The interrater agreement analysis showed high reliability (ICC = 0.89, 95% CI 0.81–0.93 for the T1WI-based score; ICC = 0.86, 95% CI 0.74–0.92 for the T2WI-based score).
Brain lesions were common in each dataset (table ). Most subjects had a moderate BALI score, and none was free of lesions (datasets 1–3: minimum = 5, 6, 3; maximum = 16, 16, 17; median and mean = 10, 10, 9). Brain lesions were seldom restricted to a single category. Across datasets, nearly 85% of the participants showed various grades of lesions involving multiple categories (i.e. BG, DWM, PV, IT and GM-SV). Except for the ‘other lesion’ category, important lesions (i.e. subscore in a category of ≥2) were observed for each category. The highest subscore was in the DWM category, with 20% of the subjects having a DWM subscore of 3 and 9% of 4, but none of 5. Higher BALI total scores were revealed by T2WI than by T1WI for the datasets (two-way ANOVA: n = 249, F = 7.65, p = 0.009) (table ), which was most obvious in PV and DWM. Regardless of imaging method or dataset, the BALI score varied significantly among different diagnoses (two-way ANOVA: F = 32.03, p < 0.001), with AD or MCI patients showing a higher BALI score than healthy controls (p = 0.016) (table ).
Total and categorical scores of BALI
Despite the difference in mean values, T1WI- and T2WI-based BALI scores were highly correlated with a linear fit (fig. ). Significant correlations between T1WI- and T2WI-based scores were also observed for each diagnostic group in each dataset (r = 0.87–0.93; p < 0.001). The T1WI-based subscores were also highly correlated with T2WI-based ones (r = 0.64–0.94; p < 0.001).
Interrelations between T1WI- and T2WI-based BALI. Solid line: linear fit y = a + bx, where a = 3.093, b = 0.797, n = 249, r = 0.904 (95% CI: 0.879–0.925) and p < 0.001. Dotted diagonal line: T1WI-based scores = T2WI-based scores.
A lower BALI total score (i.e. less lesions), either T1WI or T2WI based, was associated with a higher cognitive test result (i.e. better performance; r = 0.21–0.29, p < 0.05 for T1WI-based score; r = 0.26–0.29, p < 0.01 for T2WI-based score) (fig. ). Cognitive test scores were consistently related with brain atrophy, but not significantly related with brain lesions in most individual categories, except for BG in dataset 1 (fig. ).
Fig. 3 BALI in relation to MMSE (a, b) and to age (c, d) based on T1WI (white bars) and T2WI (gray bars) in datasets 1 (a, c) and 3 (b, d). Data presented are Pearson correlations for the total score and scores for various subcategories. Dotted lines: correlations (more ...)
Both the T1WI- and T2WI-based scores increased with age (fig. ). A strong correlation was observed for most lesion categories versus age. The slopes of the linear equation describing changes in T1WI- and T2WI-based BALI scores with age did not differ, although the intercept for the T2WI-based score was slightly higher (fig. ). When people of different diagnostic groups were considered separately, the significant correlation between age and the BALI scores still held. The BALI score was higher in men than in women (10.0 ± 3.4 vs. 9.0 ± 3.7; F = 7.06; p = 0.009).
Fig. 4 BALI as a function of age. Correlation between age and T1WI-based score (a), and T2WI-based score (b). Solid line: linear fit (y = a + bx) for all subjects, regardless of dataset – for the T1WI-based score: a = −7.750, b = 0.235, n = 293, (more ...)
T1WI- and T2WI-based BALI scores showed comparable accuracy at classifying people with AD at the individual level. At its midpoint (i.e. ≤10 vs. >10), the BALI score sensitively predicted 17–20% more AD cases, with a sensitivity of 63–64%, whereas it predicted 10–12% fewer healthy controls, with a specificity of 62–73% (table ). The area under the curve (AUC) ranged between 0.67 and 0.70 (fig. ).
Accuracy of classification of BALI at its midpoint value (n)
Fig. 5 ROC curves for identifying subjects with AD. Solid line: T1WI-based BALI in dataset 1 (AUC = 0.70; 95% CI: 0.57–0.83). Dashed line: T2WI-based BALI in dataset 1 (AUC = 0.67; 95% CI: 0.56–0.79). Dotted line: T1WI-based BALI in dataset 2 (more ...)