Patient Enrollment & Human Protection
This study was approved by the institutional review board of the University of Texas Health Sciences Center at Houston and by the Memorial Hermann Hospital Office of Research. In this retrospective image analysis study, a sub-cohort of 35 ICH patients with IVH, previously enrolled in a randomized clinical safety trial of pioglitazone for hematoma resolution in ICH (SHRINC NCT00827892), were evaluated. SHRINC enrollment inclusion-exclusion criteria and methodology were described elsewhere 37. Inclusion criteria for the present study were 1) patients aged 18-80 years diagnosed with deep ICH and IVH extension, 2) availability of MRI-DTI data for at least two-time points, 3) baseline hematoma volume <100cc, and 4) National Institutes of Health Stroke Scale (NIHSS) <25. Patients who underwent hemicraniectomy were excluded. A total of 85 MRI scans over 4 different time points were included.
Neurological and Radiological Assessments
The baseline clinical assessments included NIHSS, Glasgow Coma Score (GCS), blood pressure, and hematoma volume. The serial neurological assessments included NIHSS and modified Rankin Scale (mRS) in some patients. Serial MRIs were obtained on days 1, 14, 28, and 42. The radiological measurements included total hematoma volume (HV), edema volume (EV), lateral ventricular volume (LVV), ventricular blood volume (VBV), rate of blood clearance, ventricular FA, and MD.
Image Acquisition
Serial images were obtained using Philips and GE full-body MRI systems. Anatomical images included both the 2D and 3D T1-weighted (2D =TR/TE = 460/13 ms, imaging matrix = 256 x 128 mm2, slice thickness = 5 mm; 3D = TR/TE = 8.1/3.7 ms, 256 x 256 x 190 mm3, slice thickness = 1mm), fluid-attenuated inversion recovery (FLAIR) (2D = TR/TE = 11002/152 ms, imaging matrix = 256 x 256 mm2, slice thickness = 3 mm; 3D = TR/TE =8000/337, 256 x 256 x 190 mm3, slice thickness = 1mm), and 2D dual-echo gradient echo (GRE) (TR/TE1/TE2 = 1000/16/32 ms, matrix = 256 x 256 mm2, slice thickness = 3 mm). The quantitative DTI were acquired with TR/TE = 14.5 sec/82.8 ms, b-value = 0, 1000 s/mm2, matrix = 128 x 128, slice thickness = 3 mm.
Sub-Group Analysis of Patients With and Without External Ventricular Drain (EVD)
The patients were dichotomized into those treated with EVD (n=14) and those without EVD (n=12). The changes in ventricular CSF imaging metrics (FA, MD), rate of VBV clearance, and recovery were compared between cohorts.
Hematoma and Lateral Volume Segmentation:
A semi-automated seed growing algorithm using Analyze 12.0 (Analyze Direct Inc., KS, USA) software was used to delineate hematoma and edema volume using FLAIR images at all the time points. The hematoma and perihematomal volumes delineation were saved as a mask. The hematoma volume was further segmented between the parenchyma and ventricular blood volume using signal intensity threshold within and outside the LV. The rate of blood clearance from the ventricle was calculated by the ratio of change in VBV over time.
DTI Image Processing
The DTI image preprocessing was done using FSL (http://www.fmrib.ox.ac.uk) and processing steps are summarized in supplementary material Figure S1.
Clinical and Imaging Metrics Association
Patient NIHSS and drained CSF volumes were further associated with the imaging metrics (i.e., HV and EV, VBV, ventricular FA, and MD). We also evaluated baseline NIHSS and GCS to predict imaging metrics (FA and MD). SHRINC was a drug safety trial, so mRS was not captured and correlated with the imaging metrics.
Statistical Analysis
Bayesian generalized linear mixed modeling (GLMM) was used to evaluate changes in each outcome (total hematoma volume, total VBV, contralesional and ipsilesional FA, MD, and LVV) 38. The present analyses relied on Bayesian inference, providing a posterior distribution with a credible range of values for each effect. A subset of these values was taken as the most credible interval (95% CrI) and a posterior probability (PP) that the effect exists was quantified. The strength of evidence for predictor effects was based on heuristics for varying PP thresholds described in the literature39: no evidence (PP=50%), anecdotal evidence (PP: 51-74%), moderate evidence (PP: 75-90%), strong evidence (PP: 91-96%), very strong evidence (PP: 97-99%), and extreme evidence (PP>99%). A PP ≥ 75% was chosen to indicate a minimum level of support in favor of the existence of predictor effects, consistent with previous work on observational clinical data [R]. Models used weakly informative priors (b=~Normal [μ = 0, σ2 = 100]) to maximize the influence of the present data on PP. Each outcome was fit as a function of time (days) with a random intercept to account for multiple observations per patient. Moderator effects of the EVD group and time-varying changes in VBV (CC), HV, drained CSF volume, and NIHSS on FA and MD were examined. GLMM was able to account for the non-normal distribution of each outcome via specification of link functions: all outcomes except NIHSS and total drained CSF volume were modeled using the lognormal distribution; total drained CSF volume was modeled using the negative binomial distribution; and NIHSS total score was modeled using a zero-one-inflated beta distribution.