalamic and 18 mL in basal ganglia locations as hematoma volume cutoffs with maximal sensitivity and specificity for predicting poor outcome in the ERICH derivation cohort, and validated these cutoffs in the ATACH-2 trial population by demonstrating strong predictive power (c-statistic > 0.80) in multivariable models for poor outcome. Our findings confirm the independent association of thalamic location with poor outcome in a large, racially and ethnically diverse population of ICH.

To date, most trial enrollment criteria and clinical grading scales, including the ICH score, have used a cutoff of 30 mL to identify patients likely to have a poor outcome.^12,15 However, this cutoff was not developed with regard to location and may underestimate the devastating effects of smaller deep hemorrhages. A lack of specific volume cutoffs for use in deep ICH may have contributed to variability in enrollment criteria and lack of positive results in recent trials. Our proposed volume cutoffs in thalamic and basal ganglia ICH may be utilized in proof-of-concept trials that are likely to have small absolute effect sizes, including those targeting secondary injury. We propose basal ganglia ICH, specifically those located in the putamen, as an ideal population to target in these trials. The clinical consequences of ICH involving the putamen are likely the direct result of parenchymal hemorrhage with minimal confounding by IVH, hydrocephalus, brainstem injury, and other neuroanatomic factors.¹⁶

While volume cutoffs may be useful for guiding clinical trial inclusion and exclusion criteria, we also provide a prognostic model that incorporates ICH volume as a continuous variable for more accurate prediction of outcome for selection of patients in clinical trials. While several prognostic models for ICH have been developed and validated, these models do not account for differences in hematoma volume by location or the independent effect of specific deep location.^15,17-19 We therefore developed our prognostic model for use in deep ICH to account for these differences. Our results also confirm reports of other studies that demonstrate thalamic location is an independent predictor of poor outcome. A pooled analysis of the INTERACT studies, the largest study to assess the influence of specific location on functional outcome, found that ICH affecting the thalamus and posterior limb of the internal capsule showed the strongest association with poor outcome.²⁰ Our study adds to these findings, demonstrating that in a racially and ethnically diverse population, thalamic ICH has worse outcomes than ICH in other deep locations. The reason for this difference remains unclear. Patients with thalamic hemorrhages were older and were more likely to have associated IVH, both of which may confound outcome. However, these differences were controlled for in multivariable models. It is possible that there are underlying differences in biology or predisposition to secondary injury that may influence outcomes differentially in these locations. Further research is needed to determine factors associated with poor outcome among specific deep locations, in hopes of identifying therapies that may have location-specific effects.

The strengths of our study are the large sample size and use of two multi-center cohorts to develop and validate an outcome prediction model and volume cutoffs and in deep ICH. Our study has several limitations. First, we lack more granular location data, including data on whether hemorrhages involve multiple locations. It is possible that larger hemorrhages may involve both the thalamus and basal ganglia. Second, we are limited by use of a clinical trial population as our validation cohort. While the ERICH study did not have limitations on patient selection, the ATACH-2 study excluded patients with ICH volume greater than 60mL or pre-morbid disability requiring assistance in ambulation or activities of daily living. These restrictions in the validation cohort may limit the external validity of our prediction tools. Our results demonstrate that use of the prognostic model and identified volume cutoffs in a clinical trial population is feasible and relevant for future trial design. However, these tools have not been validated in an unselected patient population and are not suitable for prognostication in general patient populations. A third limitation is that our identified optimal volume cutoffs were selected using a method that assigns equal weight to maximizing sensitivity and specificity. While this approach mathematically balances the trade-off of more specific, larger volume cutoffs for fewer observations, it may not always be the most clinically relevant. However, our identified cutoffs have high positive predictive values for poor outcome, such that there is a high likelihood that patients with hematoma volumes above the cutoff will have a poor outcome. Finally, we lack data on variations in patient care that could have contributed to differences in outcome and mortality between the ERICH and ATACH-2 studies.

In conclusion, we developed and validated a prognostic model incorporating thalamic location and easy-to-use volume cutoffs in thalamic and basal ganglia ICH with maximal sensitivity and specificity for poor outcome for use in selecting patients for clinical trials. These predictive tools may be utilized to improve clinical trial design by targeting patients with deep ICH who are likely to obtain maximal benefit from candidate therapies.