Despite having independent association with a number of outcomes, SVD MRI features lacked prognostic value for patients with stroke.
Thomas Tourdias, MD, PhD,
Data from a new study published in Neurology suggest that cerebral small vessel disease (SVD) MRI features do not improve stroke outcome model predictions despite being associated with patient functional, cognitive, and psychological outcomes.
The first model developed in the study (model 1), based on age, baseline National Institute of Health Stroke Scale (NIHSS) score, and infarct volume, performed well at discriminating poor vs. good functional outcomes (area under the curve [AUC], 0.915) and performed fairly at identifying cognitively impaired (AUC, 0.750) and cognitively distressed (AUC, 0.688) patients. Model 2 (created by adding the total SVD score to model 1) and model 3 (adding individual SVD MRI features to model 1) did not perform significantly better than model 1, despite higher SVD scores associating with poorer outcomes for patients (odds ratio [OR], 1.30; 95% CI, 1.07–1.58; P = .009).
Senior author Thomas Tourdias, MD, PhD, professor, medical imaging, Bordeaux University, Bordeaux University Hospital, and colleagues stated that this “’negative’ result is of substantial value because [the] study included a large dataset of patients, prospectively and longitudinally followed over time, with applications of all major methodological requirements for prognostic research design and analysis.”
Tourdias and colleagues looked at 2 datasets of patient information. Of the 348 patients from the “Brain before Stroke” cohort, the mean age was 67.5 years (SD, 14.1) and the median NIHSS score was 4 (interquartile range [IQR], 2–8). Of the 137 patients from the “Strokedem” cohort, the mean age was 64.8 (SD, 12.6) and the median NIHSS score was 1 (IQR, 0–2). Both cohorts were around 60% male.
The biggest driving factor in model 1’s prediction success was NIHSS score, which was significantly associated with a poor modified Rankin score (mRS) at 3 months (odds ratio [OR] for log-transformed NIHSS, 10.56; 95% CI, 5.83–19.14) and depression. Higher age was also significantly associated with mRS and depression at 3 months.
READ MORE: Machine Learning Models Accurately Predict Outcomes in Subarachnoid Hemorrhages
Tourdias and colleagues found a significant association between SVD score and functional outcome (OR, 1.30; 95% CI, 1.07–1.58; P = .009), and cognitive outcome (OR, 1.23; 95% CI, 1.00–1.5; P = .05) and an insignificant association between SVD score and psychological outcome (OR, 1.16; 95% CI, 0.92–1.45; P = .2077). However, when added to the multivariate analyses, total SVD score did not improve outcome predictions. Combined models were not statistically different from model 1, with an AUC difference of 0.003 (95% CI, -0.006 to 0.014) for functional outcome, -0.004 (95% CI, -0.018 to 0.010) for cognitive outcome, and –0.004 (95% CI, -0.030 to 0.023) for psychological outcome.
Total SVD score alone was also investigated for its ability to predict outcomes and was found to be close to random for all outcomes, with ORs of around 0.5. Total SVD score was determined by white matter hyperintensity (WMH), lacunes, perivascular space, microbleeds and atrophy, of which WMH was consistently found to be associated with poor outcome. Tourdias and colleagues also found a significant association between brain volume loss and poor functional outcome at 3 months (OR, 1.18; 95% CI, 1.06–1.31; P = .0029).
Tourdias and colleagues hypothesize that “[one] reason why cerebral SVD does not add information to the prediction might have been that age and cerebral SVA are correlated...age, which was included in our models, [could have conveyed] most of the information.”