MRI Diffusion Connectomics-Based Characterization of Progression in Alzheimer’s Disease

Abstract

Characterizing Alzheimer’s disease (AD) progression remains a significant clinical challenge. The initial stages of AD are marked by the accumulation of amyloid-beta plaques and Tau tangles, with cognitive functions often appearing normal, and clinical symptoms may not manifest until up to 20 years after the prodromal period begins. Comprehensive longitudinal studies analyzing brain-wide structural connectomics in the early stages of AD, especially those with large sample sizes, are scarce. In this study, we investigated a longitudinal diffusion-weighted imaging dataset of 264 subjects to assess the predictive potential of diffusion data for AD. Our findings indicate the potential of a simple prognostic biomarker for disease progression based on the hemispheric lateralization of mean tract volume for tracts originating from the supramarginal and paracentral regions, achieving an accuracy of 86%, a sensitivity of 86%, and a specificity of 93% when combined with other clinical indicators. However, diffusion-weighted imaging measurements alone did not provide strong predictive accuracy for clinical variables, disease classification, or disease conversion. By conducting a comprehensive tract-by-tract analysis of diffusion-weighted characteristics contributing to the characterization of AD and its progression, our research elucidates the potential of diffusion MRI as a tool for the early detection and monitoring of neurodegenerative diseases and emphasizes the importance of integrating multi-modal data for enhanced predictive analytics.

1. Introduction

An estimated 35 million people worldwide suffered from Alzheimer’s Disease in 2022, with 7 million new cases every year [1]. The percentage of people with Alzheimer’s disease, the most common form of dementia, increases with age, where 5.0% of people aged 65 to 74, 13.1% of people aged 75 to 84, and 33.2% of people aged 85 and older have Alzheimer’s [2].

1.1. Research in Context

1.1.1. Research before This Study

Comprehensive whole-brain analyses on longitudinal data remain relatively rare. There is mixed support for the utility of diffusion-weighted imaging data in developing biomarkers. Few studies establish prognostic biomarkers to indicate disease conversion. Numerous studies conduct correlational analysis to understand the effect size of the diffusion characteristics between disease stages, and many studies develop classification models with high accuracy, often relying on clinical indicators as predictors. However, little research has been conducted to assess the ability of diffusion characteristics in isolation to characterize disease stages or conversion.

1.1.2. Added Value of This Study

To our knowledge, this study is the first to systematically evaluate the potential of diffusion-weighted imaging (DWI) metrics, in isolation, to predict a range of neuropsychological and neurobiological indicators commonly used in staging Alzheimer’s disease with an aim to develop objective, non-invasive cognitive assessment methods that could complement or potentially reduce reliance on traditional clinical testing procedures.

This study integrates tractography metrics with rich phenotypic and neuropsychological data across a substantial longitudinal dataset focused on Alzheimer’s research. This integration allows for a thorough analysis of neurodevelopmental changes over time as subjects progress in Alzheimer’s disease with the aim of identifying early-stage prognostic biomarkers. The interaction between the hemispheric lateralization of tract volumes connected to the supramarginal gyrus and paracentral regions offers a potential prognostic biomarker of Alzheimer’s disease progression, with AUROC of 74% and AUPRC of 75% being important findings of this study. While this region has been implicated in previous studies, our analysis is novel in that it uses tractography metrics to establish this connection. To the best of our knowledge, this is the first study to implicate the supramarginal gyrus in AD progression by using detailed tractography measurements.

1.1.3. Implications

Diffusion-weighted imaging measurements alone did not provide strong predictive accuracy for clinical variables, disease classification, or disease conversion. Our analysis of white matter tract features revealed moderate but notable associations with neurobiological and neuropsychological markers, opening the door to future potential models, likely based on multi-modal data capable of predicting these clinical indicators. Our findings also demonstrate the potential for a simple prognostic biomarker of disease conversion when combined with other clinical indicators.

A brief overview of this study is provided in

Table 1. Study overview.

Variable	Measurement
Participants	264 (434 sessions)
Whole-brain tractography measurements	12 diffusion measurements per tract
Disease stages	CN, MCI, and AD
Neuropsychological measurements	9
Neurobiological measurements	7
Features used	Whole-brain tractography measurements
Feature reduction	ElasticNet
Evaluation metrics	RMSE, AUROC, and AUPRC
Performance	AUROC of 74% and AUPRC of 75% based exclusively on diffusion measurements

.

A definitive diagnosis of Alzheimer’s disease (AD) can only be confirmed by a histological examination of brain tissue post-mortem [3]. The initial stages of AD are characterized by the accumulation of plaques of the protein amyloid-beta (Aβ) in the medial parietal cortex. In this prodromal stage, cognitive function appears normal, and patients may not exhibit clinical symptoms up to 20 years after the start of the prodromal stage [4]. As the condition advances, other signs of neurodegeneration, such as neuronal death [5], atrophy (depending on the subtype) [6], and gliosis [7], become discernible after a variable period of latency. These changes correlate with clinical cognitive evaluations taken over multiple years and align with a suite of biomarkers, including hippocampal volume and heightened concentrations of Aβ and Tau proteins, which were pinpointed as indicators in the timeline of AD progression as described in the established literature [8].

The prevailing consensus is that early-stage therapeutic interventions could offer the greatest potential to improve health outcomes before irreversible neuronal loss and damage to brain tissue occur. Estimates suggest that providing treatment during the disease’s preclinical phase could significantly curtail its progression. In fact, some projections indicate that “a delay of 10 years would result in virtual disappearance of the disease” [9]. The first neurons damaged are those responsible for memory, language, and cognition. However, the pathophysiological processes that cause this damage are thought to begin 20 years before symptoms are reported [1,10]. Since Alzheimer’s disease is a gradual and progressive neurodegenerative disorder, understanding the potential of biomarkers to characterize the disease’s pathology and its long-term development is a key motivation behind this study. These biomarkers may support the identification of individuals who could benefit from treatment, potentially improving health outcomes for patients with the disease.

The current gold standard for the diagnosis of Alzheimer’s disease is biopsy or autopsy [11]. Recent studies have invested considerable effort and resources in the early detection of AD in the prodromal stages of mild cognitive impairment (MCI). Altered brain asymmetry of subcortical structures, reduction in cortical thickness, and hippocampal, entorhinal, fusiform and medial temporal lobe volumes are all proposed biomarkers of AD [11]. However, the net improvement in AD diagnostic accuracy from structural MRI tests following clinical neurocognitive memory assessments has been shown to be low, +1.1% (95% CI 0.1 to 3.9) [12]. In contrast, diffusion-weighted imaging (DWI) techniques have shown promise [13]. DWI was designed to study white matter (WM) structure [14], a tissue to which AD has been associated [15,16,17]. This modality is particularly useful, given that AD exhibits degeneration of cellular barriers of neurons and fiber tracts as a result of the buildup of Tau proteins [18]. In recent years, a large body of research has focused on leveraging DWI to classify AD stages and predict disease progression [10,13,19,20,21,22,23,24].

Despite growing evidence that diffusion-weighted imaging correlates with disease severity [25], we have been unable to find a comprehensive analysis of whether diffusion-weighted imaging can be used to predict established biomarkers used for staging AD. In this study, we investigate the potential of tractography metrics to predict neuropsychological and neurobiological test results in the context of Alzheimer’s disease progression while also aiming to identify early prognostic biomarkers of AD. Our analysis addresses four key questions: (Q1) Do phenotypic characteristics predict cognitive decline within our current dataset? (Q2) Are WM tract features predictive of neuropsychological and neurobiological indicators? (Q3) Does baseline tract volume change with cognitive decline? (Q4) To what extent can tractography metrics predict cognitive decline? We present our methodology for data acquisition and analysis, followed by results corresponding to each research question. Finally, we discuss the implications of our findings for both clinical practice and future research directions in neuroimaging and Alzheimer’s disease.

2. Materials and Methods

2.1. Participants

In this study, two hundred and sixty-four participants across multiple exams totaling 434 sessions 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 the ADNI is to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment could be combined to measure the progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD). The images obtained are associated with a clinical diagnosis and a specific time point. As such, individuals may span one or more clinical diagnoses, such as progression from cognitive normal (CN) → MCI → AD, and therefore can have images at each diagnosis.

Table 2. Stratification of ADNI images and associated demographic and clinical details.

Diagnosis at Imaging	Progression Profile	Subjects (F:M)	Mean Age (F:M)	SD (F:M)
CN	CN	126:92	71:78	7.4:7.7
CN	CN → MCI	5:4	74:79	10.8:8.6
CN	CN → AD	1:0	88: -	-:-
MCI	MCI	60:75	76:76	7.7:6.9
MCI	CN → MCI	4:5	73:82	13.4:5.8
MCI	MCI → AD	2:1	82:83	8.1:-
AD	AD	28:27	78:76	7.9:8.4
AD	CN → AD	1:0	87: -	-:-
AD	MCI → AD	2:1	83:85	7.6:7.7

Note: CN = cognitively normal, MCI = mild cognitive impairment, and AD = Alzheimer’s disease.

shows the number of participants by clinical diagnosis at imaging and clinical progression.

2.2. MRI Acquisition

T1-weighted (T1w) images were acquired at 3T, 208 × 240 × 256 voxels of size 1 mm³. Diffusion MRI data were acquired by using diffusion-weighted single-shot spin-echo echo-planar imaging. Fifty-six slices of 2 mm in thickness, yielding 2 mm isotropic voxels, were obtained. Forty-nine diffusion-weighted measurements (b = 1000 s/mm²) and seven non-diffusion-weighted measurements (b = 0 s/mm²) were acquired with TR = 7200 ms, TE = 56 ms, and field of view = 232 mm × 232 mm.

The imaging sequences and parameters of the anatomical scans followed the Alzheimer’s Disease Neuroimaging Initiative 3 protocols (https://adni.loni.usc.edu/wp-content/themes/freshnews-dev-v2/documents/mri/ADNI3-MRI-protocols.pdf, accessed on 16 June 2022) and were collected across 57 imaging centers. The data used in our research included all subjects in the ADNI3 cohort with at least one T1-weighted MP-RAGE and a corresponding diffusion-weighted image (DWI).

2.3. Preprocessing of MRI Data and Estimation of Structural Networks

We developed a pipeline management software package [26] offering a robust, fault-tolerant, and extensible platform to execute the processing workflows of the ADNI data. The pipeline code was containerized by using Apptainer [27] to facilitate reproducibility as well as environment management during pipeline execution on each of the super-computing clusters in which software was executed. We sought to process all ADNI3 participant sessions that contained both a T1-weighted image and a diffusion-weighted image. A total of 961 participants, totaling 1873 exam sessions, were considered for our study. After excluding scans without DWI images or bval/bvec files, 264 participants represented in 434 sessions were processed by our pipeline.

The preprocessing pipeline to extract tractography metrics from the ADNI dataset has been described elsewhere [28,29]. For the ADNI dataset, T1w MRI images were segmented into sub-regions by using the FreeSurfer software package, version 7.2 [30], with cortical [31] and subcortical [30] labeling pipelines. The white matter volume generated by Freesurfer’s recon-all was further separated into 181 regions of interest (ROIs) by using the Freesurfer program mri_extract_label. Labels were extracted by using the Desikan--Killiany atlas [31]. The diffusion-weighted image was registered to the T1w image before orientation distribution function (ODF) estimation was performed. ODF maps were created from the preprocessed DWI images by using the Diffusion Toolkit (DTK v0.6.4) [32] software package. The HARDI/Q-ball imaging model [33] with a fiber orientation distribution function was estimated at each voxel. The Fiber Assignment by Continuous (FACT)-alike tracking algorithm [32,34] was employed for deterministic fiber tracking. Seed points for tractography were generated throughout the entire brain volume where valid diffusion data existed, using a 35° angle threshold for stopping criteria. This whole-brain tract file was constructed by using the odf_recon and odf_tracker utilities from DTK.

The generated tract file and individual white matter ROI masks were postprocessed for the extraction of 32,580 tracts (cartesian product of ROIs), with multiple measures extracted for each, including mean tract length, tract volume, mean fractional anisotropy (FA), FA standard deviation, mean diffusivity (MD) calculated by using the mean of the three eigenvalues, MD standard deviation, and the corresponding left–right asymmetries of each of these measurements (12 measurements in total). We derived these measurements for tracts that started or terminated in our ROIs as well as tracts that passed through our ROIs, for a total of 781,920 measurements per scan session. White matter tract identification was assessed by using q-ball imaging [33,35]. Although a variety of structural connectomics analytics technologies have been developed for the analysis of diffusion-enabled brain MRI examinations [34,36,37,38,39], our approach includes a thorough assessment of features with characterization potential, such as the variability (as measured with the standard deviation) in FA and MD along each given fiber pathway, as well as hemispheric asymmetry measurements for all aforementioned features.

2.4. Statistics

We applied whole-brain deterministic tractography techniques to generate a high-dimensional dataset across 434 exams. Each subject had approximately 1.3 million feature measurements on average being evaluated for potential as a biomarker for characterizing AD, depending on the number of visits a subject participated in. The ADNI3 dataset includes detailed clinical biomarkers for most participants at each imaging session, providing tremendous benefit, as they offer potential indicators of early decline in the preclinical stages or for those with mild cognitive impairment (MCI). We targeted 16 neuropsychological and neurobiological phenotypic characteristics (see

Table 3. Table of ADNI biomarkers available per scan/session.

Phenotypic Characteristic	Range	Description
ADAS-11 (neuropsychological)	0–70	A rating scale to assess the severity of cognitive and non-cognitive dysfunction from mild to severe AD. ADAS-11 assesses the cognitive domains of memory, language, and praxis. Specific tasks include Word Recall, Naming Objects and Fingers, Commands, Constructional Praxis, Ideational Praxis, Orientation, Word Recognition, and Language
ADAS-13 (neuropsychological)	0–85	A rating scale to assess cognitive domains hypothesized to be important treatment targets of antidementia drugs that are not assessed by the ADAS-11: attention and concentration, planning and executive function, verbal memory, nonverbal memory, and praxis.
AV45 (PET image analysis)		An imaging biomarker for amyloid plaque accumulation in subjects with cognitive impairment that may be attributed to the presence of Alzheimer’s disease. Average AV45 regional standardized uptake values of frontal, anterior cingulate, precuneus, and parietal cortex relative to the cerebellum.
BRAIN VOLUME (MR image analysis)		Volume (mm3) of brain
Diagnosis	[CN,MCI,AD]	Diagnosis classification at scan
ENTORHINAL VOLUME (MR image analysis)		Volume (mm3) of entorhinal cortex
FDG (PET scan information)		Average fluorodeoxyglucose PET of angular, temporal, and posterior cingulate. It reflects loss of neuropil, loss of synapse, and functional impairment of neurons. Lower FDG-PET was regarded as a signal of neuronal hypometabolism due to neurodegeneration.
FUSIFORM VOLUME (MR image analysis)		Volume (mm3) of fusiform
HIPPOCAMPUS VOLUME (MR image analysis)		Volume (mm3) of hippocampus
ICV (MR image analysis)		Intracranial volume. In patients with dementia, but not in MCI, severity of cognitive impairment and ICV were moderately correlated. The effect of ICV on cognition was not mediated by hippocampal atrophy.
MIDTEMPORAL VOLUME (MR image analysis)		Volume (mm3) of mid temporal
VENTRICLE VOLUME (MR image analysis)		Volume (mm3) of ventricles
CDR SB (neuropsychological)	0–36	Clinical Dementia Rating scale Sum of Boxes (CDR-SB) score. This score has been used to accurately stage severity of Alzheimer dementia and mild cognitive impairment (MCI).
FAQ (neuropsychological)	0–30	The Functional Activities Questionnaire (FAQ) measures instrumental activities of daily living such as preparing balanced meals and preparing finances. A cut-point of 9 (dependent in 3 or more activities) is recommended to indicate impaired function and possible cognitive impairment.
MMSE (neuropsychological)	0–30	Mini Mental State Examination. It is an 11-question measure that tests five areas of cognitive function: orientation, registration, attention and calculation, recall, and language. The maximum score is 30. A score of 23 or lower is indicative of cognitive impairment.
MoCA (neuropsychological)	0–30	Montreal Cognitive Assessment Test for Dementia. This test is a 30-item test of language, memory, visual and spatial thinking, reasoning, and orientation skills. A score of 26 or above is considered normal.
RAVLT Immediate (neuropsychological)		Rey Auditory Verbal Learning Test evaluating short-term memory, working memory, and long-term memory. RAVLT Immediate is the sum of scores from the 5 first trials.

) as response variables in our analysis of white matter tract features.

We performed a univariate statistical analysis on each feature derived within our dataset to understand range, central tendency, standard deviation, and correlation to each of the 16 selected phenotypic characteristics provided by the ADNI as they related to cognitive decline. Cognitive decline is presented in ADNI data as a categorical variable representing diagnosis. Using our dataframe containing diffusion measurements derived from dMRI and T1-weighted images, we sought to answer the following research questions.

Q1.

Do phenotypic characteristics predict cognitive decline within our current dataset?

To investigate potential associations between phenotypic biomarkers (Table 3) and diagnostic categories in our study, we employed the Kruskal–Wallis test, a non-parametric test suitable for comparing distributions across multiple groups. The Kruskal–Wallis test enabled an overall assessment of whether statistically significant differences exist in the distribution of biomarker values across diagnostic categories (CN, MCI, and AD). This approach was chosen due to its robustness in handling non-normally distributed data and its ability to discern variations in central tendencies across multiple groups. All statistical analyses were performed by using R (version 4.1.3) with a significance threshold set at p < 0.05 after Benjamini–Hochberg (BH) correction.

Following the establishment of the relationship between these phenotypic characteristics and diagnosis, our next step involved exploring diffusion-weighted imaging (DWI) measurements known to be strong predictors of these biomarkers. This exploration aims to identify additional features that could enhance subsequent predictive models.

Q2.

Are WM tract features predictive of neuropsychological and neurobiological indicators?

We divided our data into training and test subsets with an 80–20 split that sought to ensure a balance of our response variables across the training/test datasets. Given the longitudinal nature of our dataset, comprising multiple MRI sessions per patient over time, we employed a stratification approach that involved ensuring that all sessions pertaining to a single patient were exclusively included in either the training or the test set, but not both. This was achieved through a randomized allocation process until no participants were found in either the training and test sets to avoid data leakage. Due to the high dimensionality of our data, it was necessary to perform aggressive feature reduction to include only the features exhibiting the top 10% highest variance. We used mean imputation across the remaining features to make it possible to employ regularization techniques for further reduction. We normalized our data; then, for each independent neuropsychological and neurobiological response variable, we employed ElasticNet [40,41] regularization with alpha ranging from 0.5 to 0.8 to reduce the number of features. The final features expected to offer the most discriminatory power for our response variables were scaled and re-imputed by using k-NN nearest neighbour imputation [42]. Imputation using k-NN was initially unable to perform complete imputation with such a wide dataset, forcing us to initially use mean imputation until we could perform ElasticNet regularization. Subsequently, for each neuropsychological and neurobiological response variable, we employed a repeated 10-fold cross-validation method to ensure robustness with various models, including support vector machines [42,43], decision trees [44], random forest [45], multi-layer perceptron [46], and gradient boosting [47]. The repeated cross-validation approach minimized variability in performance metrics due to random partitioning. We evaluated our results by using RMSE, MAE, and R2 metrics. This multi-step process is depicted in Figure 1.

Q3.

Does baseline tract volume change with cognitive decline?

Brain atrophy is a major symptom of AD observed in vivo [48,49,50]. There is compelling evidence to suggest that Aβ facilitates the spread of Tau neurofibrillary tangles, which may then drive neurodegeneration, atrophy, and subsequent dementia [20]. While almost all aged brains show characteristic changes linked to neurodegeneration [51], Alzheimer’s disease has different neurodegenerative processes compared with normal ageing, with distinctive neuron loss profiles [6]. This atrophy is understood to begin in the entorhinal cortex, progressing then to the hippocampus, temporal, frontal, and parietal areas, before spreading to the entire cerebral cortex [19,22].

DWI tractography enables the reconstruction of white matter tract bundles by estimating the principal directions of diffusion within a voxel, thereby enabling the segmenting of tracts and providing tract-specific measures such as volume, MD, and FA. Since DWI does not directly measure neurons themselves, but rather assesses the diffusion characteristics of water in tissue, these metrics represent water, not the neurons. We investigated the changes in tract volume, MD, and FA across different stages of cognitive impairment. Our analyses aimed to elucidate the relationship between these neuroimaging biomarkers and the progression from cognitively normal (CN) status to mild cognitive impairment (MCI) and Alzheimer’s disease.

Linear mixed effects models were employed to assess the effects of diagnosis and age on tract volume, MD, and FA, accounting for random effects due to individual differences. Disease-related changes were examined by using linear mixed effects models to understand the interaction of disease stages, age, tract volume, MD, FA, and tract length. We employed the lmer function from the R package lme4 [52]. Specifically, the model was formulated as $y_{ij}={\beta }_0+{\beta }_1\times x_{1ij}+{\beta }_2\times x_{2ij}+u_j+{ϵ}_{ij}$, where $y_{ij}$ represents the tractography measurement (tract volume, MD, FA, and length were each considered separately) for the $i^{th}$ observation within the $j^{th}$ subject. The predictors $x_{1ij},x_{2ij}$ represent the fixed effects disease stage and age. The term $u_j$ is the random intercept for subject $j$, assumed to follow a normal distribution $u_j~{\rm N}\left(\mathsf{O},{\sigma }_u^2\right)$. The residual error term ${ϵ}_{\mathrm{ij}}$ is also assumed to be normally distributed with ${ϵ}_{ij}~{\rm N}\left(\mathsf{O},{\sigma }^2\right)$. This modeling approach allowed us to examine the effects of the predictors while accounting for the hierarchical structure of the data and the within-subject variability.

Q4.

To what extent can tractography metrics predict cognitive decline?

Pairwise Wilcoxon Rank Sum Tests were conducted to identify significant differences in diffusion metrics across these groups, with the Benjamini–Hochberg correction [53] applied to control the false discovery rate, acknowledging the heightened risk of type I errors due to multiple comparisons. Features exhibiting a false discovery rate (FDR) corrected p-value lower than 0.05 were considered for further investigation as potential predictors in our classification models. To explore the predictive capacity of tractography metrics for cognitive decline, we segmented our dataset by current disease sub-stage as determined by physicians based on established clinical criteria. We created a cumulative distribution plot to understand the disease stages in which features exhibit significant differences. We further elucidated our understanding of which features were exhibiting significant differences by ranking the frequency of occurrence in a simple bar chart. Our expectation is that features exhibiting significant differences between disease classifications may offer predictive potential as classifiers, as we seek to develop models that indicate the potential for disease conversion towards AD.

In our total sample population of 264 participants across 434 sessions, only 14 participants converted to the next disease stage within 3 years (these participants are henceforth referred to as Converters), and only 15 converted within 10 years. Given our significant class imbalance, we divided our data into training and test subsets with a 60–40 split that sought to ensure the balance of our response variables across the training/test datasets, leaving 5 postitive classes (Converters) in our test data. We employed Adaptive Synthetic Sampling Approach for Imbalanced Learning (ADASYN), version 1.3.1, from the smotefamily package (version 1.3.1) [54] to compensate for our significant class imbalance. Subsequently, we applied a rigorous machine learning model development process involving the use of the ElasticNet grid search strategy to perform feature reduction, using conversion to the next disease stage within three years as a response variable; we varied $\alpha$ between 0 and 1.0 to identify the most parsimonious feature set for subsequent model development.

We employed a repeated 10-fold cross-validation strategy to ensure robustness with various models, including support vector machines, decision trees, random forest, and eXtreme gradient boosting. We used a grid search technique to identify the best classification model (Normal vs. Converter) and hyper-parameters predictive of cognitive decline and used Kappa as an evaluation metric for model comparison.

3. Results

Q1.

Do phenotypic characteristics predict cognitive decline within our dataset?

In our analysis of various neuropsychological and neurobiological indicators, the Kruskal–Wallis test revealed statistically significant differences across the three Alzheimer’s disease stages: cognitively normal (CN), mild cognitive impairment (MCI), and Alzheimer’s disease (AD). Significant differences were observed for all of our neuropsychological indicators and many of our neurobiological indicators, as shown in

Table 4. Differences in the distribution of neuropsychological and neurobiological indicator values across diagnostic categories (CN, MCI, and AD).

Indicator	CN	MCI	AD	Mean	IQR	${\mathit{\chi }}^2$ (Kruskal–Wallis)	p-adj
Neurophysiological Indicators
ADAS-11	139	80	40	8.9	7.2	115.39	$1.05\times {10}^{-24}$
ADAS-13	139	80	38	13.9	10.3	131.18	$4.58\times {10}^{-28}$
AV45	92	42	25	1.2	0.4	24.96	$2.29\times {10}^{-5}$
CDR SB	140	78	41	1.4	1.5	205.51	$3.78\times {10}^{-44}$
FAQ	138	75	40	3.3	3.0	149.02	$6.56\times {10}^{-32}$
FDG	5	57	30	1.2	0.1	33.05	$5.34\times {10}^{-7}$
MMSE	138	82	40	27.5	3.0	108.51	$2.74\times {10}^{-23}$
MOCA	259	82	38	24.1	6.0	128.39	$1.72\times {10}^{-27}$
RAVLT	139	80	38	40.2	20.0	115.06	$1.14\times {10}^{-24}$
Neurobiological Indicators
Brain vol	132	76	35	1,031,622.0	140,524.0	8.49	$2.86\times {10}^{-2}$
Entorhinal vol	130	77	34	3940.6	1031.0	29.08	$3.39\times {10}^{-6}$
Fusiform vol	131	77	33	17,953.6	3295.0	22.52	$5.16\times {10}^{-5}$
Hippocampus vol	133	76	32	7060.3	1531.6	60.00	$8.41\times {10}^{-13}$
ICV	136	73	37	1,468,892.6	233,267.5	1.43	$4.90\times {10}^{-1}$
Middle temporal vol	131	77	33	20,163.5	3979.0	23.70	$3.56\times {10}^{-5}$
Ventricle vol	135	75	35	40,103.4	27,051.5	21.45	$6.60\times {10}^{-5}$

p-adj refers to Benjamini–Hochberg-corrected p-values.

.

As expected [55], the Clinical Dementia Rating-Sum of Boxes (CDR-SB) scores, a comprehensive measure of dementia severity, exhibited very high chi-squared values (${\chi }^2=205.51, p<0.001$), indicating pronounced differences across disease stages, thus supporting CDR-SB’s ability to stage severity of Alzheimer dementia and mild cognitive impairment. The observation of a high chi-squared value for the CDR-SB test along with an observed long tail depicted in the density plot found in Figure 2 suggests that the MCI and AD disease stages have more variability and more extreme values than CN, signalling a potential sensitivity for early detection of conversion from CN.

Other neuropsychological assessments, such as the Alzheimer’s Disease Assessment Scale (ADAS-11 and ADAS-13), Mini-Mental State Examination (MMSE), and the Montreal Cognitive Assessment (MOCA), showed substantial discriminative power (ADAS-11: (${\chi }^2=115.39$); ADAS-13: (${\chi }^2=131.18$); MMSE: (${\chi }^2=108.51$); MOCA: (${\chi }^2=128.39$); all have p < 0.001). These results highlight the efficacy of these tests in differentiating among the stages of Alzheimer’s disease.

The Rey Auditory Verbal Learning Test (RAVLT) immediate recall scores exhibited significant differences among the diagnostic groups (CN, MCI, and AD), as evidenced by the Kruskal–Wallis test (${\chi }^2=115.06, p<0.001$). Figure 3 demonstrates how the RAVLT scores progressively decrease from cognitively normal individuals to those with MCI and further to individuals with AD.

In terms of neurobiological markers, volumes of key brain regions, such as the hippocampus, entorhinal cortex, fusiform gyrus, and the middle temporal gyrus, showed significant differences across the disease stages (hippocampus: ${\chi }^2=60.00, p<0.001$; entorhinal: ${\chi }^2=29.08, p<0.001$; fusiform: ${\chi }^2=22.52, p<0.001$; middle temporal: ${\chi }^2=23.70, p<0.001$).

Q2.

Are WM tract features predictive of neuropsychological and neurobiological indicators?

The feature reduction strategy employed by this study substantially streamlined the initial high-dimensional feature space of DWI tractography measurements to a manageable subset. From an initial 1.3 million features, we identified high variance features to reduce our dataset to 134,340 explanatory variables that were imputed and further refined by using ElasticNet regularization. The ElasticNet model’s $\alpha$ parameter, which balances L2 and L2 penalties, was optimized independently for each response variable, with values ranging from 0.5 to 0.8. This approach aimed to balance feature retention and model complexity. The final feature count for each response variable is depicted in

Table 5. Feature reduction profile of DWI tractography data.

	Features after Elasticnet Regularization
Response	$\mathit{\alpha }=0.5$	$\mathit{\alpha }=0.75$	$\mathit{\alpha }=0.8$
ADAS-11	507	295	352
ADAS-13	175	203	164
CDR-SB	353	155	40
FAQ	77	85	101
FDG	129	45	27
MMSE	12	6	9
MOCA	36	89	96
RAVLT	181	217	197
Entorhinal Vol	69	19	94
Fusiform Vol	360	267	392
Hippocampus Vol	521	319	128
Mid Temp Vol	154	202	168

.

We used repeated 10-fold cross-validation (10 iterations) to train our models for each of the neuropsychological and neurobiological indicators identified (Table 4) as being discriminative of diagnosis classification. Despite the rigorous model development process, our predictive models showed varying levels of predictive performance across different indicators (

Table 6. Best prediction models of neuropsychological and neurobiological indicators.

Indicator	Mean	IQR	Model	$\mathit{\alpha }$	RMSE	MAE	R2
Neuropsychological Indicators
ADAS-11	8.9	7.2	Random forest	0.75	5.91	4.22	0.18
ADAS-13	13.9	10.3	Random forest	0.5	8.25	6.20	0.30
AV45	1.2	0.4	Decision trees	0.5	0.20	0.15	0.15
CDR-SB	1.4	1.5	SVM Radial	0.8	2.32	1.61	0.19
FAQ	3.3	3.0	Random forest	0.8	3.53	2.76	0.66
FDG	1.2	0.1	Multilayer Perceptron	0.5	0.09	0.07	0.08
MMSE	27.5	3.0	Linear	0.8	3.72	2.83	0.12
MOCA	24.1	6.0	Gradient boosting	0.75	4.09	3.12	0.25
RAVLT	40.2	20.0	Decision trees	0.5	11.9	9.70	0.29
Neurobiological Indicators
Entorhinal Vol	3940.6	1031.0	Decision trees	0.8	864.96	659.14	0.11
Fusiform Vol	17,953.6	3295.0	SVM Radial	0.8	2352.72	1841.73	0.17
Hippocampus Vol	7060.3	1531.6	Random forest	0.8	885.20	702.23	0.46
Mid Temporal Vol	26,163.5	3979.0	Gradient boosting	0.75	2622.42	2151.96	0.23

). For AV45, an imaging biomarker for amyloid plaque accumulation (mean: 1.2; IQR: 0.4), the model yielded error metrics (RMSE: 0.2; MAE: 0.15) that were lower than the sample mean and IQR. Hippocampal volume (mean: 7060.3; IQR: 1531.6) predictions resulted in error metrics of RMSE of 885.20 and MAE of 702.23, and the model for MoCA scores (mean: 24.1; IQR: 6.0) produced error metrics of RMSE of 4.09 and MAE of 3.12, suggesting a 13% error on average in assessing a patient’s MoCA scores based on DWI MRI analysis alone. Most of our models had R² values below 0.4, indicating limited explanatory power. The model for FDG, a marker for glucose metabolism, captured the most variability in the data, with an R² of 0.66.

While we found that there is lack of comprehensive analysis using DWI-only metrics to predict these indictors, there is alignment with several studies.

A study by Patil et. al. [56] found that no strong correlation was observed for any DWI measurements in any region with respect to MMSE. This was supported by Jokinen et al. [57], who determined that white matter ADC was not predictive of poor cognitive outcomes.

Correlational analysis is the most common approach to presenting the association between DWI and clinical scores. A recent study by Saito et al. [58] consistently reported low correlations between DWI and many of the indicators we evaluated in this study.

Q3.

Does baseline WM tract volume change with cognitive decline?

Our findings (

Table 7. Estimated effects of diagnosis and age on MD, FA, and tract volume using linear mixed effects models.

Variable	Estimate	Std. Error	df	t-Value	p-adj
Mean diffusivity
Intercept	$1.41\times {10}^{-3}$	$3.39\times {10}^{-5}$	$2.83\times {10}^2$	41.55	$2.08\times {10}^{-122}$
Diagnosis (MCI)	$-3.36\times {10}^{-5}$	$8.16\times {10}^{-7}$	$1.12\times {10}^7$	−41.19	$2.00\times {10}^{-16}$
Diagnosis (AD)	$-3.90\times {10}^{-5}$	$1.29\times {10}^{-6}$	$1.12\times {10}^7$	−30.17	$8.52\times {10}^{-200}$
Age	$-6.13\times {10}^{-6}$	$9.42\times {10}^{-8}$	$1.10\times {10}^7$	−65.05	$2.00\times {10}^{-16}$
Fractional anisotropy
Intercept	$3.94\times {10}^{-1}$	$2.79\times {10}^{-3}$	$1.30\times {10}^3$	141.16	$2.00\times {10}^{-16}$
Diagnosis (MCI)	$-2.64\times {10}^{-3}$	$2.44\times {10}^{-4}$	$7.20\times {10}^6$	−10.80	$6.75\times {10}^{-27}$
Diagnosis (AD)	$-1.97\times {10}^{-3}$	$3.86\times {10}^{-4}$	$5.35\times {10}^6$	−5.08	$3.69\times {10}^{-7}$
Age	$-2.93\times {10}^{-4}$	$2.80\times {10}^{-5}$	$1.05\times {10}^6$	−10.47	$1.60\times {10}^{-25}$
Tract volume
Intercept	$2.73\times {10}^3$	$2.19\times {10}^2$	$5.69\times {10}^3$	12.48	$1.07\times {10}^{-34}$
Diagnosis (MCI)	$2.78\times {10}^1$	$2.38\times {10}^1$	$1.00\times {10}^6$	1.17	$2.43\times {10}^{-1}$
Diagnosis (AD)	$-1.22\times {10}^2$	$3.76\times {10}^1$	$5.35\times {10}^5$	$-3.24$	$1.60\times {10}^{-3}$
Age	$2.51\times {10}^1$	$0.26\times {10}^1$	$6.76\times {10}^4$	9.37	$1.57\times {10}^{-20}$
Tract length
Intercept	$1.62\times {10}^1$	$0.19\times {10}^1$	$1.78\times {10}^3$	8.41	$8.36\times {10}^{-17}$
Diagnosis (MCI)	$0.20\times {10}^1$	$1.79\times {10}^{-1}$	$5.54\times {10}^6$	10.87	$2.27\times {10}^{-27}$
Diagnosis (AD)	$-0.36\times {10}^1$	$2.83\times {10}^{-1}$	$3.72\times {10}^6$	$-12.98$	$3.23\times {10}^{-38}$
Age	$6.35\times {10}^{-1}$	$2.05\times {10}^{-2}$	$6.01\times {10}^5$	30.97	$27.21\times {10}^{-210}$

p-adj: linear mixed effects model p-values were estimated based on Satterthwaite’s approximation, and subsequently FDR-corrected.

) show significant decreases in both MD and FA across both the MCI and AD stages compared with CN, even after adjusting for age, with tract volume analyses revealing significant increases with the progression of cognitive decline towards the later stages of the disease (AD), after adjusting for age. There was not a significant difference in tract volume between the CN and MCI stages.

The MCI group is associated with an increase in mean tract length compared with the CN group, holding age constant.

Q4.

To what extent can tractography metrics predict cognitive decline?

Among the 1.3 million tract measurements assessed, 5394 tract measurements (0.3%) exhibited statistically significant differences among groups after performing pairwise Wilcoxon Rank Sum Tests with Benjamini–Hochberg correction [53] (p < 0.05). In the cumulative distribution shown in Figure 4, we note the curve deviation of the statistics from the null distribution increases modestly as participants transition into later stages of cognitive impairment, suggesting tract anomalies may be more pronounced in later stages of the disease. The curve deviations of the comparison of mild cognitive impairment (MCI) to late mild cognitive impairment (LMCI) (blue) are larger than other groups, suggesting potentially higher effect sizes [59].

Figure 5 presents a frequency plot of measurement types identified as significant during the transition across the MCI → LMCI → AD cognitive impairment stages. The plot ranks measurement types by their frequency of occurrence, emphasizing which measurements are most prevalent in highlighting tract anomalies associated with cognitive decline. The higher-frequency measurements relate to the morphometrics of detected tracts (length and volume) as well as differences in MD as expected [60].

We employed four machine learning algorithms, including support vector machine, random forest, XGBoost, and MARS, to predict progression from a normal disease stage to a “Converter” status, indicative of advancement to a more severe disease stage within three years of imaging. These models were evaluated by using a repeated 10-fold cross-validation technique to ensure the reliability and stability of our predictions.

All four models exhibited nearly identical performance metrics across the evaluation scheme. The accuracy for each model was observed to be between 0.4836 and 0.4985, with a 95% confidence interval ranging from 42.89% to 55.33%.

3.1. Comparison with Similar Studies

Despite many studies claiming that diffusion metrics offer potential for prognostic biomarkers of AD (

Table 8. Comparative analysis of our proposed model with other models applied to the ADNI dataset.

Authors	Highlights	Participants	Performance
Classification of current disease stage
Chen (2023) [61]	The study investigated white matter alterations in the Alzheimer’s continuum by using diffusion tensor imaging, finding widespread changes correlated with Tau pathology, particularly in the cingulum, which may serve as a promising biomarker for preclinical Alzheimer’s disease.	236 ADNI3 subjects (176 CN, 36 MCI, and 24 AD)	74% Acc, 69% AUC, 58% Sens, and 78% Spec
Chen (2023) [62]	A model that enhances multi-modal AD diagnosis by using orthogonal latent space learning, feature weighting, and graph learning to improve discriminative information retention and relationship encoding among samples.	757 ADNI2 subjects (283 CN, 330 MCI, and 144)	67% Acc, 69% Sens, 64% Spe, and 71% AUC
Deng (2023) [63]	The Fully Connected Multi-Kernel Convolutional Neural Network model accurately diagnoses Alzheimer’s disease and mild cognitive impairment from diffusion tensor imaging (DTI) data while also generating fiber probability maps to assist in clinical diagnosis.	413 subjects (162 CN, 130 MCI, and 121 AD)	96% Acc, 97% Sens, 100% Spec, and 98% Auc
Khan (2022) [64]	Developed a 3-tiered cognitive hybrid machine learning algorithm for disease prediction.	818 ADNI1 subjects (229 CN, 396 MCI, and 193 AD)	95% Acc, 95% Sens 97% Spe, and 99% Auc
Razzak (2022) [65]	Proposes an integrative deep ensemble learning framework called PartialNet, tailored for Alzheimer’s detection using brain MRIs, demonstrating improved predictive performance and efficiency compared with DenseNet, with notable gains in both multiclass and binary class AD detection on benchmark datasets.	350 subjects (95 AD, 146 MCI, and 95 CN)	98% Acc (mean of CN, MCI, and AD)
Hazarika (2022) [66]	The study discusses various deep learning models for Alzheimer’s disease classification, highlighting DenseNet-121’s strong performance and computational inefficiency, and proposes a modified DenseNet-121 with depth-wise convolutions.	210 (70 CN, 70 MCI, and 70 AD)	98% Acc (mean of CN, MCI, and AD)
Prediction of future conversion
Stone (2021) [67]	This study identified diffusivity measures from specific white matter tracts, particularly axial diffusivity, by using only DTI data.	87 subjects: 34 Converted and 53 Not converted	72% Acc and 67% AUC
Velazquez (2022) [68]	Prediction of conversion from mild cognitive impairment (MCI) to AD using DTI data with clinical variables from health records.	384 subjects: 49 Converted and 335 Not converted	98% Acc and AUC 99%

), many of these studies highlight significant effect sizes but do not actually attempt prediction. Most studies that classify current disease stages involve the consolidation of clinical indicators, health record data, and data from multiple imaging modalities to achieve maximum accuracy. Our analysis has determined that clinical variables alone are sufficient to achieve an averaged balanced accuracy of 88%, a specificity of 92%, and a sensitivity of 84% with our current dataset. The addition of individual tract-specific diffusion data contributed very little to our models (accuracy of 88%, specificity of 93%, and sensitivity of 85%). When tract-specific measurements are used in isolation, DWI data appear to offer relatively weak performance when classifying disease stages of AD.

3.2. Adaptions to Methodology

After reflecting on the results obtained while answering Questions 1–4 in our research methodology, it was important to consider alternative mechanisms for understanding the probability with which a participant will exhibit further disease progression. We found limited evidence that tractography metrics could be useful to predict established neuropsychological and neurobiological biomarkers (Table 6), and we found considerable evidence by using linear mixed effects modeling that tractography metrics exhibit an effect on disease stages while accounting for age (Table 7).

Given the high dimensionality of our data and probable loss of meaning resulting from aggressive feature selection, we considered an approach whereby we condensed our features into aggregate measures at the seed level to provide a representative feature as an alternative to feature reduction strategies. Formula (1) represents the Z-score for a given region measurement of a participant’s session.

$$Z_{ijkl}={\displaystyle \frac{X_{ijkl }-{\mu }_k}{{\sigma }_k}}$$

(1)

where ${\mathrm{x}}_{ijkl}$ is the $k^{\mathrm{th}}$ measurement (e.g., those identified in Figure 5) for the $i^{\mathrm{th}}$ participant in the $j^{\mathrm{th}}$ session from the $l^{\mathrm{th}}$ seed region targeting a specific region. ${\mu }_{kl}$ is the mean of the $k^{\mathrm{th}}$ measurement, and ${\sigma }_k$ represents the standard deviation for the $k^{\mathrm{th}}$ measurement.

After scaling the measurements, we determined the mean Z-score for each participant, session, and ROI (Formula (2)). A value at this aggregated level gives insights into how a participant’s ROI may be different from the same region in other participants. For example, mean tract length of all tracts connected to the paracentral cortex for a given participant’s session.

$${\overline{Z}}_{ijk}={\displaystyle \frac{1}{M}}{\displaystyle \sum }_{l=1}^M{Z}_{ijkl}$$

(2)

With a more condensed set of features to move forward with, we continued to rely on an ElasticNet grid search to perform feature selection, resulting in a reduction from 2994 features to 2, relying on an $\alpha$ value of 0.6. These two features identified were the features representing the hemispheric lateralization of mean tract volume for tracts originating from the supramarginal and paracentral regions.

Relying on a dataset with only two features representing these diffusion-derived anatomical measurements, we applied several machine learning models to predict the classification of individuals into two groups: Normal (no conversion) and those who would convert to Alzheimer’s within three years. The models tested included eXtreme gradient boosting (XGB), multivariate adaptive regression splines (MARS), support vector machine (SVM), and random forest (RF), with results presented in

Table 9. Best prediction models of Normal vs. Converter within 3 years.

Model	AUROC	AUPRC	F1	Accuracy	Sensitivity	Specificity	Kappa	p-adj
RF 1	0.74	0.75	0.64	0.71	0.52	0.90	0.42	$1.11\times {10}^{-14}$
XGBTree 2	0.77	0.72	0.57	0.67	0.44	0.90	0.34	$3.78\times {10}^{-10}$
MARS 3	0.64	0.56	0.50	0.61	0.48	0.74	0.26	$7.20\times {10}^{-5}$
SVM 4	0.54	0.50	0.24	0.50	0.16	0.84	−0.00	$0.05\times {10}^1$

¹ Random forest mtry = c(2, floor(sqrt(num_features)), floor(num_features/3)) ² eXtreme gradient boosting nrounds = seq(from = 25, to = 100, by = 25); max_depth = seq(from = 5, to = 35, by = 10); eta = seq(from = 0.2, to = 1, by = 0.2); gamma = seq(from = 1, to = 10, by = 1); colsample_bytree = seq(from = 0.6, to = 1, by = 0.2); min_child_weight = seq(from = 2, to = 5, by = 1); subsample = 1 ³ Multivariate adaptive regression splines; degree = seq(from = 1, to = 3, by = 1); nprune = seq(from = 1, to = 10, by = 1) ⁴ Support vector machine; sigma = (0.001,0.01,0.1,1,10,100); C = (0.001,0.01,0.1,1,10,100).

.

Overall, the random forest model performed the best, achieving an Area Under the Receiver Operating Characteristic (AUROC) of 0.74, indicating a good ability to differentiate between “Converter” and “Normal” classes. The Area Under the Precision–Recall Curve (AUPRC) was 0.75, reflecting a strong performance in capturing the “Converter” class, which is particularly important given the class imbalance of our original data (see Figure 6). The F1-score was 0.64, providing a harmonic mean of precision and recall, and the overall accuracy of predicting the correct class was 71%. The model’s ability to detect “Converters” (true positive) has room for improvement, with a sensitivity of 0.52. The specificity was much better, 0.90, suggesting a strong ability to identify Normal subjects. This was expected given the large class imbalance. The model exhibited a Cohen’s Kappa of 0.42, indicating that there is moderate agreement between the predicted classifications and the actual classifications. These results are consistent with other studies who attempted to predict future disease conversion by using DWI-only data (Table 8).

To enhance the predictive power of this model, we integrated these two diffusion metrics with traditional clinical variables, including MMSE, MoCA, RAVLT, CDR-SB, FAQ, hippocampal volume, entorhinal volume, and Aβ and Tau indicators. This hybrid model achieved a significant improvement, yielding an accuracy of 86%, a sensitivity of 86%, and a specificity of 93%. These results surpass the diagnostic performance of current clinical assessments, where the sensitivity ranges from 70.9% to 87.3% and the specificity from 44.3% to 70.8% [69]. Our findings emphasize the value of a hybrid machine learning approach that combines advanced neuroimaging techniques with conventional clinical assessments.

4. Discussion

Alzheimer’s disease is a complex and widespread [20] neurodegenerative disease that manifests itself in multiple ways [6,70], including atrophy of the hippocampus, entorhinal region, and middle temporal regions, as well as accumulation of Aβ and Tau proteins. Tractography measurements derived from diffusion-weighted images appear to show limited potential as a single scan test capable of offering predictions for many of the traditional neuropsychlogical and neurobiological assessment metrics used today; however, they may still contribute as an important tool to a comprehensive approach to understanding the complexity of AD and characterization of disease staging for some of these clinical variables.

Assessing and diagnosing Alzheimer’s disease remain complex and challenging due to the lack of a complete model that can identify the disease in any stage. Currently, diagnosis often relies heavily on subjective assessments derived from neuropsychological and neurobiological tests carried out in primary care settings. These tests, while valuable, can be influenced by various factors, such as the examiner’s expertise, the time of day when the test is administered, the testing environment, and the patient’s physical and emotional state at the time of testing [71].

The inherent variability and subjectivity in these evaluations can lead to inconsistent diagnoses, particularly in the early or preclinical stages of Alzheimer’s disease, where symptoms may be subtle or overlap with other conditions, such as ageing. Clinic pathological studies have shown that the diagnostic sensitivity of clinicians is between 70.9% and 87.3% and the specificity is between 44.3% and 70.8% [69]. Additionally, traditional diagnostic methods may fail to capture the full spectrum of neuropathological changes associated with Alzheimer’s, limiting their effectiveness in early detection and intervention. In light of this, the aim of our study was to elucidate the contribution that diffusion-weighted imaging can make to improving model development towards the early detection of Alzheimer’s. Machine learning models hold significant promise for improving the evaluation of Alzheimer’s disease by offering more objective, accurate, and scalable diagnostic tools. Machine learning can uncover subtle and complex relationships within the data that may not be apparent through traditional methods, potentially leading to earlier and more accurate diagnoses, a better monitoring of disease progression, and personalized treatment plans.

4.1. Novel Contributions

Our study identified the hemispheric lateralization of tract volumes connected to the supramarginal gyrus and paracentral regions as a potential prognostic biomarker of AD disease. The supramarginal gyrus is part of the parietal lobe and plays a role in language perception and processing [72], spatial orientation and tool use [73], emotion recognition [74], writing and word recognition [75], and the integration of sensory information [76]. This region’s association with AD and dementia has been reported in the literature [77,78,79,80], typically in later stages of the disease. In the most recent study referenced, the authors used magnetoencephalography to reveal that decreased beta-band intensity in the left supramarginal gyrus is associated with decreased neuropsychological assessment scores and increased clinical severity of cognitive impairment, suggesting its importance in assessing cognitive status. Our findings support this observation by providing complimentary evidence that changes in volume asymmetries of tracts connecting the supramarginal gyrus may be associated with cognitive impairment and dementia. The deterioration of white matter tracts connecting the supramarginal gyrus may lead to reduced efficiency of the default mode network, to which the supramarginal gyrus belongs.

While our final model for identifying “Converters” leaves some room for improvement, overall, the model demonstrates commendable ability in distinguishing Converter from Normal subjects, particularly by achieving good AUROC and AUPRC scores based on only two features. However, the moderate sensitivity and low specificity suggest that there is more work to do in terms of enhanced feature selection, alternative model development, or the fine tuning of the current models. This analysis could be expanded by integrating DWI and resting state functional MRI data around the regions implicated in our model. Existing resting state studies indicate that significant differences in signal intensity exist for the same regions our models use [81,82]. Graph theoretical metrics could also be included in this analysis to understand if changes in nodal efficiency across different disease stages could also strengthen our model. There do appear to be imaging data available for the ADNI3 cohort that calculate the network failure quotient from resting state functional MRI images, which may encompass these two potential enhancements to our analysis.

In addition, the application of proven machine learning algorithms with a consolidated dataset of whole-brain tractography, phenotypic, and neuropsychological data for early biomarker identification in Alzheimer’s disease (AD) represents a thorough and integrative approach. While tractography-focused predictive analytics has been widely used in neuroscience research, comprehensive whole-brain analyses on longitudinal data remain relatively rare. Our study reinforces the utility of this approach by demonstrating its application in a comprehensive longitudinal dataset. Utilizing well-established machine learning techniques in combination with exhaustive tractography and neuropsychological data provides a robust methodology for investigating early biomarkers of AD. Our findings add to the existing body of evidence elucidating the potential of diffusion MRI as a tool for the early detection and monitoring of neurodegenerative diseases, highlighting the importance of integrating multi-modal data for enhanced predictive analytics.

Our findings confirm the significant predictive value of existing neurobiological and neuropsychlogical biomarkers in detecting Alzheimer’s disease. The biomarkers identified in Table 8 demonstrate that while some biomarkers may be more effective in different stages of the disease, they collectively provide a robust toolbox for disease detection. The RAVLT is of particular interest, given that it is one of the earliest indicators of conversion from CN to MCI [8]. Our findings highlight the pronounced impact of Alzheimer’s disease on memory function. Notably, the RAVLT is recognized as a critical diagnostic tool, particularly due to its sensitivity in detecting early memory deficits that often signify the transition from mild cognitive impairment (MCI) to Alzheimer’s disease (AD) [8]. Furthermore, Figure 3 provides a visual depiction of these differences, illustrating a clear distinction among the CN, MCI, and AD groups. This separation is indicative of the progressive nature of memory impairment in AD pathology.

Our analysis of white matter tract features revealed moderate but notable associations with neurobiological and neuropsychological markers. Our results based on predicting cognitive test scores indicate some potential for relying on DWI-based MRI to non-subjectively assess cognitive progression in AD. The results demonstrate an MAE of 3.12 for the MoCA, a test which is on a scale of 0 to 30, implying a 13% error on average in assessing a patient’s cognitive outcomes based on DWI MRI analysis alone. This is an interesting finding, which implies that one day, we may be able to create predictive technologies informed by MRI that may be able to accurately predict a patient’s cognitive test scores. Many patients with AD despise taking cognitive tests, implying that technologies developed to monitor their disease progression may be a welcome development in AD patient management.

A significant finding of our study is the relationship between tract characteristics and cognitive decline (Table 7). Our results suggest that the brain might be undergoing specific microstructural changes that both restrict diffusion (lower MD) and disrupt the coherence of white matter tracts (lower FA), while at the same time, the increases in tract volume during later disease stages might reflect underlying processes, such as the cellular proliferation of astrocytes [70,83], or changes in the extracellular matrix [84], potentially confounding volume measurement, although further studies are needed to explore this. Conceivably, an increase in volume may be symptomatic of inflammation leading to edema, which could increase the extracellular space. A corresponding increase in MD would have supported this hypothesis [85]; however, that was not observed in our data.

Our results highlight that changes in tract length may offer a useful biomarker for disease staging. A significant increase ($p=2.266\times {10}^{-27}$) suggests that on average, MCI diagnosis is associated with a longer tract length than observed in CN participants, whereas the AD group is associated with a decrease in a mean tract length while holding age constant, which is consistent with expectations of neurodegeneration leading to tract deterioration. There is support in the existing literature of increased tract length with age [86], and future research may explore the potential that early or mild stages of cognitive decline could trigger compensatory mechanisms [87] in the brain, potentially leading to an increase in tract length as the brain attempts to maintain connectivity.

Our attempts to leverage tract-specific measurements from diffusion-weighted images that were correlated with existing neurobiological and neuropsychlogical biomarkers as a means to identify cognitive impairment were initially unsuccessful. Our model accuracy aligned with the No Information Rating, indicating that our models’ predictions were no better than random chance. Further, the Cohen’s Kappa statistic for each model was 0, reflecting the absence of agreement beyond chance between the predicted outcomes and the actual disease progression status.

There are several practical implications from these findings. The identification of tract volume asymmetries in the supramarginal gyrus and paracentral regions offers a nascent but promising potential prognostic biomarker as a non-invasive method for the detection and monitoring of disease progression. If sufficiently advanced, this approach could be integrated into routine clinical practice, providing clinicians with a valuable tool to assess disease progression and inform treatment plans.

4.2. Limitations

Limited sample sizes in neuroimaging studies can compromise the reliability and validity of the findings reported [88]. Small sample sizes reduce the statistical power, increasing the likelihood of Type 1 (false positive) and Type 2 (false negative) errors. While the original ADNI-3 cohort is larger than many studies (960 subjects across 6050 scans), many subjects lacked diffusion-weighted imaging data. Our study included 264 participants with a limited number of scan sessions per participant (between one and five).

We acknowledge the potential source of error resulting from the smoothing effects of interpolation as a result of registering DWI images to T1. Our pipeline strategy was initially developed for a large dataset of noisy clinical data [26,29], where it was determined after many approaches that registration of DWI to T1-weighted images before ODF reconstruction offered the most reliable alignment approach with the highest number of successful registrations. It was felt that lower rates of successful registration were highly undesirable and could potentially skew the results of analyses more so than the error associated with the simple smoothing that results from the interpolation process.

We relied on single-shell DWI, where diffusion measurements are acquired with a single b-value. Single-shell DWI has been shown to underperform in resolving complex fiber configurations within a voxel compared with multi-shell DWI [89]. The limitations of single-shell DWI mean that our study might not fully capture the complexity of white matter architecture, particularly in regions where multiple fiber pathways intersect [90]. This can potentially lead to the mischaracterization of fiber tract integrity and connectivity, especially with respect to the tract length measurement we considered in our study. Consequently, our findings regarding tract length might be less reliable than if we had access to more sensitive multi-shell images that offer opportunities for improved delineation of crossing fibers.

Given the breadth of our data, we relied on aggressive feature reduction strategies to make machine learning feasible. This included only considering the features with the highest variability (top 10%) after removing features with low variance or highly correlated redundant features. This may have resulted in the exclusion of potentially important data that were never introduced during model training. As a consequence, there may be significant characteristics within our data that could have enhanced model performance but were excluded early in the process.

The results of our adapted machine learning strategy are promising, though they offer room for improvement. While it is encouraging to achieve this classification accuracy with only two measures based on the hemispheric lateralization of mean tract volume for tracts originating from the supramarginal and paracentral regions, there are specific limitations that should be addressed. Our highly imbalanced proportion of Converter to Normal participants (14/434) necessitated a reliance on synthetic data to better balance for reliable predictions. This deficiency is likely a contributing factor to our low sensitivity scores, as synthetic data do not perfectly capture the complexity and variability of real-world data. The limited sample size restricts the statistical power and generalizability of our findings. More data would enhance the model training process of our ML models, allowing for better feature learning and reducing the risk of overfitting. In particular, increasing the number of individuals who exhibit progressive disease pathology would help provide a more balanced dataset.

4.3. Conclusions

Overall, our results align with the existing literature on the neurodegenerative patterns characteristic of Alzheimer’s disease [5,91]. The observed microstructural changes and their impact on cognitive function highlight the importance of integrating advanced neuroimaging techniques with traditional neuropsychological assessments. Future research should focus on refining these predictive models, exploring additional biomarkers, and validating our findings in larger, more balanced cohorts to enhance the robustness and generalizability of our conclusions.