Exploring the power of structural brain MRI and clinical measures in predicting AD neuropathology: a machine learning approach
==============================================================================================================================

* Farooq Kamal
* Cassandra Morrison
* Michael D. Oliver
* Mahsa Dadar

## Structured Abstract

**Importance** Vascular and structural brain changes are increasingly recognized for their role in cognitive decline and progression of neurodegenerative conditions including Alzheimer’s disease (AD). Despite advances in imaging technologies, the exact contribution of these brain changes to disease processes remains a subject of ongoing research.

**Objective** To apply machine learning techniques to determine critical features of AD-related neuropathologies in vivo.

**Main Outcomes and Measures** A total of 127 participants (95 females, mean age=87.3) from the RUSH dataset and 65 participants (17 females, mean age=79.0) from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset were included. In the RUSH dataset, machine learning models were applied towards feature selection of MRI, clinical, and demographic data to identify the best performing set of variables that could predict neuropathology outcomes (e.g., Braak neurofibrillary tangle stage, neurofibrillary tangle burden; NFT). The best-performing neuropathology predictors using the top seven MRI, clinical, and demographic features were then validated in ADNI to compare results and ensure that the feature selection process did not lead to overfitting. For continuous measures, gradient boosting, bagging, support vector regression, and linear regression were implemented. For binary outcomes, logistic regression, gradient boosting, support vector machine, and bagging classifiers were utilized.

**Results** Applying feature ranking methods using similar information criteria, four machine learning models consistently ranked white matter hyperintensity (WMHs), gray matter (GM), and white matter (WM) volumes as important features in predicting all neuropathology measures. In the RUSH dataset, prediction accuracy was highest for Braak stage, NFT, and tangles (i.e., cross-validated correlation between actual measures and predictions was above 0.8). The best-performing model achieved *r*=0.83 (RMSE=0.50) in predicting tangles. The best-performing binary classifier achieved 82% accuracy, 86% sensitivity, and 78% specificity in predicting NIA-Reagan (measure of neurofibrillary tangles and neuritic plaques). Similar results were observed in the ADNI dataset.

**Conclusion and Relevance** These results highlight the efficacy of machine learning models, particularly when incorporating structural MRI features (e.g., GM, WM) alongside WMHs, in accurately predicting AD neuropathology. The use of machine learning may prove beneficial in early detection of AD pathology.

Keywords
*   Older adults
*   Alzheimer’s disease
*   Cognitive decline
*   White Matter Hyperintensities
*   Machine learning
*   Neuropathology

## 1. Introduction

Neurodegenerative diseases, characterized by a progressive decline in neurological function, pose a significant challenge to global healthcare systems, with Alzheimer’s disease (AD) standing out as a major public health concern. AD is characterized by pathological accumulation of neurofibrillary tau tangles and amyloid beta deposits.1 In addition to these pathological changes, structural brain alterations, including gray matter (GM) and white matter (WM) atrophy, and white matter hyperintensities (WMHs) contribute to cognitive decline in AD.2–5 GM changes (particularly volume reduction) are closely associated with cognitive impairment in AD, highlighting their importance in the neurodegeneration process.6–8 Research has also shown a strong relationship between WM volume changes and cognitive impairments across the spectrum of AD progression.9 Similarly, WMHs are markers of cerebrovascular disease burden and are increasingly recognized for their role in cognitive decline and the progression of neurodegenerative diseases.10–12

People with AD are more likely to present with structural and cerebrovascular changes such as GM and WM volume loss and WMH burden at autopsy.13–16 The volume of the entorhinal cortex6 and hippocampus7,8 are commonly studied to determine AD stage/progression as they are two of the first brain regions to show abnormal volume loss in AD. Although neurodegeneration has been included as a marker for staging AD, the Alzheimer’s Association Workgroup has recently proposed a revised criteria for diagnosing and staging of AD. The updated guidelines now also include vascular brain injury (often measured using MRI markers of cerebrovascular disease) as a common non-AD co-pathology,1 further highlighting the prevalence of vascular pathology in AD. Taking markers of both structural and cerebrovascular pathology into consideration, the present study employs a novel approach to examine whether these in vivo abnormalities are predictive of neuropathology severity in two samples, one predominantly with cognitively normal individuals, and the other predominantly with AD dementia.

Previous neuropathological studies in the literature have provided insights into early risk factors for AD by exploring the correlations between neuropathological and clinical or brain volume measures individually. However, these studies did not investigate the potential of using a combination of WMH, WM, and GM volumes as features to predict neuropathology. Thus, a comprehensive approach that combines these features to predict neuropathological outcomes remains unexplored. For instance, Kaur et al.17 explored the association between neuropathology (e.g., neurofibrillary tangle burden (NFT)) and total and regional brain volume and reported significant correlations between neuropathological measures and brain volume. A recent study by Zhang et al.18 further explored the relationship between clinical and neuropsychological measures and neuropathological data in people who reported normal cognitive functioning. They found that neuropathological measures such as Braak-NFT stage were significantly correlated with clinical measures of memory, executive function, and visuospatial function domains. These studies however predominantly concentrate on either clinical or MRI-based markers that correlate with neuropathology.17–20

With their ability to handle large datasets and complex nonlinear patterns, machine learning models have been widely utilized for disease diagnosis and prognosis applications.21 The integration of machine learning has opened new avenues for understanding and diagnosing neurodegenerative diseases. For example, support vector machines (SVMs) have been used to classify different stages of AD using MRI data, helping to outline disease progression.22 Similarly, linear regression has been instrumental in identifying biomarkers and predicting disease outcomes in neurodegenerative conditions.18 Despite these advances, accurately predicting neuropathology remains a significant challenge that has not been previously addressed.

Current research has yet to fully incorporate vascular and structural brain changes, including WMH, WM, and GM volumes, with clinical factors in machine learning models for neuropathology prediction. This gap underscores a critical area for exploration, as WMHs represent a potentially valuable marker for early disease detection that is yet to be fully leveraged. Given the limitations of the current research, the present study aims to integrate these features with other AD-related brain measures to predict post-mortem neuropathology. To achieve this, machine learning methods are applied to determine the importance of different features (e.g., cognitive, regional and global brain volumes, and WMH measures) in predicting AD-related neuropathologies. We hypothesize that in addition to typical MRI measures (i.e., brain volume) such as WM, GM, WMH burden will also be a critical feature that contributes to AD-related post-mortem neuropathology prediction.

## 2. Methods

### 2.1 Rush Alzheimer’s Disease Center (RADC) Participants

Data was used from the RADC Research Resource Sharing Hub ([www.radc.rush.edu](http://www.radc.rush.edu)). Participants provided informed written consent to participate in one of three cohort studies on aging and dementia: 1) Minority Aging Research Study,22 2) Rush Alzheimer’s Disease Center African American Clinical Core23, or 3) the Rush Memory and Aging Project.24

The participants had a minimum baseline age of 55 and were identified as cognitively normal (CN), MCI, or dementia at baseline. Cognitive diagnosis was determined using a three-stage process including computer scoring of cognitive tests, clinical judgment by a neuropsychologist, and diagnostic classification by a clinician based on the National Institute on Aging and the Alzheimer’s Association.25

There were a total of 192 people in the RUSH dataset that had postmortem neuropathology information and MRI measures. We only included those participants with MRI under 4 years apart from death (mean years in time between MRI and neuropathology = 2.35). A total of 127 participants (95 females, mean age=87.3) from the RUSH dataset had postmortem neuropathology information and MRI sequences from which volumetric measures and WMHs could be extracted were included. From the final clinical consensus diagnosis at time of death, there were 71 cognitively healthy older adults, 22 people with mild cognitive impairment (MCI), and 34 people with AD.

#### 2.1.1 Cognitive Battery

A comprehensive neuropsychological battery comprising 19 tests were administered to assess episodic memory, semantic memory, working memory, processing speed, and visuospatial ability.26,27 For detailed information on specific cognitive tests please see supplemental material or [https://www.radc.rush.edu/](https://www.radc.rush.edu/).

#### 2.1.2 MRI and post-mortem measures

Standard procedures determined by RADC and RUSH researchers were used to calculate MRI and assess post-mortem neuropathology. T1-weighted (T1w) 3D Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE) and T2-weighted 2D Fluid-Attenuated Inversion Recovery (FLAIR) were employed for structural assessments. T1w images were processed using FreeSurfer28 and regional GM volumes (i.e., hippocampal, amygdala, entorhinal cortex, putamen, caudate, thalamus, entorhinal, and parahippocampal gyrus) were extracted. Total GM and WM volumes as well as intracranial volumes (used for normalization) were calculated using the Computational Anatomy Toolbox (CAT)29 toolbox from SPM.30 WMHs were segmented using sysu31, a previously validated deep learning based automated WMH segmentation tool.

Neurofibrillary tangle burden, diffuse plaque burden, neuritic plaque density, global Alzheimer’s disease pathology, cerebral atherosclerosis rating, CERAD score for neuritic plaques, Braak staging for neurofibrillary tangles, presence of Alzheimer’s disease based on NIA-Reagan criteria, and assessments of chronic microinfarcts, and gross chronic cerebral infarcts were determined at post-mortem through neuropathological evaluations by the RUSH RADC investigators. Detailed are presented in the supplementary materials.

### 2.2 Alzheimer’s Disease Neuroimaging Initiative (ADNI) Participants

Data used in the preparation of this article were also obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). Participants from ADNI had baseline ages between 55 and 90. Participants were included if they had information for comparable variables to the RUSH dataset. That is, MRIs from which regional white and gray matter volume measurements and WMH volume could be extracted, completed the Clinical Dementia Rating-Sum of Boxes (CDR-SB, as a measure of global cognition), digit span (as a measure of working memory), trail making test – part B (as a measured of processing speed), everyday cognition scale (as a measure of subjective cognitive decline), and logical memory (as a measure of episodic memory). A total of 65 participants (17 females, mean age=79.0) from ADNI had postmortem neuropathology information and MRI measures from which volumetric measures could be extracted were included. There were 8 cognitively healthy older adults, 10 people with MCI, and 45 people with AD.

#### 2.2.1 Structural MRI acquisition and processing

All longitudinal scans were downloaded from the ADNI website (see [http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/](http://adni.loni.usc.edu/methods/mri-tool/mri-analysis/) for detailed MRI acquisition protocol). T1w scans for each participant were pre-processed through our standard pipeline including noise reduction,32 intensity inhomogeneity correction,33 and intensity normalization into range [0-100]. The pre-processed images were then linearly (9 parameters: 3 translation, 3 rotation, and 3 scaling)34 registered to the MNI-ICBM152-2009c average.35

#### 2.2.2 WMH measurements

A previously validated WMH segmentation technique was employed to generate participant WMH measurements36 (detailed in supplementary material).

#### 2.2.3 Freesurfer Measurements

T1w images were processed using FreeSurfer and quality controlled by the UCSF group, and regional GM volumes (i.e., hippocampal, amygdala, entorhinal cortex, putamen, caudate, thalamus, entorhinal, and parahippocampal gyrus) as well as total GM and WM volumes were extracted. 1.5T and 3T data were processed with FreeSurfer versions 4.3 and 5.1, respectively, as appropriate.28

## 3. Predictive modelling

### 3.1 Predictive modelling with multiple feature sets in RUSH

Analyses were performed using Python software version 3.11. The predictor variables included global cognition, episodic memory, subjective cognitive decline, semantic memory, perceptual speed, perceptual orientation, working memory, and structural brain volumes (WMHs, WM, total and regional GM) that were obtained based on the last available visit (within 4 years prior to death). Neurofibrillary tangle burden, diffuse plaque burden, neuritic plaque density, global Alzheimer’s disease pathology, cerebral atherosclerosis, CERAD score, Braak staging, NIA-Reagan criteria assessment, chronic microinfarcts, and gross chronic cerebral infarcts scores were the outcomes of interest for the prediction models.

All continuous values were standardized using the *StandardScaler* from the *sklearn.preprocessing* package in Python prior to analyses, aligning with best practices in machine learning research. 37–39 We incorporated a range of linear and nonlinear machine learning models, including gradient boosting classifier and regressor, support vector machines, linear regression, and bagging techniques. These models were selected due to their robustness in handling diverse datasets. To evaluate the model performance, we implemented a 10-fold cross-validation scheme using KFold from *sklearn.model_selection*, minimizing bias and overfitting.40 Each model within our framework was trained on a subset of the predictors, ranging from individual features to combinations of up to seven features. We evaluated models based on their performance in predicting continuous and binary outcomes. For binary outcomes, we calculated accuracy, sensitivity, and specificity, while for continuous outcomes, we measured root mean square error (RMSE) and correlation. These metrics provided a comprehensive assessment of model performance.

We examined combinations of predictors using *itertools.combinations*, testing each possible combination to identify the most informative feature set for predicting neuropathology outcomes. To further confirm the validity of our initial results, we employed a bootstrap method, which involved resampling our dataset and performing the same k-fold evaluation on each bootstrap sample. We conducted 50 bootstrap iterations, each involving a 10-fold cross-validation process. This technique, implemented using the resample function, is increasingly recognized for its efficacy in assessing model uncertainty and variance.41,42

Regression plots were used to visualize and evaluate the predictive performance of machine learning models by comparing neuropathology outcome values against the cross-validated predicted neuropathology values. Higher R2 values indicated a better fit of the model to the data, while lower root mean squared error (RMSE) values suggested higher accuracy in predictions.

Receiving operating characteristic (ROC) curve analysis was completed to evaluate the predictive performance of different machine learning models for the neuropathological outcome NIA-Reagan criteria. ROC curve analysis complements the accuracy measures by providing insight into the models’ ability to balance true positive and false positive rates. All results were calculated based on cross-validated predictions.

### 3.1 Predictive modelling with multiple feature sets in ADNI

Predictive modeling with multiple feature sets was also carried out in the ADNI dataset and compared to similar features in the RUSH dataset. In other words, to demonstrate the generalizability of the predictive power of the selected features across datasets, no additional feature selection process was performed for the ADNI. Instead, the closest available features from the ADNI to those selected in the feature selection step performed in the RUSH cohort were used to perform the same prediction task in the ADNI dataset. Examining the data in ADNI provides broader context and validation for our feature selection, mitigating the risk of overfitting and ensuring the generalizability of findings. All continuous values were z-scored within the population prior to analyses.

## 4. Results

Demographic information for both cohorts is presented in table 1. While the core results focus on three neuropathological measures with the highest correlation (i.e., Tangles, NFT, and Braak), a summary of all results including additional neuropathologies are presented in supplemental materials. Most informative predictors of neuropathology outcomes for Tangles, NFT, and Braak scores across each model are presented in Table 2 and results for NIA-Reagan Criteria are presented in Table 3. Informative predictors of neuropathology outcomes for Braak in the RUSH dataset and Braak in the ADNI dataset are presented in Table 4. Tangles and NFT measures were not available in the ADNI dataset. Therefore, the comparison of predictive power for Tangles and NFT between the two datasets was not included in Table 4.

View this table:
[Table 1:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T1)

Table 1: Demographic Information for both cohorts

View this table:
[Table 2:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T2)

Table 2: Most informative Predictors of neuropathology outcomes Tangles, NFT, and Braaksc across each model

View this table:
[Table 3:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T3)

Table 3: Most informative Predictors of neuropathology outcome NIA-Reagan Criteria across each model

View this table:
[Table 4:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T4)

Table 4: Most informative predictors of neuropathology outcomes NFT and Braak for the RUSH dataset and Braak in the ADNI dataset

### 4.1 Neuropathology Prediction in RUSH

For the outcome Tangles, the bagging model featuring predictors WM volume, caudate volume, entorhinal volume, global cognition, perceptual speed, APOE genotype demonstrated the highest retrained correlation of 0.80, with an RMSE of 6.34. Notably, the gradient boosting model, which included WMHs, hippocampal, amygdala, parahippocampal volume, global cognition, episodic memory, APOE genotype achieved a marginally higher correlation of 0.82, albeit a slightly higher RMSE of 7.30. Linear regression and support vector machine models yielded moderate correlations of 0.63 and 0.68, respectively.

For NFT, the gradient boosting model produced a robust correlation of 0.82 and the lowest RMSE of 0.50, suggesting its high predictive ability amongst the evaluated algorithms. This model incorporated WMHs, hippocampal volume, global cognition, episodic memory, perceptual orientation, APOE genotype, and sex as predictors, reflecting the multifaceted nature of AD pathology. The bagging model, with predictors total GM volume, WMHs, entorhinal cortex volume, global cognition, subjective cognitive decline, APOE genotype, sex, also exhibited a high correlation of 0.78. In contrast, linear regression and SVM models yielded lower correlations of 0.56 and 0.73, respectively, indicating less predictive accuracy.

In predicting Braak scores, the gradient boosting model showed a strong correlation of 0.81 and a low RMSE of 0.87, indicating this model’s ability to predict Braak pathology. This model utilized predictors WMHs, hippocampal volume, global cognition, subjective cognitive decline, perceptual orientation, APOE genotype, and age at MRI visit. The bagging model, which incorporated the predictors GM, caudate, thalamus, and entorhinal cortex volumes, global cognition, APOE genotype, and age at MRI visit, achieved a significant correlation of 0.78 and an RMSE of 0.87. On the other hand, linear regression and SVM models yielded lower correlations of 0.62 and 0.65, respectively.

#### 4.1.2 Scatter plots

Figure 1 shows the performance of the best model using scatter plots comparing actual versus predicted values for three neuropathological outcomes (Tangles, NFT, Braak). Among the models predicting tangles, the bagging model achieved the highest R2 value of 0.37, indicating that approximately 37% of the variance in the actual tangle measurements was accounted for by the model. The corresponding RMSE was 4.13, reflecting the average squared difference between actual and predicted values. For the NFT outcome, the SVM model showed the best performance, with an R2 of 0.55 and an RMSE of 0.29, indicating a strong predictive capability for this model in estimating neurofibrillary tangle measurements based on the selected features. When predicting Braak, the gradient boosting model achieved an R2 of 0.70 and an RMSE of 0.64.

![Figure 1:](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/02/29/2024.02.28.24303519/F1.medium.gif)

[Figure 1:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/F1)

Figure 1: Scatter plots showing the performance of the best models comparing actual versus predicted values for three neuropathological outcomes

![Figure 2:](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/02/29/2024.02.28.24303519/F2.medium.gif)

[Figure 2:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/F2)

Figure 2: ROC curves showing the predictive performance of SVM, logistic regression, gradient boosting, and bagging machine learning models for AD-related neuropathology using NIA-Reagan criteria, with gradient boosting achieving the highest AUC of 0.98.

### 4.2 Neuropathology Prediction in RUSH for binary outcome

In predicting the binary outcome NIA-Reagan Criteria, the bagging model demonstrated robust performance, achieving an accuracy of 0.87, with high sensitivity at 0.89 and specificity of 0.85. This model leveraged predictors including GM, WM, WMHs, putamen and entorhinal volume, perceptual speed, and APOE, indicating its comprehensive approach to capturing neuropathological changes. The gradient boosting model also showed strong predictive ability, with an accuracy of 0.87, the highest sensitivity among the models at 0.90, and a specificity of 0.82. It utilized an array of predictors which included GM, WM, amygdala and caudate volume, perceptual speed, APOE genotype, and sex. The Logistic regression model had lower performance metrics with an accuracy of 0.73, sensitivity of 0.80, and specificity of 0.63. The SVM model, yielded an accuracy of 0.78, sensitivity of 0.87, and specificity of 0.65. This result demonstrates the superior capability of nonlinear machine learning techniques, such as bagging and gradient boosting, to accurately differentiate the presence from the absence of the disease.

#### 4.2.1 ROC curve

In assessing the predictive performance of machine learning models for the binary neuropathological outcome defined by NIA-Reagan criteria, we also employed ROC curve analysis. The gradient boosting model demonstrated superior performance with an AUC of 0.98, suggesting an outstanding ability to distinguish between the presence and absence of AD-related neuropathologies. This model included GM, WM, amygdala, caudate volume, perceptual speed, APOE genotype, and sex as its primary predictors, underscoring a comprehensive set of factors contributing to AD pathology. Following closely, the bagging model achieved an AUC of 0.95, with GM, WM, WMHs, putamen, entorhinal volume, perceptual speed, and APOE genotype as the most informative predictors. This ensemble method effectively captured the complexity of neuropathological manifestations, achieving an accuracy of 0.87, with high sensitivity (0.89) and specificity (0.85). The SVM model also presented robust predictive power with an AUC of 0.88. In contrast, logistic regression displayed a modest AUC of 0.82. Although SVM and logistic regression did not include WM or WMHs, both models incorporated other MRI and clinical features.

### 4.3 Neuropathology Prediction with clinical predictors

To examine whether MRI features enhance predictive performance beyond that achieved with clinical features alone, we conducted a subsequent analysis focusing solely on clinical features. For Tangles, the gradient boosting model produced a moderate correlation. The model incorporated global cognition, perceptual speed, memory trouble, episodic memory, perceptual orientation, APOE genotype, and age at visit to achieve a correlation of 0.68 and an RMSE of 7.30. However, when compared to models solely based on clinical scores, those incorporating MRI features demonstrated superior performance. For the NIA-Reagan criteria, the gradient boosting model demonstrated an accuracy of 0.77, sensitivity at 0.83 and specificity at 0.71. Please see tables 5 and 6 for performance of clinical predictors.

View this table:
[Table 5:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T5)

Table 5: Most informative clinical predictors of neuropathology outcomes Tangles, NFT, and Braaksc across each model

View this table:
[Table 6:](http://medrxiv.org/content/early/2024/02/29/2024.02.28.24303519/T6)

Table 6: Most informative clinical predictors of neuropathology outcome NIA-Reagan criteria across each model

### 4.4 Neuropathology Prediction in ADNI

When comparing RUSH and ADNI datasets, predictors were chosen based on availability in both datasets. For Braak, predictive modeling within the RUSH dataset revealed that the gradient boosting model, incorporated the predictors WMHs, entorhinal cortex volume, global cognition, subjective cognitive decline, APOE genotype, sex, and age at visit, achieved a correlation of 0.80 and an RMSE of 0.53 (see table 4). When this model was adapted to the ADNI dataset, using the closest analogous predictors, it demonstrated a slight increase in performance, with a correlation of 0.77 and an RMSE of 0.62. The bagging model displayed a correlation of 0.76 and an RMSE of 0.65 for Braak in RUSH, and when applied to ADNI data, it maintained a strong correlation of 0.74 and an RMSE of 0.67.

## 5. Discussion

Previous literature has shown the importance of neurofibrillary tangles, amyloid beta deposits, and alterations in brain morphology, such as GM atrophy and WM abnormalities, in the development of AD.1,3 Although these studies have recognized the importance of brain changes, the integration of such features into machine learning models for neuropathology prediction analysis has been less explored. In this study, we investigated the predictive power of machine learning models in identifying neuropathology outcomes in aging individuals on the spectrum of AD, with a focus on the importance of MRI measures in addition to clinical measures. Our findings reveal that models incorporating MRI predictors, alongside clinical measures demonstrate significant predictive ability for key neuropathological measures such as Tangles, NFT, and Braak scores. Notably, the models including MRI features achieved higher performance than clinical measures alone. The gradient boosting and bagging models demonstrated superior predictive capabilities, achieving high correlations and low RMSE across all outcomes. When analyzing only clinical predictors, the performance metrics indicated lower predictive accuracy for AD neuropathology, reinforcing the added value of incorporating MRI features alongside clinical scores to enhance the predictive capability of machine learning models. These results are consistent with previous literature highlighting the role of structural and vascular brain changes in cognitive decline and AD progression.2–5 Results from the present study demonstrate the predictive utility of integrating vascular markers, specifically WMHs, with structural brain changes through machine learning models.

Furthermore, we analyzed two distinct datasets involving participants across a wide age range and at various stages of AD, including normal cognition. It is important to clarify that the clinical measures used in the RUSH and ADNI datasets were not identical. Adjustments were necessary due to the non-identical clinical measures used across the RUSH and ADNI datasets. We adapted our models to the available measures in ADNI, without repeating the feature selection step, to assess the generalizability of the selected features. Despite differences in age and disease severity between the RUSH and ADNI datasets, our models with matching predictors between datasets demonstrated similar accuracies. The matching predictors included WMH, regional and total WM, and GM volumes in addition to clinical scores. This ability to maintain high performance levels, irrespective of the dataset’s demographic and clinical diversity, underscores the model’s robust generalizability. This consistency in feature importance across datasets reinforces the reliability and relevance of these features in predicting AD neuropathology, aligning with, and expanding on previous machine learning research. 17,18

Our findings also revealed that nonlinear machine learning models outperformed linear models in predicting AD neuropathology. MRI features do not necessarily influence outcomes in a linear manner, therefore the nonlinear methods are able to better capture the variability observed in neuropathology. This finding underscores the significance for clinical settings and early prediction research, as it reveals that traditional methods, which assume linear patterns of change, may overlook crucial nuances in neuropathological progression due to the nonlinear influence of predictor variables like GM, WM, WMHs. Nonlinear models, capable of navigating these complexities, can combine information from a set of features to offer more accurate estimation of AD neuropathology and progression than linear models.

The application of ROC curve analysis post cross-validation revealed that the machine learning models employed in our study exhibited notably high AUC values in predicting the neuropathological outcome as defined by the AD-reagan criteria. The gradient boosting model, in particular, achieved an AUC of 0.98, indicative of its exceptional discriminative capacity. This high level of predictive performance, achieved using MRI features such as GM, WM, and WMHs, is particularly remarkable. The strong performance of these models demonstrates that machine learning can effectively leverage standard MRI features to provide insights into the underlying neuropathological state, complementing direct clinical assessments. However, the few studies that have examined neuropathology prediction have been in relatively small samples and report conflicting findings. For instance, a study by Toledo et al.43 utilized cerebrospinal markers, MRI features, and clinical data from N = 22 ADNI participants and found that non-MRI features were more effective than MRI features in predicting dementia with Lewy bodies in addition to Alzheimer’s pathology. They further highlighted that hippocampal volume was not a reliable feature for predicting pathologies. However, Kautzky et al.20 applied machine learning with MRI data in N = 49 participants to predict neuropathological changes in AD, achieving an accuracy of 77%. Important MRI features highlighted in their study included the entorhinal cortex and anterior cingulate cortex.

Despite the promising findings of our study, there are limitations that warrant attention. Due to the limited availability of longitudinal MRI measures in Rush dataset, our study relied on cross-sectional data from the closest MRI timepoint available (under four years from time of death). This limited our ability to assess whether and how much the time difference between the MRI and neuropathology assessments impact the performance of the models. Future studies within longitudinal settings could also assess the longitudinal progression of the neuropathology predictions, as well as their relation to clinical outcomes. Lastly, the clinical measures and biomarkers used in this study, though carefully selected, do not encompass all possible factors influencing AD neuropathology. Emerging biomarkers and novel imaging techniques continue to be identified, suggesting that integrating these new data sources could potentially enhance the predictive power of future models. Future studies should aim to address these challenges by incorporating longitudinal data, exploring models that offer greater interpretability, and including a broader range of biomarkers to improve prediction accuracy and clinical utility.

Our study demonstrates the feasibility and efficacy of using machine learning models, equipped with a carefully selected set of MRI features, to predict Alzheimer’s disease neuropathology across diverse datasets. These findings pave the way for further research and potential clinical applications, offering hope for improved early detection and management of Alzheimer’s disease.

## Supporting information

Supplementary Materials [[supplements/303519_file02.pdf]](pending:yes)

## Data Availability

All data produced are available online at RADC Research Resource Sharing Hub ([www.radc.rush.edu](http://www.radc.rush.edu)) and the Alzheimers Disease Neuroimaging Initiative (ADNI) database ([adni.loni.usc.edu](http://adni.loni.usc.edu/))

[https://adni.loni.usc.edu/](https://adni.loni.usc.edu/)

[https://www.radc.rush.edu/](https://www.radc.rush.edu/)

## Conflict of Interest

The authors declare no competing interests.

## Funding information

Alzheimer’s Disease Neuroimaging Initiative; This research was supported by a grant from the Canadian Institutes of Health Research.

## Financial Disclosures

Dr. Kamal is supported by a recruitment scholarship from Quebec Bio-Imaging Network. Dr. Morrison is supported by a postdoctoral fellowship from Canadian Institutes of Health Research, Funding Reference Number: MFE-176608. Dr. Dadar reports receiving research funding from the Healthy Brains for Healthy Lives (HBHL), Alzheimer Society Research Program (ASRP), and Douglas Research Centre (DRC).

## Consent Statement

Written informed consent was obtained from participants or their study partner

## Acknowledgments

Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through 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 Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health ([www.fnih.org](http://www.fnih.org)). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.

## Appendix: Collaborators

### Alzheimer’s Disease Neuroimaging Initiative

The data used for this article were from the ADNI database (adni.loni.usc.edu), which was established in 2003 as a public-private venture with Michael W. Weiner, MD as the Principal Investigator. ADNI’s primary aim is to determine if serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be utilized to measure the progression of MCI and AD. The research was granted ethical approval from the review boards of all the involved institutions, and written consent was obtained from participants or their study partner. The participants for this study were taken from all the ADNI cohorts (ADNI-1, ADNI-2, ADNI-GO, and ADNI-3).

## Footnotes

*   ‡ Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database ([adni.loni.usc.edu](http://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](http://adni.loni.usc.edu/wp-content/uploads/how\_to_apply/ADNI_Acknowledgement_List.pdf)

*   Received February 28, 2024.
*   Revision received February 28, 2024.
*   Accepted February 29, 2024.


*   © 2024, Posted by Cold Spring Harbor Laboratory

This pre-print is available under a Creative Commons License (Attribution-NonCommercial 4.0 International), CC BY-NC 4.0, as described at [http://creativecommons.org/licenses/by-nc/4.0/](http://creativecommons.org/licenses/by-nc/4.0/)

## References

1.  1.Jack CR, Bennett DA, Blennow K, et al. NIA-AA Research Framework: Toward a biological definition of Alzheimer’s disease. Alzheimer’s Dement. 2018;14(4):535–562. doi:10.1016/j.jalz.2018.02.018
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.jalz.2018.02.018&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

2.  2.Bilello M, Doshi J, Nabavizadeh SA, et al. Correlating cognitive decline with white matter lesion and brain atrophy magnetic resonance imaging measurements in Alzheimer’s disease. J Alzheimer’s Dis. 2015;48(4):987–994. doi:10.3233/JAD-150400
    
    [Abstract/FREE Full Text](http://medrxiv.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6NjoiaW9zamFkIjtzOjU6InJlc2lkIjtzOjg6IjQ4LzQvOTg3IjtzOjQ6ImF0b20iO3M6NTA6Ii9tZWRyeGl2L2Vhcmx5LzIwMjQvMDIvMjkvMjAyNC4wMi4yOC4yNDMwMzUxOS5hdG9tIjt9czo4OiJmcmFnbWVudCI7czowOiIiO30=) 

3.  3.Dadar M, Manera AL, Ducharme S, Collins DL. White matter hyperintensities are associated with grey matter atrophy and cognitive decline in Alzheimer’s disease and frontotemporal dementia. Neurobiol Aging. 2022;111:54–63. doi:10.1016/j.neurobiolaging.2021.11.007
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neurobiolaging.2021.11.007&link_type=DOI) 

4.  4.Kamal F, Morrison C, Maranzano J, Zeighami Y, Dadar M. Topographical differences in white matter hyperintensity burden and cognition in aging, MCI, and AD. GeroScience. 2023;45(1):1–16. doi:10.1007/s11357-022-00665-6
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s11357-022-00665-6&link_type=DOI) 

5.  5.Morrison C, Dadar M, Villeneuve S, Collins DL. White matter lesions may be an early marker for age-related cognitive decline. NeuroImage Clin. 2022;35(January):103096. doi:10.1016/j.nicl.2022.103096
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.nicl.2022.103096&link_type=DOI) 

6.  6.Igarashi KM. Entorhinal cortex dysfunction in Alzheimer’s disease. Trends Neurosci. 2023;46(2):124–136. doi:10.1016/j.tins.2022.11.006
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.tins.2022.11.006&link_type=DOI) 

7.  7.Fjell AM, McEvoy L, Holland D, Dale AM, Walhovd KB, Initiative. & ADN. What is normal in normal aging? Effects of Aging, Amyloid and Alzheimer’s Disease on the Cerebral Cortex and the Hippocampus. Prog Neurobiol. 2014;117:20–40. doi:10.1016/j.pneurobio.2014.02.004.What
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.pneurobio.2014.02.004&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24548606&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

8.  8.Risacher SL, Saykin AJ, West JD, et al. Baseline MRI predictors of conversion from MCI to probable AD in the ADNI cohort. Curr Alzheimer Res. 2009;6(4):347–361.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.2174/156720509788929273&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=19689234&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000271344300007&link_type=ISI) 

9.  9.Chen Y, Wang Y, Song Z, Fan Y, Gao T, Tang X. Abnormal white matter changes in Alzheimer’s disease based on diffusion tensor imaging: A systematic review. Ageing Res Rev. 2023;87(5):101911. doi:10.1016/j.arr.2023.101911
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.arr.2023.101911&link_type=DOI) 

10. 10. Van Den Berg E, Geerlings MI, Biessels GJ, Nederkoorn PJ, Kloppenborg RP. White Matter Hyperintensities and Cognition in Mild Cognitive Impairment and Alzheimer’s Disease: A Domain-Specific Meta-Analysis. J Alzheimer’s Dis. 2018;63(2):515–527. doi:10.3233/JAD-170573
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3233/JAD-170573&link_type=DOI) 

11. 11.Gouw AA, Seewann A, Vrenken H, et al. Heterogeneity of white matter hyperintensities in Alzheimer’s disease: Post-mortem quantitative MRI and neuropathology. Brain. 2008;131(12):3286–3298. doi:10.1093/brain/awn265
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1093/brain/awn265&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=18927145&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000261996700020&link_type=ISI) 

12. 12.Piguet O, Double KL, Kril JJ, et al. White matter loss in healthy ageing: A postmortem analysis. Neurobiol Aging. 2009;30(8):1288–1295. doi:10.1016/j.neurobiolaging.2007.10.015
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neurobiolaging.2007.10.015&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=18077060&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

13. 13.Bangen KJ, Nation DA, Delano-Wood L, et al. Aggregate effects of vascular risk factors on cerebrovascular changes in autopsy-confirmed Alzheimer’s disease. Alzheimer’s Dement. 2015;11(4):394–403.e1. doi:10.1016/j.jalz.2013.12.025
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.jalz.2013.12.025&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=25022538&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

14. 14.Di Patre, P. L., Read, S. L., Cummings, J. L., Tomiyasu, U., Vartavarian, L. M., Secor, D. L., & Vinters H V. Progression of clinical deterioration and pathological changes in patients with Alzheimer disease evaluated at biopsy and autopsy. Arch Neurol. 1999;56(10):1254–1261.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/archneur.56.10.1254&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=10520942&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000083012500010&link_type=ISI) 

15. 15.Iordanishvili E, Schall M, Loução R, Zimmermann, M., Kotetishvili K, Shah NJ, Oros-Peusquens AM. Quantitative MRI of cerebral white matter hyperintensities: a new approach towards understanding the underlying pathology. Neuroimage. 2019;202:116077.
    
    

16. 16.Nishiyama Y, Kanayama H, Mori H, et al. Whole brain analysis of postmortem density changes of grey and white matter on computed tomography by statistical parametric mapping. Eur Radiol. 2017;27:2317–2325.
    
    

17. 17.Kaur B, Himali JJ, Seshadri S, et al. Association between neuropathology and brain volume in the framingham heart study. Alzheimer Dis Assoc Disord. 2014;28(3):219–225. doi:10.1097/WAD.0000000000000032
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/WAD.0000000000000032&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24614264&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

18. 18.Zhang M, Ganz AB, Rohde S, et al. The correlation between neuropathology levels and cognitive performance in centenarians. Alzheimer’s Dement. 2023;19(11):5036–5047. doi:10.1002/alz.13087
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/alz.13087&link_type=DOI) 

19. 19.Boyle PA, Yu L, Wilson RS, Schneider JA, Bennett DA. Relation of neuropathology with cognitive decline among older persons without dementia. Front Aging Neurosci. 2013;5(SEP):1–8. doi:10.3389/fnagi.2013.00050
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3389/fnagi.2013.00001&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=23430962&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

20. 20.Kautzky A, Seiger R, Hahn A, et al. Prediction of autopsy verified neuropathological change of Alzheimer’s disease using machine learning and MRI. Front Aging Neurosci. 2018;10(December):1–11. doi:10.3389/fnagi.2018.00406
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3389/fnagi.2018.00021&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

21. 21.Mateos-Pérez JM, Dadar M, Lacalle-Aurioles M, Iturria-Medina Y, Zeighami Y, Evans AC. Structural neuroimaging as clinical predictor: A review of machine learning applications. NeuroImage Clin. 2018;20(August):506–522. doi:10.1016/j.nicl.2018.08.019
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.nicl.2018.08.019&link_type=DOI) 

22. 22.Barnes LL, Shah RC, Aggarwal NT, Bennett DA, Schneider JA. The Minority Aging Research Study: Ongoing efforts to obtain brain donation in African Americans without dementia. Arriola Barnes, Barnes, Beekly, Bennett, Bennett, Bennett, Bonner, Boulware, Boyle, Brookmeyer, Buchman, Cooper, de la Monte, Dressler, Froehlich, Fyffe, Garte, Grann, Gravlee, Green, Herring, James, Jorde, Keyzer, Krieger, Kuzawa, Lambe, LaVeist, Levkof A, ed. Curr Alzheimer Res. 2012;9(6):734–745. doi: 10.2174/156720512801322627
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.2174/156720512801322627&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=22471868&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

23. 23.Schneider JA, Aggarwal NT, Barnes L, Boyle P, Bennett DA. The neuropathology of older persons with and without dementia from community versus clinic cohorts. J Alzheimer’s Dis. 2009;18(3):691–701. doi:10.3233/JAD-2009-1227
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3233/JAD-2009-1227&link_type=DOI) 

24. 24.Bennett DA, Buchman AS, Boylea PA, Barnesa LL, S. RW, Schneidera JA. Religious Orders Study and Rush Memory and Aging Project. J Alzheimer’s Dis. 2018;64(Suppl 1):S161-S189. doi:10.3233/JAD-179939.Religious
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.3233/JAD-179939.Religious&link_type=DOI) 

25. 25.McKhann GM, Knopman DS, Chertkow H, et al. The diagnosis of dementia due to Alzheimer’s disease: Recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s Dement. 2011;7(3):263–269. doi:10.1016/j.jalz.2011.03.005
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.jalz.2011.03.005&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21514250&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000291239600003&link_type=ISI) 

26. 26.Wilson RS, CF DL, LL B, et al. Participation in cognitively stimulating activities and risk of incident Alzheimer disease. JAMA. 2002;287(6):742–748.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/jama.287.6.742&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=11851541&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000173809500031&link_type=ISI) 

27. 27.Barnes LL, Yumoto F, Capuano A, Wilson RS, Bennett DA, Tractenberg RE. Examination of the factor structure of a global cognitive function battery across race and time. J Int Neuropsychol Soc. 2015;22(1):66–75. doi:10.1017/S1355617715001113
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1017/S1355617715001113&link_type=DOI) 

28. 28.Fischl B. Freesurfer. Neuroimage. 2012;62(5):774–781.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neuroimage.2012.01.021&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=22248573&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000306390600031&link_type=ISI) 

29. 29.Gaser C, Dahnke R. CAT: A Computation Anatomy Toolbox for SPM.
    
    

30. 30.Friston KJ, Ashburner J, Frith CD, Poline JB, Heather JD, Frackowiak RSJ. Spatial registration and normalization of images. Hum Brain Mapp. 1995;3(3):165–189.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/hbm.460030303&link_type=DOI) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=A1995TY22200002&link_type=ISI) 

31. 31.Li H, Jiang G, Zhang J, et al. Fully convolutional network ensembles for white matter hyperintensities segmentation in MR images. Neuroimage. 2018;183:650–665.
    
    

32. 32.Coupe P, Yger P, Prima S, Hellier P, Kervrann C, Barillot C. An optimized blockwise nonlocal means denoising filter for 3-D magnetic resonance images. IEEE Trans Med Imaging. 2008;27(4):425–441. doi:10.1109/TMI.2007.906087
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1109/TMI.2007.906087&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=18390341&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000254812600001&link_type=ISI) 

33. 33.Sled JG, Zijdenbos AP, Evans AC. A nonparametric method for automatic correction of intensity nonuniformity in mri data. IEEE Trans Med Imaging. 1998;17(1):87–97. doi:10.1109/42.668698
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1109/42.668698&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=9617910&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000073646700008&link_type=ISI) 

34. 34.Dadar M, Fonov VS, Collins DL. A comparison of publicly available linear MRI stereotaxic registration techniques. Neuroimage. 2018;174(March):191–200. doi:10.1016/j.neuroimage.2018.03.025
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neuroimage.2018.03.025&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=29548850&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

35. 35.Fonov V, Evans AC, Botteron K, Almli CR, McKinstry RC, Collins DL. Unbiased average age-appropriate atlases for pediatric studies. Neuroimage. 2011;54(1):313–327. doi:10.1016/j.neuroimage.2010.07.033
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neuroimage.2010.07.033&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=20656036&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000283825000033&link_type=ISI) 

36. 36.Dadar M, Maranzano J, Ducharme S, Collins DL. White matter in different regions evolves differently during progression to dementia. Neurobiol Aging. 2019;76:71–79. doi:10.1016/j.neurobiolaging.2018.12.004
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neurobiolaging.2018.12.004&link_type=DOI) 

37. 37.Bhattarai P, Taha A, Soni B, Thakuri DS, Ritter E, Chand GB. Predicting cognitive dysfunction and regional hubs using Braak staging amyloid-beta biomarkers and machine learning. Brain Informatics. 2023;10(1):33.
    
    

38. 38.Feurer, M., Klein, A., Eggensperger K, Springenberg JT, Blum M, Hutter F. Auto-sklearn: efficient and robust automated machine learning. Autom Mach Learn. Published online 2019:113–134.
    
    

39. 39.Kumari R, Singh J, Gosain A. International Conference on Information and Communication Technology for Intelligent Systems. Singapore Springer Nat Singapore. Published online 2023:145–153.
    
    

40. 40.Varoquaux G, Raamana PR, Engemann DA, Hoyos-Idrobo A, Schwartz Y, Thirion B. Assessing and tuning brain decoders: cross-validation, caveats, and guidelines. Neuroimage. 2017;145:166–179.
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neuroimage.2016.10.038&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=27989847&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

41. 41.Tekin S, Mega MS, Masterman DM, et al. Orbitofrontal and anterior cingulate cortex neurofibrillary tangle burden is associated with agitation in Alzheimer disease. Ann Neurol. 2001;49(3):355–361. doi:10.1002/ana.72
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/ana.72&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=11261510&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 
    
    [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000167250100012&link_type=ISI) 

42. 42.Duchesne S, Caroli A, Geroldi C, Collins DL, Frisoni GB. Relating one-year cognitive change in mild cognitive impairment to baseline MRI features. Neuroimage. 2009;47(4):1363–1370. doi:10.1016/j.neuroimage.2009.04.023
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neuroimage.2009.04.023&link_type=DOI) 
    
    [PubMed](http://medrxiv.org/lookup/external-ref?access_num=19371783&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F02%2F29%2F2024.02.28.24303519.atom) 

43. 43.Toledo JB, Cairns NJ, Da X, et al. Clinical and multimodal biomarker correlates of ADNI neuropathological findings. Acta Neuropathol Commun. 2014;2(1):1–13. doi:10.1186/20515960165
    
    [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1186/20515960165&link_type=DOI)