We retrospectively studied our database for patients with ICH aged ≥18 years between January 2010 and August 2015. As inclusion criteria we defined: 1) spontaneous ICH confirmed on NCCT and 2) NCCT and CTA performed on admission within 6 hours after onset of symptoms. Patients were excluded if they had head trauma, brain tumor or secondary ICH from hemorrhagic transformation of ischemic infarction. We defined a binary outcome (good vs. poor): as poor outcome we defined modified Rankin Scale (mRS) on discharge >3, good outcome was defined as mRS ≤3. Additionally, we obtained vascular risk factors (hypertension, diabetes mellitus) from patients’ clinical records.

The study was approved by the Ethics Committee of the University Muenster and the Westfalian Chamber of Physicians, Muenster, Germany. All study protocols and procedures were conducted in accordance with the Declaration of Helsinki.

The CT scans were performed using standard clinical parameters with axial 5-mm section thickness. The images were obtained and stored for further evaluation. The location of the hematoma was assessed and documented. Hemorrhage locations were classified as basal ganglia, lobe, brain stem and cerebellum. Two experienced readers independently first evaluated the presence of BHS and BS in all patients’ NCCT and then independently evaluated the presence of SS in the corresponding CTA. Both readers were blinded to all clinical information and the other scans at the time of the ratings. Discrepancies about the occurrence of BHS, BS and SS were settled by joint discussion of the 2 readers.

The NCCT BHS was defined as recently published [8]. Briefly, it consists of a relatively hypodense area (black hole) which is encapsulated within a hyperdense area and which is not connected with the adjacent brain tissue. The relatively hypodense area has an identifiable border and a difference of at least a 28 Hounsfield units (HU) between the 2 density regions. The BHS is illustrated in the Supplementary Figure 1.

The hematoma blend sign [6] represents a hematoma with a hyperdense and a hypodense area, which show a well-defined margin that is easily recognized by the naked eye. There should be a difference of at least 18 HU between the two regions and the hypodense area should not be encapsulated by the hyperdense part of the hematoma (Supplementary Figure 1).

The ICH volumes of the baseline CT were segmented using Analyze (Analyze 10.0; AnalyzeDirect Inc., Overland Park, KS, USA). We therefore extracted the Digital Imaging and Communications in Medicine (DICOM) data of CT scans and implemented them into Analyze. Here the hematoma volume was segmented slice by slice.

Univariable distribution of metric variables is described by median and interquartile range (IQR). Absolute and relative frequencies are given for categorical data. In order to compare two independent samples regarding a metric or categorical outcome we used Mann-Whitney U-test or Fisher’s exact test, respectively. Mc Nemar’s test was used to assess differences between paired portions.

Sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of BHS, BS and SS are given with exact 95% confidence interval (CI). The underlying goldstandard was outcome (good or poor) as defined above in the patients section. We used McNemar’s test to compare sensitivity or specificity of BHS and BS with that of SS, and we used the test of Leisenring et al. [12] for analogous comparisons regarding PPV or NPV. In order to measure inter-rater agreement we used Cohen’s kappa [13].

Association between the clinical and radiological parameters and outcome (good or poor) was assessed by logistic regression analysis. For multivariable model building, stepwise forward selection was used with two-fold interactions being assessed in a second block (inclusion criterion: P-value of the score test ≤0.05, exclusion criterion: P-value of the likelihood ratio test >0.1). Variables considered for multivariable model building are given in Table 1. Given for selected variables are odds ratio (OR) with 95% CI and P-value of likelihood ratio test. For non-selected variables P-value of score test is displayed. Odds were calculated as ratio of the probability for poor outcome to the probability for good outcome. Degree of agreement between imaging parameters was quantified using Cohen’s kappa and visualized by Venn diagram.

No adjustment for multiple testing was performed. Accordingly, P-values are displayed for descriptive reasons in order to detect and study meaningful effects. The P-values are considered as noticeable in case of P≤0.05.

Statistical analyses were performed in SPSS version 24 (IBM Corp., Armonk, NY, USA) and in SAS 9.4 (SAS Institute, Cary, NC, USA).

A total of 182 patients fulfilled the inclusion criteria, and were included in this analysis. The median age of the patients was 68 years (IQR 54-79). Eighty three (45.6%) were female and 99 (54.4%) were male. Poor outcome was observed in 104 patients (57.1%). The median hematoma volume was 22.4 mL (IQR 8.3-42.7) and differed noticeably between patients with poor (median 31.5 mL) and good outcome (median 11.1 mL) (P<0.001). The hematoma was mostly located in the basalganglia (n=87, 47.8%), followed by cerebral lobes (n=81, 44.5%), cerebellum (n=8, 4.4%) and brain stem (n=6, 3.3%). Analyses revealed statistically noticeable differences in the distribution of intraventricular hemorrhage (P=0.003), hematoma volume (P<0.001) and the imaging parameters SS, BS and BHS (all P<0.001) between patients with poor and good outcome (Table 2).

Of 182 patients with spontaneous ICH, 26 (14.3%) presented with BHS, 37 (20.3 %) with BS and 39 (21.4%) with SS on initial imaging (Table 2). Fourteen (7.7 %) patients had all of BHS, BS and SS, and all of them had a poor outcome. In contrast, of the remaining 168 patients, 68 (40.5%) had a poor outcome. Inter-rater agreement for identifying BHS, BS and SS was very high between the 2 readers (BHS: κ=0.934, BS: κ=0.965, and SS: κ=0.984). Discrepancy between the 2 readers was observed in 1 patient for SS, 3 for BHS, and 2 for BS.

The sensitivity, specificity, PPV, and NPV for predicting poor outcome were 23.1%, 97.4%, 92.3%, and 48.7% for BHS, 31.7%, 94.9%, 89.2%, and 51.0% for BS, and 34.6%, 96.2%, 92.3%, and 52.4% for SS (Table 3). SS showed a noticeably higher sensitivity for predicting outcome than BHS in any location (34.6% vs. 23.1%; P=0.029) and in the lobe (45.7% vs. 23.9%; P=0.013). Sensitivity of SS and BS were of comparable magnitude in any location (34.6% vs. 31.7%; P=0.607) and in the lobe (45.7% vs. 43.5%; P=1.000). However, differences between SS vs. BHS and SS vs. BS regarding specificity, PPV and NPV were not statistically noticeable (Table 3).

Logistic regression analysis was performed to assess the association between various clinical and radiological parameters and poor outcome. In univariable logistic regression, higher baseline hematoma volume (P<0.001), intraventricular hemorrhage (P=0.002) and the presence of BHS/BS/SS (all P<0.001) on admission CT scan were associated with poor outcome (Table 1).

Multivariable logistic regression analysis identified intraventricular hemorrhage (OR 2.22, P=0.002), higher baseline hematoma volume (OR 1.03 per mL, P<0.001) and the presence of SS on baseline CT (OR 13.68, P<0.001) as independent predictors of poor outcome (Table 4). With these variables included in the model, presence of BHS and/or BS on baseline CT were not selected as further independent predictors for outcome in multivariable analysis (P=0.116 and P=0.128). This reflects the high degree of association between imaging parameters that was substantial between BS and SS (κ=0.701) and moderate between BHS and SS (κ=0.424) (Table 5). Outcome predictions were discrepant between BHS/BS and SS for only 31 (17.0%)/18 (9.8%) out of 182 patients (Figure 1). These results suggest good performance of BHS, BS and SS in prediction of poor outcome, with SS however remaining the strongest independent predictor of poor outcome compared to BS and BHS, and with BS performing slightly better than BHS.

According to the multivariable model, the probability of poor outcome is relevantly affected by the presence of SS (Figure 2). Whereas prognosis in patients without SS is poor only in the case of high hematoma volume, presence of SS is an indicator for poor outcome even for small hematoma volumes. For example, for a patient without intraventricular haemorrhage and hematoma volume of 20 mL the presence of SS increases the probability of poor outcome to 85%, compared to 33% with absent SS, whereas for a patient without intraventricular haemorrhage and hematoma volume of 80 mL the presence of SS increases the probability of poor outcome to 97%, compared to 76% with absent SS.