MARBLE: An in vivo MRI marker of limbic‐predominant age‐related TDP‐43 encephalopathy neuropathologic change (LATE‐NC) Authors: Mahir Tazwar, Arnold M Evia, Abdur R Ridwan, Sue E Leurgans, Maude Wagner, Alzheimer's Disease Neuroimaging Initiative, David A Bennett, Julie A Schneider, Konstantinos Arfanakis Published: 2026 ABSTRACT AbstractINTRODUCTIONLimbic‐predominant age‐related transactive response DNA binding protein 43 kDa encephalopathy neuropathologic change (LATE‐NC) is common in older adults, yet definitive diagnosis is only possible at autopsy.METHODSThis work developed a fully automated in vivo marker of LATE‐NC named MARBLE (MARker of Brain LatE) by applying machine learning methods on brain magnetic resonance imaging (MRI) and neuropathology data from 918 community‐based older adults.RESULTSTesting MARBLE in an independent group of older adults, with ante mortem intervals between MRI and death of < 3 years, achieved an area under the curve = 0.92 for in vivo prediction of LATE‐NC. Moreover, a higher MARBLE score was associated with higher LATE‐NC stages (β = 0.083, 95% confidence interval = [0.053, 0.113], p < 10−4), and a lower level and accelerated decline of cognition, with the strongest effects observed in episodic and semantic memory, and perceptual speed (p < 10−4).DISCUSSIONMARBLE is an automated in vivo MRI‐based marker of LATE‐NC holding significant promise for enhancing diagnosis, refining clinical trial enrollment, and advancing targeted therapeutic approaches.Keywords: autopsy, biomarker, cognitive decline, community cohort, diagnosis, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change, machine learning, magnetic resonance imaging, neuroimaging, neuropathology, transactive response DNA binding protein 43 BACKGROUND Limbic‐predominant age‐related transactive response DNA binding protein 43 (TDP‐43) encephalopathy (LATE) is a recently defined neurodegenerative disease entity prevalent in older adults with a significant and generally under‐recognized impact on public health. 1 LATE neuropathological change (LATE‐NC) is characterized by the accumulation of abnormal TDP‐43 proteins, primarily in the amygdala, hippocampus, and other cortical areas, with or without hippocampal sclerosis. 1 , 2 LATE plays a major role in late‐life cognitive decline, often manifesting as an amnestic syndrome that closely resembles Alzheimer's disease (AD) clinical phenotypes. 1 , 3 , 4 Community‐based autopsy studies have reported LATE‐NC in > 25% of individuals > 80 years of age in absence of AD, increasing to > 50% in those with comorbid AD. 5 LATE‐NC can present as an independent pathology and is associated with a more indolent cognitive decline trajectory compared to cases with concomitant AD. 1 , 4 , 6 , 7 However, it also co‐exists with other neurodegenerative and vascular pathologies, making accurate diagnosis difficult due to shared clinical features. 5 , 8 , 9 , 10 Currently, definitive diagnosis of LATE is only possible at autopsy. 2 Given the prevalence and deleterious effects of LATE in older adults, it has become critical to develop a reliable marker that can assist in detecting this disease during life. In recent years, considerable research efforts have been directed toward identifying in vivo markers of LATE‐NC, including neuroimaging, biofluid, and other markers. 11 , 12 , 13 , 14 , 15 , 16 , 17 At the molecular level, loss of normal nuclear TDP‐43 function, a hallmark of TDP‐43 proteinopathies, has been linked to the encoding of cryptic exons in specific genes, leading to the translation of novel protein isoforms such as the neoepitope within hepatoma‑derived growth factor‑related protein 2 (HDGFL2), which have emerged as candidate molecular markers of LATE‐NC in biofluids and tissues. 14 , 18 , 19 Among neuroimaging markers, studies using fluorodeoxyglucose positron emission tomography (FDG PET) have revealed a temporo‐limbic hypometabolism pattern, particularly inferior‐to‐medial temporal (IMT) hypometabolism ratio and medial temporal lobe hypometabolism, that can help differentiate between LATE‐NC and AD neuropathologic change (ADNC). 11 , 13 Efforts to develop TDP‑43–specific PET tracers are advancing, and candidate tracers now show selective affinity for TDP‐43 pathology in autoradiography studies. 17 Magnetic resonance imaging (MRI) studies have similarly reported pronounced hippocampal atrophy, higher inferior middle temporal to hippocampal volume ratios, reduced medial temporal lobe volumes, lower transverse relaxation rates, and altered diffusion metrics as potential predictors of LATE‐NC. 6 , 12 , 16 , 20 , 21 , 22 , 23 , 24 Some of the above findings have recently informed the proposal of probabilistic clinical diagnostic criteria for LATE, integrating age, cognitive profile, and neuroimaging features. 25 , 26 Despite these significant strides, a marker of LATE‐NC with satisfactory diagnostic accuracy is not yet available, highlighting the critical need for continued innovation in this space. 25 , 27 , 28 The objective of this study was to develop and validate an in vivo marker of LATE‐NC using machine learning on multi‐modal MRI and neuropathology data from a large number of community‐based older adults. First, two machine learning classifier models were separately trained using ex vivo or in vivo MRI and pathology data from longitudinal cohort studies of aging conducted at Rush University Medical Center. Each classifier's performance in predicting LATE‐NC was evaluated through nested cross‐validation. The ex vivo classifier was additionally translated to in vivo using a subset of the training data with available in vivo and ex vivo MRI on the same persons. Predictions from the in vivo classifier and the translated ex vivo classifier were ensembled to form the final LATE‐NC prediction score. Next, these classifiers were packaged into a fully automated software container named MARBLE (MARker of Brain LatE), which takes as input raw in vivo MRI data, performs all necessary preprocessing, extracts features, and enters them into the classifiers to generate a score reflecting the likelihood that a person suffers from LATE‐NC. MARBLE was validated for in vivo prediction of LATE‐NC in external data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) and National Alzheimer's Coordinating Center (NACC). Finally, the association of baseline MARBLE scores with the baseline level and decline of cognition was examined in older adults who were not included in training or testing. METHODS Study participants RESEARCH IN CONTEXT Systematic review: The authors performed a literature review using PubMed, Google Scholar, and recent conference proceedings focused on in vivo markers of limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change (LATE‐NC). No study to date has developed and validated an in vivo magnetic resonance imaging (MRI)‐based LATE‐NC marker by combining MRI and detailed pathology data from large community‐based cohorts. Systematic review: The authors performed a literature review using PubMed, Google Scholar, and recent conference proceedings focused on in vivo markers of limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change (LATE‐NC). No study to date has developed and validated an in vivo magnetic resonance imaging (MRI)‐based LATE‐NC marker by combining MRI and detailed pathology data from large community‐based cohorts. Interpretation: This work introduced MARker of Brain LatE (MARBLE), a novel, fully automated in vivo marker of LATE‐NC that is based solely on MRI data. MARBLE offers a non‐invasive, user‐friendly solution for predicting LATE‐NC in vivo, and may have implications in enhancing diagnosis; refining participant selection in clinical trials for LATE, Alzheimer's disease, and other dementias; as well as in advancing targeted disease‐specific therapeutic approaches. Interpretation: This work introduced MARker of Brain LatE (MARBLE), a novel, fully automated in vivo marker of LATE‐NC that is based solely on MRI data. MARBLE offers a non‐invasive, user‐friendly solution for predicting LATE‐NC in vivo, and may have implications in enhancing diagnosis; refining participant selection in clinical trials for LATE, Alzheimer's disease, and other dementias; as well as in advancing targeted disease‐specific therapeutic approaches. Future directions: Future studies will validate MARBLE in independent cohorts to further evaluate its generalizability and clinical utility. Future directions: Future studies will validate MARBLE in independent cohorts to further evaluate its generalizability and clinical utility. Older adults participating in longitudinal studies of aging were included in this work, specifically from the Rush Memory and Aging Project (MAP), Religious Orders Study (ROS), 29 Minority Aging Research Study (MARS), 30 African American Clinical Core (AACC), 31 and Latino Core (LATC) of the Rush Alzheimer's Disease Research Center (ADRC); 32 ADNI (https://adni.loni.usc.edu); and NACC (https://naccdata.org). The institutional review board of Rush University Medical Center approved MAP, ROS, MARS, AACC, and LATC. The five studies enroll community‐based older adults without known dementia. Participants provided written informed consent, with all MAP, ROS, and a subset of MARS, AACC, and LATC participants also signing an anatomical gift act agreeing to post mortem brain donation. ADNI and NACC collectively provide a rich repository of de‐identified clinical, neuropsychological, neuropathological, and imaging data from multiple ADRCs across the United States. Ethical approval was obtained by the ADNI and NACC investigators from the institutional review boards at each participating site, and all participants and/or their authorized representatives provided written informed consent to use their data for research purposes. Rush MAP, ROS, MARS, AACC, and LATC participants were included in training, translation, and testing of MARBLE. More specifically, 850 Rush participants who (1) had undergone autopsy and detailed neuropathologic examination; (2) had available pathology data that met quality standards; (3) had available ex vivo MRI data that passed quality checks; and (4) did not have amyotrophic lateral sclerosis (ALS), frontotemporal lobar degeneration (FTLD), tumors, large hemorrhages, or large infarcts, were included in training a LATE‐NC classifier, and will be referred to as the “ex vivo classifier training group,” or “ECT group” (Figure 1A). A subset (n = 40) of the ECT group also having in vivo in addition to ex vivo MRI data, and an ante mortem interval (AMI; time between in vivo MRI data acquisition and death) of < 3 years, was used for translation of the ex vivo classifier to in vivo, and will be referred to as the “translation group” (Figure 1A). Another group of 188 Rush participants with available in vivo MRI and pathology data that passed quality checks, an AMI < 5 years, and the same exclusions as above, was used in training a separate LATE‐NC classifier, and will be referred to as the “in vivo classifier training group,” or “ICT group” (Figure 1A). Finally, 934 Rush participants who (1) had completed baseline clinical evaluation and more than one cognitive function testing session, (2) had in vivo MRI data that passed quality checks, and (3) were not included in the training of the two classifier models (i.e., not part of ECT or ICT), were included in testing the hypothesis that a higher MARBLE score is associated with a lower baseline level and accelerated decline of cognition, and will be referred to as the “cognitive testing group” (Figure 1A). Venn diagram illustrating participant groups. A, Rush participants in the ECT (pink‐yellow) and ICT (yellow‐blue) groups had available ex vivo or in vivo MRI and pathology data and were included in the development of the ex vivo and in vivo classifiers, respectively. The translation group (yellow) had both in vivo and ex vivo MRI data on the same Rush participants and was used to translate the ex vivo classifier to in vivo. Rush participants in the cognitive testing group were included in testing the hypothesis that a higher MARBLE score is associated with a lower baseline level and accelerated decline of cognition. There was no overlap between the cognitive testing group and the ECT or ICT groups. B, ADNI and NACC participants in the pathology testing group (light red) had available in vivo MRI and pathology data, and were included in validation of MARBLE. ADNI, Alzheimer's Disease Neuroimaging Initiative; ECT, ex vivo classifier training; ICT, in vivo classifier training; MARBLE, MARker of Brain LatE; MRI, magnetic resonance imaging; NACC, National Alzheimer's Coordinating Center. ADNI and NACC participants were included in validation of MARBLE and will be referred to as the “pathology testing group.” Specifically, this group included 17 ADNI (as of November 25, 2024) and 86 NACC (as of December 2024 data freeze) participants (Figure 1B) who (1) had available data on LATE‐NC; (2) had in vivo MRI data that passed quality checks; (3) had diffusion‐weighted MRI data in > 26 diffusion gradient directions and slice thickness < 3 mm; (4) did not have TDP‐43 pathology in the spinal cord, FTLD, tumors, large hemorrhages, or large infarcts; and (5) were one of at least 5 participants imaged on the same MRI scanner (important for data harmonization across scanners). Clinical evaluation and cognitive function testing Rush MAP, ROS, MARS, AACC, and LATC participants underwent annual clinical evaluations, including uniform neurological examination, detailed cognitive assessments, and medical history interviews. 29 , 30 Each participant completed 21 neuropsychological tests. Of these, 19 were used to evaluate performance in five cognitive domains: episodic memory, semantic memory, working memory, perceptual speed, and visuospatial ability, 29 , 30 , 33 while the Mini‐Mental State Examination (MMSE) and complex ideation material informed clinical impressions and were used in descriptive statistics. Composite measures of global cognition and each of the five cognitive domains were calculated by converting individual test scores to z scores using the mean and standard deviation at baseline visit and then averaging these z scores, as previously described. 4 , 33 Clinical diagnosis of mild cognitive impairment (MCI) and dementia was determined by an experienced clinician based on National Institute of Neurologic and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association (NINCDS‐ADRDA) criteria, 34 after reviewing the cognitive and clinical findings. 33 Participants without dementia or MCI were categorized as no cognitive impairment (NCI). ADNI participants completed well‐established neuropsychological test batteries, including MMSE, and were classified into three clinical groups (NCI, MCI, AD) according to the ADNI Clinical Core diagnostic protocol (adni.loni.usc.edu). 35 NACC participants underwent clinical, neurological, and neuropsychological evaluations following the standardized Uniform Data Set collection protocol. 36 Clinical diagnosis of NCI, MCI, or AD for NACC participants was determined by consensus of a multidisciplinary team at each participating site. 37 MRI data In vivo MRI data acquisition, processing, and feature extraction Longitudinal in vivo 3 Tesla (T) MRI data were collected on the Rush MAP, ROS, MARS, AACC, and LATC participants included in the ICT, translation, and cognitive testing groups, as well as on the ADNI/NACC participants in the pathology testing group. Three‐dimensional (3D) T1‐weighted (T1w) magnetization prepared rapid acquisition gradient echo (MPRAGE) and diffusion‐weighted imaging (DWI) data were collected using the parameters shown in Table SA1 in supporting information for Rush participants and in Tables SA2–SA3 in supporting information for ADNI/NACC participants. For training the in vivo classifier, translating the ex vivo classifier, and testing the performance of MARBLE in predicting LATE‐NC, we used in vivo data proximate to death. For testing the association of MARBLE score with subsequent cognitive trajectories, we used baseline MRI data (Figure 1). Raw T1w images were corrected for bias field inhomogeneity, were skull stripped using HD‐BET, 38 and segmented into gray matter (GM), white matter (WM), and intracranial space using CAT12. 39 The total intracranial volume (ICV) was calculated. Then the brain was divided into lobes and GM was segmented into multiple regions, using an in‐house multi‐atlas segmentation pipeline based on 30 atlases with lobar and Desikan–Killiany labels generated via FreeSurfer with manual corrections. 40 Lobar GM and WM volumes, and regional GM volumes were calculated. Moreover, local brain morphometry was assessed using deformation‐based morphometry (DBM). 41 Specifically, skull‐stripped T1w images were registered to the MIITRA T1w template using ANTs non‐linear registration, 42 , 43 and maps of the logarithm of the Jacobian determinant (LogJ) of the deformation fields were produced and subsequently smoothed using a Gaussian filter with a 4 mm full width at half maximum (FWHM). Voxel‐wise linear regression was then used to identify clusters of voxels showing negative association between LogJ and LATE‐NC, controlling for all other neuropathologies (see section 2.4), age, sex, years of education, scanner, and AMI between in vivo MRI and death. Regions showing the strongest negative association of LogJ with LATE‐NC were identified and median LogJ values were extracted from these regions. The DWI data were processed using TORTOISE, 44 which included eddy‐current distortion correction, bulk motion correction, B‐matrix reorientation, calculation of diffusion tensors, and generation of fractional anisotropy (FA) and trace maps. These FA and trace maps were then transformed to the space of the IIT Human Brain Atlas v.5.0 using FA‐based non‐linear registration and were projected onto the WM skeleton of the IIT atlas. 45 , 46 Median FA and trace values were extracted from regions defined by the intersection of 8 × 8 × 8 mm3 non‐overlapping cubes with the WM skeleton. Median FA and trace values were also extracted from each WM connection of the structural connectome of the IIT atlas, after thresholding the track density image (TDI) of each connection at 5% of the maximum streamline count per TDI. 46 , 47 Furthermore, voxel‐wise linear regression was used to identify clusters of voxels showing negative association between FA and LATE‐NC, or positive association between trace and LATE‐NC, controlling for all other neuropathologies (see section 2.4), age, sex, years of education, white matter hyperintensities (segmented in other studies 48 ), scanner, and AMI between in vivo MRI and death. Regions showing the strongest association were identified and median FA and trace values were extracted from these regions. Ex vivo MRI data acquisition, processing, and feature extraction When Rush participants came to autopsy, the brain was removed and the cerebrum divided into hemispheres. The hemisphere with more visible pathology was selected for ex vivo MRI and pathologic examination; in the absence of visible pathology, the selection was made non‐systematically. The selected hemisphere was immersed in phosphate‐buffered 4% formaldehyde solution 4 , 49 and imaged ex vivo on a 3T scanner with a spin‐echo echo‐planar DWI sequence within 24 hours post mortem, and with a multi‐echo spin‐echo (SE) sequence ≈ 30 days post mortem, using the parameters shown in Table SA4 in supporting information. T2‐weighted (T2w) ex vivo SE images collected with an echo time between 49.5 and 55 ms were bias‐field corrected and segmented into GM, WM, different lobes, and multiple GM regions using an in‐house multi‐atlas segmentation pipeline based on the same atlases used in vivo. 40 , 50 Lobar and regional volumes were calculated. DBM was used to assess local brain morphometry by registering the T2w images to an ex vivo brain hemisphere template, generating LogJ maps from the resulting deformation fields, and smoothing the LogJ maps with a 4 mm FWHM Gaussian filter. 41 , 51 Voxel‐wise linear regression was used to identify clusters of voxels showing negative association between LogJ and LATE‐NC, controlling for all other neuropathologies (see section 2.4), age, sex, years of education, scanner, post mortem interval to immersion in fixative, and post mortem interval to ex vivo MRI. Regions showing the strongest negative association of LogJ with LATE‐NC were identified and median LogJ values were extracted. The ex vivo DWI data were processed using TORTOISE similar to the in vivo data. 44 FA maps were generated, were non‐linearly registered to an ex vivo brain hemisphere FA template using ANTs, 42 and projected onto the WM skeleton. 45 Similar to what was done for the in vivo data, median FA values were extracted from regions defined by the intersection of small cubes with the WM skeleton, from each WM connection of the structural connectome of the IIT atlas transformed to the ex vivo hemisphere template, and also from regions showing the strongest negative association of FA with LATE‐NC in voxel‐wise linear regression controlling for all other neuropathologies (see section 2.4), age, sex, years of education, white matter hyperintensities (segmented in other studies 52 ), scanner, post mortem interval to immersion in fixative, and post mortem interval to ex vivo MRI. In vivo/ex vivo MRI feature translation Before a classifier trained ex vivo can be used in vivo, translation between in vivo and ex vivo MRI features is necessary, because brain MRI characteristics are different ex vivo compared to in vivo due to the effects of fixation and the required differences in scanning parameters. 50 Translation models were developed for the MRI features extracted in section 2.3.2 using data from the translation group (Figure 1). Regions identified in section 2.3.2 based on voxel‐wise analyses were spatially transformed to in vivo image space (in the same hemisphere as ex vivo for each participant separately) and corresponding in vivo values were extracted. Extraction of values for other regions that are segmented in both sections 2.3.1 and 2.3.2 was described in those sections. Linear mixed regression was then applied for the different MRI features to map in vivo values to the corresponding ex vivo values, after correction for scanner, post mortem interval to fixation, and post mortem interval to imaging. The resulting translation models were used to translate in vivo MRI features to ex vivo before using them in a classifier trained ex vivo (see sections 2.5, 2.6). Neuropathologic examination Cerebral hemispheres from Rush MAP, ROS, MARS, AACC, and LATC participants that came to autopsy underwent detailed neuropathologic examination overseen by a board‐certified or expert neuropathologist following uniform standard procedures. 4 , 53 LATE‐NC was assessed using immunohistochemistry with a phosphorylated monoclonal TAR5P‐1D3 (pS409/410; 1:100, Ascenion) TDP‐43 antibody to detect TDP‐43 cytoplasmic inclusions in neurons and glia in eight brain regions, involving limbic (amygdala, entorhinal cortex, hippocampus CA1, dentate gyrus) and neocortical (temporal pole, middle temporal cortex, orbitofrontal cortex, middle frontal cortex) structures. 54 Based on the presence of TDP‐43 inclusions for each region, LATE‐NC was graded into four stages: stage 0 (no TDP‐43 inclusions), stage 1 (TDP‐43 inclusions in amygdala only), stage 2 (TDP‐43 inclusions extend to entorhinal cortex and/or hippocampus CA1), and stage 3 (TDP‐43 inclusions extend to neocortex). 1 Furthermore, ADNC was diagnosed using the National Institute on Aging–Alzheimer's Association (NIA‐AA) criteria. 55 Pathological diagnoses of other neurodegenerative (including Lewy bodies) and cerebrovascular (including cerebral amyloid angiopathy, gross and microscopic infarcts, atherosclerosis, arteriolosclerosis) diseases were determined using established procedures as described in previous studies. 4 , 49 , 53 Neuropathologic examination procedures in ADNI/NACC have been documented elsewhere. 56 , 57 For this work, pathology data were collected from the NACC Neuropathology Data Form, which included variables indicating the presence of TDP‐43 immunoreactive inclusions in the amygdala, hippocampus, entorhinal/inferior temporal cortex, and neocortex (specifically midfrontal cortex). These variables were categorized to form the same 4‐stage LATE‐NC staging scheme described above. Development of LATE‐NC classifiers Data from Rush MAP, ROS, MARS, AACC, and LATC participants in the ECT and ICT groups (see section 2.1; Figure 1) were used to train ex vivo and in vivo classifier models that differentiate LATE‐NC stages 2 and 3 from stages 0 and 1. Stage 1 was grouped together with stage 0 because, according to prior LATE MRI and clinico‐pathologic studies, stage 2 was the earliest stage associated with measurable brain abnormalities in MRI, cognitive impairment, and increased dementia risk. 21 , 52 , 54 Preprocessing steps for the ex vivo classifier involved regressing out of the ex vivo MRI features (listed in section 2.3.2) the confounding effects of post mortem interval to fixation, post mortem interval to imaging, and scanners. All features were then standard normalized to have zero mean and unit variance. To reduce dimensionality and remove redundant features, a recursive feature elimination (RFE) method with a random forest or extreme gradient boosting (XGBoost) classifier was used to select a subset of the most relevant variables within each MRI feature type (the three types were volumetric, morphometric, and diffusion‐based features). Next, a model was trained based on different combinations of MRI features using a two‐level stacking classifier method (Figure 2A). At level 1, an XGBoost classifier (base estimator) generated risk scores from a single MRI feature type. At level 2, an L2‐regularized logistic regression classifier produced the final LATE‐NC prediction based on the aggregated risk scores from all feature types (Figure 2A). Model performance was evaluated using 100 repeats of stratified shuffle split cross‐validation (CV), randomly subsampling 80% of the total data for training and 20% for testing in each repeat (outer CV loop), thereby preserving the proportion of LATE‐NC positive and negative cases in each split. Hyperparameters of the final estimator were tuned using a 10‐fold CV in the training set of each repeat (inner CV loop), and balanced class weights were applied during classifier training to account for class imbalance. The average area under the receiver operating characteristic curve (AUC) across 100 repeats was calculated on the test set for each model. Pairwise model comparisons were performed using two‐sided Wilcoxon signed rank test. The model having significantly higher average AUC was selected as the final model and was retrained on all available data from the ECT group. A, Schematic of the two‐level ex vivo classifier set‐up. The final estimator (logistic regression classifier) in level 2 was trained on the cross‐validated predictions (risk scores) obtained from each of the base estimators (XGBoost classifier) or level 1 classifiers. A similar set‐up was used for training the in vivo classifier. B, Overview of MARBLE. The ex vivo and in vivo classifiers were combined in MARBLE. DBM, deformation‐based morphometry; DTI, diffusion tensor imaging; DWI, diffusion‐weighted images; LATE‐NC, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change; RS, risk score; T1w, T1‐weighted; XGBoost, extreme gradient boosting. For the in vivo classifier, preprocessing steps included: regressing out of the in vivo MRI features (listed in section 2.3.1) the confounding effects of AMI and scanner, standard normalization of features, and RFE feature selection. Model training and evaluation was conducted on data from the ICT group and mirrored the process followed for the ex vivo classifier. Machine learning models were implemented using Python 3.11 with Scikit‐learn v1.7.1 and XGBoost v3.0.4 libraries. Development of MARBLE The two classifiers of LATE‐NC trained ex vivo and in vivo were combined in a fully automated software container named MARBLE. MARBLE (1) takes as input raw in vivo T1w and DWI data (NIfTI or DICOM format), (2) performs all necessary image processing steps and extracts MRI features as described in section 2.3, (3) averages across hemispheres and translates to ex vivo a subset of these features for use in the ex vivo classifier (while features from the two hemispheres are used without averaging in the in vivo classifier), and finally (4) applies both the in vivo and ex vivo classifier models and averages the two classifier predictions to provide an ensemble diagnostic score, the MARBLE score, describing the likelihood that an individual suffers from LATE‐NC (Figure 2B). MARBLE is a Singularity container, enabling a consistent and reproducible runtime environment. In vivo validation of MARBLE: prediction of LATE‐NC MARBLE's performance in predicting in vivo LATE‐NC stage 2 or 3 was evaluated in ADNI/NACC data from the pathology testing group (see section 2.1; Figure 1). To mitigate scanner and protocol‐related variations across ADNI, NACC, and Rush in vivo MRI data, ComBat was used for harmonization, using data from 50 Rush participants not included in training as reference batch. 58 , 59 Performance was assessed by calculating the AUC. Furthermore, because longer AMI between in vivo MRI and death increases the chance of additional pathology developing after in vivo MRI, which can lower the estimated performance of MARBLE in predicting pathology, the AUC was also calculated for portions of the pathology testing group with different ranges of AMI. Additionally, linear regression was used to test the association of MARBLE score (dependent variable) with pathologically assessed LATE‐NC stages (independent variable), controlling for AMI; and in a separate analysis also controlling for ADNC severity. In vivo validation of MARBLE: associations with cognitive trajectories In the cognitive testing group (see section 2.1; Figure 1), we tested the hypothesis that higher baseline MARBLE scores were associated with lower initial levels and accelerated changes in global cognition over time, using linear mixed‐effects models. In this analysis, we considered all cognitive assessments at and after baseline MRI. Changes in global cognition were modeled using natural cubic splines with two inner knots located at the tertiles of measurement times; the non‐linear trajectories approximated by splines provided a substantially better fit than linear trajectories (e.g., Akaike information criteria = 4521.5 [splines] and = 5325.9 [linear], respectively). Covariates included age at baseline MRI, sex, years of education, scanner (both as simple effect and in interaction with splines functions of time), as well as a time‐varying term for mode of cognitive administration (generally in person, however, during COVID, via telephone); interindividual variability was captured by correlated random intercept and slopes. The same modeling approach was repeated for each of the five cognitive domains: episodic memory, semantic memory, working memory, perceptual speed, and visuospatial ability. These analyses were conducted using R software version 4.0.3 (R Foundation for Statistical Computing), and mixed models were estimated using the hlme function of lcmm R package version 2.2.1. 60 Other statistical analyses To compare demographic, clinical, and neuropathological characteristics between the classifier training groups (ECT or ICT) and the testing group (ADNI/NACC), Mann‐Whitney U tests were used for continuous/ordinal variables and chi‐squared/Fisher exact tests were used for nominal variables. Moreover, differences in participant characteristics across LATE‐NC stages were further explored using Kruskal–Wallis or chi‐squared tests as appropriate. The statistical tests in this section were implemented using Python 3.11 with Scipy v1.16.1 and statsmodels v0.14.5 libraries. RESULTS Demographic, clinical, and neuropathological characteristics Demographic, clinical, and neuropathological characteristics of Rush participants in the ECT and ICT groups and ADNI/NACC participants in the pathology testing group are listed in Table 1. The 850 community‐based older adults in the ECT group were 73% female, 90% non‐Latino White, 7% non‐Latino Black, 3% Latino, and their mean age at death was 91 (standard deviation [SD] = 6) years. Proximate to death, 33% were NCI, 23% MCI, and 44% had dementia. Also, 38% of participants had LATE‐NC stage 2 or 3, and 69% had intermediate or high ADNC. TABLE 1Demographic, clinical, and neuropathological characteristics of participants of the ex vivo and in vivo classifier training groups (ECT, ICT) and pathology testing group.CharacteristicsEx vivo classifier training (ECT) groupIn vivo classifier training (ICT) groupPathology testing group (ADNI/NACC) n 850188103Age at MRI, years, mean ± SD— 87.8 ± 6.3 (p < 0.0001) 76.1 ± 8.7Age at death, years, mean ± SD 90.8 ± 6.4 (p < 0.0001) 90.5 ± 6.3 (p < 0.0001) 81.5 ± 8.9Male, n (%)233 (27.4)44 (23.4)65 (63.1)Education, years, mean ± SD15.7 ± 3.515.8 ± 3.115.7 ± 2.8At least one copy of the ε4 allele, n (%)208 (24.5)38 (20.2)54 (52.4)Race and ethnicity, n (%) p = 0.0152 Non‐Latino White766 (90.1)150 (79.8)97 (94.2)Non‐Latino Black58 (6.8)32 (17.0)4 (3.9)Latino24 (2.8)4 (2.1)1 (1.0)Clinical diagnosis, n (%) p = 0.0014 p < 0.0001 No cognitive impairment280 (32.9) a 121 (64.4) b 19 (18.4) b Mild cognitive impairment198 (23.3) a 50 (26.6) b 24 (23.3) b Dementia371 (43.6) a 16 (8.5) b 60 (58.3) b MMSE, mean ± SD 20.0 ± 9.6 a (p = 0.0130) 27.2 ± 2.3 b (p < 0.0001) 23.7 ± 5.4 b Ante mortem interval between in vivo MRI and death, years, median (range)— 2.6 (0.2−5.0) (p < 0.0001) 5.0 (0.3−11) Post mortem interval to fixation, hours, median (interquartile range)7.4 (4.8)7.9 (6.2)10.5 (13.7)LATE‐NC, n (%) p < 0.0001 p = 0.0157 Stage 0 (no TDP‐43 inclusions)375 (44.1)95 (50.5)67 (65.0)Stage 1 (TDP‐43 inclusions in amygdala only)155 (18.2)39 (20.7)9 (8.7)Stage 2 (extended to hippocampus CA1 and/or entorhinal cortex)90 (10.6)16 (8.5)22 (21.4)Stage 3 (extended to neocortex)230 (27.1)38 (20.2)5 (4.9)High or intermediate ADNC (NIA‐AA criteria), n (%) 586 (68.9) (p < 0.0001) 124 (66.0) (p < 0.0001) 81 (78.6)Presence of Lewy bodies, n (%)238 (28.0)47 (25.0)44 (42.7)Presence of gross infarcts, n (%)341 (40.1)81 (43.1)12 (11.7)Presence of microscopic infarcts, n (%)336 (39.5)87 (46.3)24 (23.3)Moderate or severe CAA, n (%)315 (37.1)68 (36.2)31 (30.1)Moderate or severe atherosclerosis, n (%) 187 (22.0) (p = 0.0069) 42 (22.3) (p = 0.0165) 31 (30.1)Moderate or severe arteriolosclerosis, n (%) 210 (24.7) (p < 0.0001) 57 (30.3) (p < 0.0001) 59 (57.3)MRI scanner, n (%)3T Siemens Trio82 (9.6)155 (82.4)21 (20.4)3T Philips Achieva287 (33.8)33 (17.6)15 (14.6)3T Siemens Verio481 (56.6)—6 (5.8)3T GE Discovery——61 (59.2) Notes: Comparisons between the ECT or ICT groups and the pathology testing group were performed using Mann–Whitney U test for continuous/ordinal variables and chi‐squared/Fisher exact test for nominal variables. All statistical tests were two sided. Individuals with missing data were excluded from calculations. Significant p values are shown in bold font.Abbreviations: ADNC, Alzheimer's disease neuropathological change; ADNI, Alzheimer's Disease Neuroimaging Initiative; CAA, cerebral amyloid angiopathy; ECT, ex vivo classifier training group; ICT, in vivo classifier training group; LATE‐NC, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change; MMSE, Mini‐Mental State Examination; MRI, magnetic resonance imaging; NACC, National Alzheimer's Coordinating Center; NIA‐AA, National Institute on Aging and Alzheimer's Association; SD, standard deviation; TDP‐43, transactive response DNA binding protein 43. a Proximate to death. b Proximate to MRI. Demographic, clinical, and neuropathological characteristics of participants of the ex vivo and in vivo classifier training groups (ECT, ICT) and pathology testing group. 87.8 ± 6.3 (p < 0.0001) 90.8 ± 6.4 (p < 0.0001) 90.5 ± 6.3 (p < 0.0001) 20.0 ± 9.6 a (p = 0.0130) 27.2 ± 2.3 b (p < 0.0001) 2.6 (0.2−5.0) (p < 0.0001) 586 (68.9) (p < 0.0001) 124 (66.0) (p < 0.0001) 187 (22.0) (p = 0.0069) 42 (22.3) (p = 0.0165) 210 (24.7) (p < 0.0001) 57 (30.3) (p < 0.0001) Notes: Comparisons between the ECT or ICT groups and the pathology testing group were performed using Mann–Whitney U test for continuous/ordinal variables and chi‐squared/Fisher exact test for nominal variables. All statistical tests were two sided. Individuals with missing data were excluded from calculations. Significant p values are shown in bold font. Abbreviations: ADNC, Alzheimer's disease neuropathological change; ADNI, Alzheimer's Disease Neuroimaging Initiative; CAA, cerebral amyloid angiopathy; ECT, ex vivo classifier training group; ICT, in vivo classifier training group; LATE‐NC, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change; MMSE, Mini‐Mental State Examination; MRI, magnetic resonance imaging; NACC, National Alzheimer's Coordinating Center; NIA‐AA, National Institute on Aging and Alzheimer's Association; SD, standard deviation; TDP‐43, transactive response DNA binding protein 43. Proximate to death. Proximate to MRI. The 188 community‐based older adults in the ICT group were 77% female, 80% non‐Latino White, 17% non‐Latino Black, 2% Latino, and their mean age at MRI proximate to death was 88 (SD = 6) years, while their mean age at death was 91 (SD = 6) years (Table 1). Proximate to the time of MRI, 64% were NCI, 27% had MCI, and 9% had dementia. Also, 29% had LATE‐NC stage 2 or 3, and 66% had intermediate or high ADNC. The 103 individuals in the ADNI/NACC pathology testing group were 37% female, 94% non‐Latino White, 4% non‐Latino Black, 1% Latino, and their mean age at MRI proximate to death was 76 (SD = 9) years, while their mean age at death was 82 (SD = 9) years (Table 1). Proximate to the time of MRI, 18% were NCI, 23% had MCI, and 58% had dementia. Also, 26% had LATE‐NC stage 2 or 3, and 79% had intermediate or high ADNC. Compared to the ECT group, the pathology testing group had twice the proportion of male participants, was on average 9 years younger at the time of death, included more individuals with dementia and fewer with NCI, included fewer individuals with LATE‐NC stage 2 or 3 but more people with intermediate or high ADNC and moderate or severe atherosclerosis and arteriolosclerosis (Table 1). Compared to the ICT group, the pathology testing group had almost three times the proportion of male participants and fewer non‐Latino Black participants, was on average 12 years younger at the time of MRI proximate to death and 9 years younger at the time of death, included 7 times the proportion of individuals with dementia, had an AMI that was ≈ 3 years longer, included more people with intermediate or high ADNC and moderate or severe atherosclerosis and arteriolosclerosis. Detailed participant characteristics per LATE‐NC stage are provided in Tables SB1−SB3 in supporting information. Participant characteristics of the cognitive testing group are shown in Table 2. Of the 934 community‐dwelling individuals, 80% were female, 53% non‐Latino White, 36% non‐Latino Black, 11% Latino, and their mean age at baseline MRI was 78 (SD = 7) years. At the time of MRI, 82% of participants were NCI, 16% MCI, and 1% had dementia. The mean clinical follow‐up time from baseline MRI through the last visit was 7 years (SD = 3), with a mean of 8 (SD = 3) cognitive assessments over time. TABLE 2Demographic and clinical characteristics of the cognitive testing group at baseline MRI (n = 934).CharacteristicsAge a , years, mean ± SD77.5 ± 6.9Male, n (%)192 (20.6)Education, years, mean ± SD15.6 ± 3.6At least one copy of the ε4 allele, n (%)167 (17.9)Race and ethnicity, n (%)Non‐Latino White492 (52.7)Non‐Latino Black334 (35.8)Latino100 (10.7)Clinical diagnosis a , n (%)NCI979 (82.2)MCI194 (16.3)Dementia16 (1.3)MMSE a , mean ± SD28.2 ± 1.9Global cognition score a , mean ± SD0.22 ± 0.56Episodic memory score a , mean ± SD0.31 ± 0.68Semantic memory score a , mean ± SD0.24 ± 0.75Working memory score a , mean ± SD0.13 ± 0.76Perceptual speed score a , mean ± SD0.21 ± 0.75Visuospatial ability score a , mean ± SD0.19 ± 0.78Scanner for in vivo MRI, n (%)3T Siemens Trio530 (56.7)3T Philips Achieva404 (43.3)Abbreviations: MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; MRI, magnetic resonance imaging; NCI, no cognitive impairment; SD, standard deviation. a At the time of baseline MRI. Demographic and clinical characteristics of the cognitive testing group at baseline MRI (n = 934). Abbreviations: MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; MRI, magnetic resonance imaging; NCI, no cognitive impairment; SD, standard deviation. a At the time of baseline MRI. Evaluation of ex vivo and in vivo classifiers of LATE‐NC The top‐performing ex vivo classifier model was based on the following features: (1) normalized volumes (temporal, frontal, parietal and subcortical GM volumes divided by total GM volume, and temporal GM volume divided by total hemisphere volume), (2) FA values from three WM regions (located in the anterior temporal lobe, orbitofrontal, and cingulate gyrus), (3) DBM values from five GM regions (amygdala, hippocampus, entorhinal cortex, parahippocampal cortex, and nucleus accumbens) and thirteen WM regions (spanning temporal and frontal lobes and basal ganglia; Figure 3A). In the ECT group, this model achieved an average AUC of 0.85 ± 0.05 (95% confidence interval [CI = 0.75, 0.95]; balanced accuracy = 0.74 ± 0.05; sensitivity = 0.80 ± 0.10; specificity = 0.69 ± 0.10; positive predictive value [PPV] = 0.62 ± 0.07; negative predictive value [NPV] = 0.85 ± 0.06; Figure 4). This classifier model demonstrated significantly higher AUC than models trained only on T1w‐derived features (normalized volumes and DBM; AUC = 0.79 ± 0.03) or individual feature types, specifically DBM (AUC = 0.75 ± 0.03), FA (AUC = 0.76 ± 0.07), normalized volumes (AUC = 0.64 ± 0.04), or demographic data (age at death, sex; AUC = 0.62 ± 0.03; Bonferroni‐corrected p < 0.0001, two‐sided Wilcoxon signed rank test; Figure 4). Brain regions from which features were extracted and used in MARBLE. A, Features for the top‐performing classifier trained on ex vivo MRI data included normalized GM volumes, median FA values from WM regions, and median log‐Jacobian values from GM and WM regions derived from DBM. B, Features for the top‐performing classifier trained on in vivo MRI data included normalized GM and WM volumes, median FA or trace values from WM regions, and median trace values from selected WM connections (chord color indicates maximum streamline count in TDI map and node color indicates the lobe containing the GM region). The two classifiers trained ex vivo and in vivo were combined in MARBLE, and therefore, MARBLE uses all the features presented in this figure. Notably, feature values were averaged across hemispheres only for the ex vivo classifier, whereas the in vivo classifier used left and right hemisphere features separately. DBM, deformation‐based morphometry; FA, fractional anisotropy; GM, gray matter; MARBLE, MARker of Brain LatE; MRI, magnetic resonance imaging; TDI, track density image; WM, white matter. Receiver operating characteristic curves for ex vivo (left) and in vivo (right) classification of LATE‐NC in the ECT group and in the ICT group, respectively, based on different combinations of features. The black line represents the top‐performing ex vivo classifier (left) or in vivo classifier (right) across 100 cross‐validation repeats, trained using the features shown in Figure 3 with a two‐level stacking classifier architecture. Colored lines represent classifiers based on a single type of MRI features or demographics (age, sex). The top‐performing ex vivo and in vivo classifiers were combined in MARBLE. AUC, area under receiver operating characteristic curve; DBM, deformation‐based morphometry; ECT, ex vivo classifier training; FA, fractional anisotropy; ICT, in vivo classifier training; LATE‐NC, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change; MARBLE, MARker of Brain LatE; MRI, magnetic resonance imaging. The top‐performing in vivo classifier model incorporated the following features: (1) normalized volumes (15 different ratios involving lobar GM, WM, or GM plus WM volumes), (2) FA values from 22 WM regions, (3) trace values from 20 WM regions, and (4) trace values from 16 WM connections involving both commissural and association tracts (Figure 3B). It achieved an average AUC of 0.94 ± 0.05 (95% CI = [0.83, 1.0], balanced accuracy = 0.86 ± 0.08, sensitivity = 0.83 ± 0.13, specificity = 0.89 ± 0.07, PPV = 0.75 ± 0.13, NPV = 0.94 ± 0.05; Figure 4). This classifier model demonstrated significantly higher AUC than models trained only on T1w‐derived features (normalized volumes and DBM; AUC = 0.82 ± 0.07) or individual feature types, specifically FA and trace (AUC = 0.87 ± 0.06), trace from WM connections (AUC = 0.77 ± 0.08), normalized volumes (AUC = 0.74 ± 0.08), or demographics (age at MRI, sex; AUC = 0.58 ± 0.09; Bonferroni‐corrected p < 0.0001, two‐sided Wilcoxon signed rank test; Figure 4). In vivo validation of MARBLE: prediction of LATE‐NC The MARBLE software container including both the translated ex vivo classifier and the in vivo classifier of LATE‐NC was applied to the in vivo MRI scans from ADNI/NACC participants in the pathology testing group (see section 2.1; Figure 1). The resulting AUC was 0.92 for participants with shorter AMI (< 3 years, n = 22), 0.82 for those with AMI < 5 years (n = 41), while the AUC across all participants (AMI from < 1 year to 11 years) was 0.76 (n = 103; Figure 5A). An exploratory assessment of MARBLE performance in participants of the ADNI/NACC pathology testing group with intermediate/high ADNC, and separately in those with none/low ADNC can be found in Supplementary Material SC1 in supporting information. Furthermore, linear regression in the entire pathology testing group revealed a significant association of MARBLE scores with LATE‐NC stages (β = 0.083, 95% CI = [0.053, 0.113], p < 0.0001; Figure 5B). This association remained significant even after controlling for ADNC severity, with < 4% variation in the estimated β‐coefficient (p < 0.0001). In vivo MARBLE performance in the pathology testing group (ADNI/NACC). A, AUC values for in vivo prediction of LATE‐NC using MARBLE in a subgroup of individuals with AMIs less than the threshold value shown on the horizontal axis. The numbers above each circle represent the number of individuals with AMI below that threshold, and, in parentheses, the number of individuals with LATE‐NC stage 2 or 3 within that subgroup. B, In vivo MARBLE scores across different LATE‐NC stages (ground truth) assessed pathologically in the pathology testing group. The shaded inlay shows the results of linear regression examining the association of MARBLE scores with LATE‐NC stages, controlling for AMI. ADNI, Alzheimer's Disease Neuroimaging Initiative; AMI, ante mortem interval; AUC, area under receiver operating characteristic curve; β, coefficient; LATE‐NC, limbic‐predominant age‐related transactive response DNA binding protein 43 encephalopathy neuropathologic change; MARBLE, MARker of Brain LatE; NACC, National Alzheimer's Coordinating Center; SE, standard error. In vivo validation of MARBLE: association with cognitive trajectories In the cognitive testing group (see section 2.1; Figure 1), linear mixed‐effects models demonstrated that higher MARBLE scores at baseline were associated with lower initial levels of global cognition (mean difference [MD] = −0.82 units, 95% CI = [–1.08, –0.57], p < 0.0001) and each of the five cognitive domains, including episodic memory (MD = −1.05, 95% CI = [–1.37, –0.73], p < 0.0001), semantic memory (MD = −1.00, 95% CI = [–1.37, –0.63], p < 0.0001), perceptual speed (MD = −0.79, 95% CI = [–1.15, –0.43], p < 0.0001), visuospatial ability (MD = −0.46, 95% CI = [–0.84, –0.07], p = 0.02), and working memory (MD = −0.43, 95% CI = [–0.82, –0.03], p = 0.03; Figure 6). Moreover, higher MARBLE scores were associated with accelerated decline in global cognition (p ≤ 0.0001 for MARBLE × splines time function) and in all cognitive domains (all p ≤ 0.01 for MARBLE‐by‐splines; Figure 6). Among the cognitive domains, cognitive decline was particularly pronounced in episodic memory, semantic memory, and perceptual speed (all p < 0.0001 for MARBLE‐by‐splines; Figure 6). These associations remained statistically significant after Bonferroni correction, except baseline levels of visuospatial ability and working memory. Mean estimated trajectories of change in global cognition and five cognitive domains over time in the cognitive testing group (N = 934), for high and low levels of baseline MARBLE score. Estimated trajectories are derived from six separate linear mixed models, with follow‐up time approximated using natural cubic splines. Each model included the continuous MARBLE score, covariates (age at baseline MRI, sex, education, scanner, mode of cognitive assessment), and the interactions of all variables with the splines time function. Curves represent the mean estimated trajectories (solid/dashed lines) and 95% confidence intervals (shaded areas) of a typical participant profile (a woman, aged 78 years at first eligible Siemens MRI scan, and with 16 years of education). The choice of profile has no influence on the differences in trajectories estimated by the model. p‐values are derived from multivariate Wald tests. MARBLE, MARker of Brain LatE; MRI, magnetic resonance imaging. DISCUSSION This work developed and validated MARBLE, a fully automated in vivo marker of LATE‐NC, using machine learning on multi‐modal MRI and neuropathology data from > 900 community‐based older adults. MARBLE combines two classifier models trained separately using ex vivo or in vivo MRI and pathology data. MARBLE takes as input raw in vivo T1w and diffusion‐weighted MRI data, performs all necessary preprocessing, extracts features and enters them into the two classifiers (after appropriate translation for the ex vivo classifier), and generates an ensemble prediction score reflecting the likelihood that a person suffers from LATE‐NC. In vivo validation of MARBLE in an independent group of persons with in vivo MRI and pathology data (ADNI/NACC) showed excellent performance in detecting LATE‐NC (AUC = 0.92) for data collected with short AMI (< 3 years). Moreover, higher MARBLE scores in the same independent group were associated with more advanced LATE‐NC stages assessed pathologically. Furthermore, higher MARBLE scores in older adults who were largely dementia free were associated with lower baseline levels and accelerated decline of cognition, particularly with pronounced decline in episodic memory, semantic memory, and perceptual speed. MARBLE offers a non‐invasive, automated solution in predicting LATE‐NC in vivo, and holds promise for refining clinical trial enrollment, and advancing targeted therapies. The performance of MARBLE in predicting LATE‐NC in vivo was excellent, with an AUC = 0.92 in the independent ADNI/NACC pathology testing group, for participants with AMI < 3 years. Notably, this level of performance was achieved in the presence of comorbid pathologies such as AD and other neurodegenerative and vascular pathologies. In addition, the MARBLE score was associated with LATE‐NC stages assessed pathologically in the entire independent pathology testing group. Interestingly stages 1 and 3 exhibited higher median MARBLE scores than stages 0 and 2, respectively, suggesting that despite training on a binary split (stages 0 and 1 vs. stages 2 and 3), MARBLE is sensitive to incremental, stage‐wise changes across the entire disease spectrum. Beyond its predictive performance, MARBLE minimizes patient burden through its non‐invasive nature and short MRI scan duration (< 15 minutes), while its fully automated framework makes it user friendly. These findings and characteristics of MARBLE are highly encouraging in light of the prevalence and deleterious impact of LATE‐NC and the fact that a molecularly specific marker is not available. One of the potential clinical implications is that by incorporating MARBLE in the clinical criteria of LATE, we may be able to enhance the accuracy and simplify LATE diagnosis. Second, MARBLE may allow refined participant selection in clinical drug and prevention trials not only for LATE, but also for AD and related dementias where it may help exclude patients with LATE or allow studying the efficacy of treatments in the presence of LATE. Third, MARBLE provides a benchmark for in vivo prediction of LATE‐NC and may be used to support the development, testing, and validation of additional imaging‐based or biofluid‐based markers of LATE‐NC. Of note, MARBLE had an even higher performance in predicting LATE‐NC, with AUC > 0.92, in a smaller portion of the pathology testing group with AMI < 2 years, but on the other hand, a lower AUC of 0.82 in a larger portion of the pathology testing group with AMI < 5 years (Figure 5A). This demonstrates the negative impact of long AMI in testing in vivo markers of pathology (see additional details in Supplementary Material SC2), 61 as well as the great value of continuing collection of longitudinal MRI and pathology data on older adults to increase the number of datasets with short AMI. Analyses in the large cognitive testing group that was > 98% dementia free at baseline MRI demonstrated that MARBLE is also associated with clinically meaningful outcomes. Higher baseline MARBLE scores were associated with lower baseline levels and accelerated decline in global cognition and individual cognitive domains. Associations with cognitive decline were more pronounced in episodic memory, semantic memory, and perceptual speed. These findings hold particular significance, as prior clinico‐pathologic studies have documented independent associations of LATE‐NC with lower levels and faster decline in global cognition, primarily involving episodic memory, semantic memory, and perceptual speed, while sparing other domains until later stages of LATE‐NC. 4 , 26 , 54 , 62 , 63 MARBLE may therefore be used to identify individuals that are free of dementia at the time of MRI but at higher risk of cognitive decline and therefore in higher need of treatment. The integration of two LATE‐NC classifiers within MARBLE, leveraging both ex vivo and in vivo MRI training data, represents a novel approach that exploits the complementary nature of the two data sources. More specifically, the ex vivo classifier was trained on data from a large (n = 850) community‐based group of older adults spread quite evenly across clinical diagnoses of NCI, MCI, and dementia, and a broad range of neuropathological burden (Table 1, Table SB1, Figure SC1 of Supplementary Material SC in supporting information). The ex vivo classifier also benefited from several distinct advantages of ex vivo MRI, including motion‐free data, inclusion of frail individuals who are normally not included in community‐based in vivo MRI studies, and the fact that no additional pathology can develop after MRI data collection, ensuring that MRI and pathology data correspond to the same condition of the brain. The above constitute an ideal environment for training a LATE‐NC classifier. However, the ex vivo classifier required translation, and the linear mixed regression that was used may have not fully captured any non‐linear effects that may have been introduced by death and tissue fixation. Also, the ex vivo classifier considered features from only one hemisphere per person, though the ground truth of pathology was also assessed only in that same hemisphere. On the other hand, the classifier trained on in vivo MRI data leveraged brain features from both hemispheres, accounting for hemispheric asymmetry often reported in LATE‐NC, 13 , 23 , 24 and obviously did not require translation. In addition, the older adults in the in vivo classifier training group included a higher proportion of minorities and had a broad range of neuropathological burden (Table 1, Table SB2 in supporting information). However, the in vivo classifier training group was smaller (n = 188), 90% dementia free, and the AMI between in vivo MRI and death may have allowed for additional pathology to develop after MRI data collection. To mitigate the impact of AMI on training, the in vivo classifier training group was specifically designed to include only individuals with AMI < 5 years. MARBLE ensembles the predictions from the two classifiers trained ex vivo and in vivo, potentially balancing their respective trade‐offs and generating a more robust LATE‐NC score. Finally, the predictive features used in the two classifiers may capture brain MRI changes associated not only with LATE‐NC, but also with combinations of LATE‐NC and commonly co‐occurring pathologies. Leveraging the combined brain MRI effects of LATE‐NC and its common co‐pathologies reflects the complex pathophysiological landscape of the aging brain, thereby fulfilling the intended purpose of the marker. The present study has multiple strengths and also a few limitations. The strengths include: (1) training LATE‐NC classifiers based on detailed pathology data from (2) a large number of community‐based older adults with (3) a broad range of clinical diagnoses and neuropathological burden. (4) Using a two‐level stacking classifier model enabled efficient integration of multi‐modal MRI features, and (5) implementing regularization mitigated the risk of overfitting. Furthermore, the generated MARBLE software container (6) is fully automated, (7) combines two complementary LATE‐NC classifiers, (8) and offers a patient friendly approach for predicting LATE‐NC in vivo. Finally, another strength of the present work is that (9) it tested MARBLE in independent cohorts using both pathology and cognitive data. One limitation was the small sample size of the external pathology testing group, especially when considering only data with short AMI. Continued acquisition of MRI and pathology data on older adults remains imperative to facilitate robust testing of this and other emerging MRI‐based markers. Another limitation was the small proportion of non‐Latino Black and Latino individuals especially in the ex vivo training and pathology testing groups. Additional efforts are needed to increase the number of autopsies in these populations who are at higher risk for AD and related dementias. This limitation was addressed in the cognitive testing group, and the meaningful associations of MARBLE with level and decline of cognition are encouraging about its applicability in diverse populations. Another limitation was that the definition of ex vivo FA features from regions of the ex vivo FA signature of LATE‐NC, which was generated with voxel‐wise analysis in the entire ECT group, may have introduced data leakage and led to higher estimates of the ex vivo classifier CV performance. However, this was true for only one of the features used to train the ex vivo classifier, does not apply to the in vivo classifier, and most importantly does not impact the evaluation of the performance of MARBLE in the independent ADNI/NACC pathology testing dataset. Another limitation of this and other MRI‐based markers is that the scores they generate depend somewhat on the scanner and protocol used. To address this, reference feature data have been included in MARBLE, allowing users to harmonize their data using tools such as ComBat. 58 , 59 Also, recommended MRI acquisition protocols have been provided in Table SA1. Future studies should extend validation of MARBLE to a broader range of scanner and acquisition protocols and assess the reproducibility of MARBLE scores longitudinally. In summary, this work developed MARBLE, a novel, fully automated, non‐invasive in vivo marker of LATE‐NC, based on multimodal MRI features and detailed pathology data from > 900 community‐based older adults. In vivo validation of MARBLE in an independent group (ADNI/NACC) showed excellent performance in detecting LATE‐NC (AUC = 0.92) for data collected with short AMIs (< 3 years). Moreover, higher MARBLE scores in the full independent group were associated with more advanced LATE‐NC stages assessed pathologically. Furthermore, higher baseline MARBLE scores in older adults that were largely dementia free were associated with clinically meaningful outcomes such as lower baseline levels and accelerated decline of cognition. Considering that definitive diagnosis of LATE‐NC is currently only possible at autopsy, MARBLE holds strong potential for advancing clinical diagnosis, refining participant selection in clinical drug and prevention trials, and facilitating the development of treatment strategies for this common neurodegenerative condition affecting the aging brain. CONFLICT OF INTEREST STATEMENT The authors have no financial interests or relationships to disclose with regard to the subject matter of this manuscript. Author disclosures are available in the Supporting Information. CONSENT STATEMENT All datasets used in this work were approved by each center's respective institutional review board, and written informed consent was obtained from all participants. Supporting information Supporting information Supporting information Supporting information Supporting information ACKNOWLEDGMENTS The authors would like to thank the participants and staff of the Rush University Memory and Aging Project, Religious Orders Study, Minority Aging Research Study, African American Clinical Core, Latino Core of the Rush Alzheimer's Disease Research Center, Alzheimer's Disease Neuroimaging Initiative, and National Alzheimer's Coordinating Center. Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 as a public–private partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of MCI and early AD. Data collection and sharing for ADNI is funded by the National Institute on Aging (National Institutes of Health Grant U19 AG024904). The grantee organization is the Northern California Institute for Research and Education. In the past, ADNI has also received funding from the National Institute of Biomedical Imaging and Bioengineering, the Canadian Institutes of Health Research, and private sector contributions through the Foundation for the National Institutes of Health (FNIH) including generous contributions from the following: AbbVie; Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol‐Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann‐La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC; Johnson & Johnson Pharmaceutical Research &Development LLC; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The NACC database is funded by NIA/NIH Grant U24 AG072122. NACC data are contributed by the NIA‐funded ADRCs: P30 AG062429 (PI James Brewer, MD, PhD), P30 AG066468 (PI Oscar Lopez, MD), P30 AG062421 (PI Teresa Gomez‐Isla, MD), P30 AG066509 (PI Thomas Grabowski, MD), P30 AG066514 (PI Mary Sano, PhD), P30 AG066530 (PI Helena Chui, MD, Arthur Toga, PhD), P30 AG066507 (PI Marilyn Albert, PhD), P30 AG066444 (PI David Holtzman, MD), P30 AG066518 (PIs Lisa Silbert, MD, Kevin Duff, PhD), P30 AG066512 (PI Thomas Wisniewski, MD), P30 AG066462 (PI Scott Small, MD), P30 AG072979 (PI David Wolk, MD), P30 AG072972 (PIs Charles DeCarli, MD, Rachel Whitmer, PhD), P30 AG072976 (PI Andrew Saykin, PsyD), P30 AG072975 (PI Julie Schneider, MD, MS), P30 AG072978 (PI Ann McKee, MD), P30 AG072977 (PI Robert Vassar, PhD), P30 AG066519 (PI Joshua Grill, PhD), P30 AG062677 (PIs Brad Boeve, MD, Ronald Petersen, MD, PhD), P30 AG079280 (PI Jessica Langbaum, PhD), P30 AG062422 (PI Gil Rabinovici, MD), P30 AG066511 (PI Allan Levey, MD, PhD), P30 AG072946 (PI Linda Van Eldik, PhD), P30 AG062715 (PI Sanjay Asthana, MD, FRCP), P30 AG072973 (PI Russell Swerdlow, MD), P30 AG066506 (PIs Glenn Smith, PhD, ABPP, David Lowenstein, PhD, Ranjan Duara, MD), P30 AG066508 (PIs Stephen Strittmatter, MD, PhD, Christopher Van Dyck, MD), P30 AG066515 (PI Victor Henderson, MD, MS), P30 AG072947 (PI Suzanne Craft, PhD), P30 AG072931 (PI Henry Paulson, MD, PhD), P30 AG066546 (PIs Sudha Seshadri, MD, Gladys Maestre, MD, PhD), P30 AG086401 (PI Erik Roberson, MD, PhD), P30 AG086404 (PI Gary Rosenberg, MD), P30 AG086403 (PI Angela Jefferson, PhD), P30 AG072958 (PIs Heather Whitson, MD, Gwenn Garden, MD, PhD), P30 AG072959 (PI Jagan Pillai, MD, PhD), P30 AG092752 (Ihab Hajjar, MD, MS). This study was supported by the following grants from National Institutes of Health: R01AG064233, R01AG067482, R01AG017917, R01AG015819, RF1AG022018, R01AG056405, P30AG010161, P30AG072975 (National Institute on Aging), and U01NS100599 (National Institute of Neurological Disorders and Stroke). Tazwar M, Evia AM, Ridwan AR, et al. MARBLE: An in vivo MRI marker of limbic‐predominant age‐related TDP‐43 encephalopathy neuropathologic change (LATE‐NC). Alzheimer's Dement. 2026;22:e71832. 10.1002/alz.71832 Data used in preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp‐content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf DATA AVAILABILITY STATEMENT The MARBLE software package can be downloaded by submitting a request to https://www.nitrc.org/projects/marble. REFERENCES Associated Data Supplementary Materials Supporting information Supporting information Supporting information Supporting information Data Availability Statement The MARBLE software package can be downloaded by submitting a request to https://www.nitrc.org/projects/marble.