EVALUATING SAMPLE-SIZE EFFICIENCY AND SENSITIVITY OF TRACTOMETRY IN ALZHEIMER’S DISEASE - PMC

2.1. Data & Preprocessing

We analyzed dMRI from 3,049 participants in the HABS-HD dataset [10](1,912 F/1,1137 M, mean age 65.55 ± 8.70), collected on two scanners (1,724 Skyra/1,325 Vida). The dataset included 2,238 cognitively normal (CN) participants, 612 with mild cognitive impairment (MCI), and 199 with dementia. Using amyloid-sensitive positron emission tomography (PET), global brain amyloid-$\beta$ burden was derived from $\left(\hat{1}8\text{F}\right)$ florbetaben SUVR values (FBB_CL) normalized by the whole cerebellum as the reference region and converted to centiloid units for 2,044 participants (1,487 CN/434 MCI/123 with Dementia). The centiloid scale standardizes amyloid PET measurements, where 0 represents the mean signal in young amyloid-negative controls and higher values represents that of typical Alzheimer’s patients, enabling direct comparison across tracers and analysis methods[11]. Details of dMRI preprocessing may be found in [3]. We generated 87 WM bundles using a bundle tractography method [12] with a modified version of the population-averaged HCP-1065 Young Adult atlas [13] as reference. We discarded 3 bundles (the left parahippocampal parietal cingulum and bilateral medial forebrain bundles) as they had less than 30% of subjects with successful reconstructions. We used BUndle ANalytics (BUAN) [9] to derive along-tract microstructural profiles using a weighted mean approach[14], where for each subject’s bundle, each point on a streamline is assigned to a segment based on the closest centroid from the atlas bundle, and DTI measures projected onto these points are weighted by their distance to the atlas bundle centroid and averaged within each segment. For each bundle, we computed along-tract profiles with 4 along-tract resolutions at 5-mm, 10-mm, 25-mm and 50-mm segment length $\left(l\right)$, where the number of segments $\left(S_l\right)$ was calculated as the average streamline length of the atlas bundle divided by $l$. Bundles shorter than 50-mm remain as one segment. The total number of segments across all 84 tracts $\left(S\right)$ for each $l$ were 1619, 791, 310, and 150, respectively.

2.2. Tractometry Bootstrapping

In our first analysis, we assessed diagnostic group (DX) differences in 4 DTI (DTI) metrics —FA, MD, RD, and AxD—using a linear mixed-effects model (LMM) fit at each bundle segment. These are all measures that show associations with dementia in both along-tract and region of interest analyses[15, 2]. CN participants were labeled as controls, and, for the purposes of this analysis, cognitively impaired participants (those with MCI or dementia) were labeled as cases. The model was defined as

where ${\text{M}}_{i,j}$ is the DTI metric for subject $i$ at segment $j$, ${\beta }_0$ is the intercept, ${\beta }_{\mathrm{1..3}}$ are the fixed effect, $u_{\text{Protocol}\left(\text{i}\right)}$ is the random intercept for MRI protocol, ${ϵ}_{i,j}$ is the residual error. In the second analysis, we modeled DTI metrics as a function of amyloid burden, quantified by Centiloid (FBB_CL), among CN participants (N=1487):

For both analyses, bootstrap resampling was used to evaluate the stability and sensitivity of along-tract effects across varying experiment conditions. Specifically, for each sample size $S$ and segment length $l$, subjects were sampled with replacement for 500 bootstrap iterations. In the case-control analysis, $S$ subjects per group were sampled, where $S=100,200,\dots ,800$. In the amyloid analysis, $S$ CN subjects with available amyloid measures are sampled where $S=100,200,\dots ,1200$. After fitting the LMM, multiple comparisons across all segments and bundles were corrected using the Benjamini–Hochberg FDR procedure $\left(q<0.05\right)$ [6].

In this procedure, all segment-level $p$-values across bundles are first sorted in ascending order:

For a target false discovery rate $q=0.05$, the Benjamini–Hochberg threshold is defined as the largest $p_{\left(k\right)}$ satisfying

The global FDR proportion was then defined as the fraction of all segments with $p_{\left(i\right)}\le p_{\left(k\right)}$, representing the proportion of bundle segments surviving global FDR correction.

The proportion of bundle segments surviving global FDR correction, or global FDR proportion, was computed for each bootstrap, sample size, and segment length. For each tract segment, standardized effect sizes, or partial d, were computed as the ratio of the fixed effect coefficient to the residual standard deviation, which quantifies the standardized magnitude associated with the predictor conditional on age, sex, and protocol. At the bundle level, we computed bundle-wise FDR proportion, the mean and max partial $d$ across all segments within a bundle. These metrics summarize the spatial extent and magnitude of along-tract effects, which enables comparing sensitivity patterns across bundles. We note that bundle-level and global FDR are not always reported in tractometry publications, but we report both for completeness.