Sella gradeA higher intracranial pressure was associated with an emptier sella, showing a stepwise increase in ICP between each grade (Kruskal–Wallis statistic = 57.2, p < 0.0001, Fig. 1B). Significant differences were observed between sella grade I and all other grades, as well as between sella grade II and IV. Abnormal sella morphology (Yuh grade 3–5) was linked with higher ICP (9.5 mmHg) compared to those with ‘normal’ sella shape (Yuh grades 1–2, 3.5 mmHg) [Mann–Whitney U = 8596, p < 0.0001].Vertical tortuosityPatients with tortuous optic nerves had a higher intracranial pressure (KW = 17.2, p < 0.001), with a significant pairwise increase between none and unilateral and none and bilateral signs (Fig. 2A, B). Those with either one or two tortuous optic nerves had significantly higher median ICP (6.7 mmHg) as compared to those without signs of tortuosity (3.5 mmHg, MWU = 12,358, p < 0.0001).Optic nerve sheath distensionPatients with a distended optic nerve sheath had a higher intracranial pressure (KW = 29.4, p < 0.0001), with a stepwise increase between those with no ONSD and unilateral ONSD or bilateral ONSD (Fig. 2C, D). Those with one or two optic nerves exhibiting nerve sheath distension had significantly higher median ICP (10.1 mmHg) as compared to those without signs (3.7 mmHg, MWU = 5605.5, p < 0.0001).Globe flatteningPatients with optic globe flattening had a higher intracranial pressure (KW = 30.2, p < 0.0001), with a significant pairwise increase between none, unilateral and bilateral signs (Fig. 2E, F). Patients with one or two optic globes showing flattened appearances had significantly raised ICP (10.1 mmHg) as compared to those without signs (3.7 mmHg, MWU = 6839.5, p < 0.0001).Optic disc protrusionPatients with protruded optic discs had a higher intracranial pressure (KW = 21.1, p < 0.0001), with a significant pairwise increase between none and unilateral and none and bilateral signs (Fig. 2G,H). Those with evidence of uni- or bilateral optic disc protrusion had significantly raised median ICP (16.2 mmHg) as compared to those without signs (4.0 mmHg, MWU = 3278, p < 0.0001).After exclusion of patients who had surgical CSF diversion, the significance of the above results with were maintained (Supplementary Table 1), except for vertical tortuosity.Reversibility36 patients (28 female) met the criteria for subgroup reversibility analysis with a mean age of 40 years (SD = 18.0). The most frequent diagnosis prior to ICP monitoring and intervention was a Chiari malformation (17%) followed by IIH (14%). Four patients (11%) had no formal diagnosis or underlying structural abnormality and were being investigated for persistent, non-specific high or low intracranial pressure symptoms. This smaller cohort was statistically equivalent to the remainder of the cohort with respect to age, sex and diagnosis. Here, the median interval between intervention and post-intervention MRI scan was 235 days (IQR: 74–529).Among patients who were surgically naive and then who had CSF diverting treatment following ICP monitoring, only the sella grade was found to be reversible (pre-intervention Yuh median = 2 (IQR: 1–3), post-intervention Yuh median = 1 (IQR: 1–2), Wilcoxon Sign-Rank, p < 0.001) (Supplementary Table 2). ONSD demonstrated a trend toward reversibility (p < 0.1), although this value was uncorrected.Multivariate prediction of ICP using radiological markersThree multivariate models were fitted to best characterise the association between radiological markers and intracranial pressure. In all three models the use of multiplicative interaction terms showed no improvement in model scoring and are therefore reported without interaction.The first model, a sparse ordinary least squares (OLS) model, simulated a clinician reviewing an MRI scan without additional clinical information. Pairwise Cramér’s V associations addressed collinearity, leading to the removal of ONSD due to its strong correlation with PGF (Supplementary Fig. 1), resulting in no change in adjusted R2. This model, explaining 26% of the variance, demonstrated all four remaining radiological markers significantly and independently associated with ICP, outperforming univariate OLS regression (Table 2, Supplementary Table 3).Table 2 Comparison of multivariate models used to predict 24-h median intracranial pressure.In the second, a more complete OLS model was fitted that included salient demographic and clinical variables including patient age, sex, suspected diagnosis and whether they had a previous surgical intervention for CSF diversion. Variables were randomly permuted and the model which minimised the Bayesian Information Criterion the most was selected (i.e., the most parsimonious). In this way, patient age and diagnosis were retained, and previous surgery was not found to be informative. This model explained around a third of the variance (adjusted R2 = 0.33). Here, sella grade, VT and ODP remained significant after clinical confounder adjustment (Table 2, Supplementary Table 4). Individual diagnoses were not significantly associated with ICP after multivariate adjustment; however, age did remain significant.In the third, a Bayesian linear model was fitted using the previous clinical and radiological variables, however here no specification was made on the relationship between independent variables and median ICP. This had a higher adjusted R2 (0.37) as compared to the previous model, a and a mean absolute error of 2.71 mmHg. Further, one could assess the degree of credibility in the coefficients of each imaging variable when other variables were held at the sample mean (Fig. 3).Fig. 3Conditional adjusted predictions for individual radiological biomarkers. All other covariates, including other radiological markers, are constant at their sample mean. Dark blue line represents mean with 94% credible intervals shown in the pale blue bars with interpolation. ODP = optic disc protrusion, PGF = posterior globe flattening, VT = vertical tortuosity, Yuh = pituitary sella grade.
