[Paper] Amyloid-β prediction machine learning model using source-based morphometry across neurocognitive disorders (2024)

18 Apr 2024     #bio #brainImaging #demensia #atn #amyloid

Momota, Yuki, et al. “Amyloid-β prediction machine learning model using source-based morphometry across neurocognitive disorders.” Scientific Reports 14.1 (2024): 7633.

Paper Link

Points

Objective

• Investigated MRI-based machine learning models to predict Alzheimer’s disease (AD), with a focus on diverse patient populations.
• Utilized source-based morphometry (SBM) to assess Amyloid-beta deposition.

Methodology

• Preprocessed 3D T1 weighted images into voxel-based gray matter (GM) images, then subjected them to SBM.
• Implemented a support vector machine (SVM) as a classifier.
• Employed SHapley ADditive exPlanations (SHAP) for model interpretability and accountability.

Results

• Achieved a final model accuracy of 89.8% when incorporating MR images, cognitive test results, and apolipoprotein E status.
• Attained an 84.7% accuracy with the model based solely on MR images.

Background

• AD is a neurodegenerative disorder characterized by the presentce of A$\beta$ plaques, neurofibrillary tangles, and brain atrophy.
• A$\beta$ is a defining characteristics of AD, but detecting it is not covenient in routine clinical practice.
  • Methods such as position emission tomography (PET), cerebrospinal fluid (CSF) testing, and Blood biomarkers are used for A$\beta$ detection but are not yet applicable in routine clinical practice.
• MRI-based A$\beta$ prediction may serve as a useful screening tool before definitive diagnosis through the aforementioned methods.

Method

Features

Participants and clinical measurements

• Recruited in Jury 2018 ~ August 2021 from the memory clinic at Keio University Hospital.
• AD / MCI / HC

Cognitive assessment (9 measures)

• Global cognitive function: Mimi-mental state examination (MMSE), Clinical dementia rating (CDR), Functional activity questionnaire (FAQ)
• Memory: Wechsler Memory Scale-Revised (WMS-R) Lgical Memeory immediate recall (LM I) and delayed recall (LM II)
• Executive function and attention: Word Fluency, Trail Making Test (TMT)
• Specific cognitive abilities: Japanese version of Alzheimer’s Disease Assessment Scale-Cognitive subscale (ADAS-cog-J), Japanese Adult Reading Test (JART)

Apolipoprotein E (APOE) genotyping

• Magnetic nanoparticle DNA extraction kit (EX1 DNA Blodd 200 $\mu$L Kit)
• real-time polymerase chain reaction (PCR)

[¹⁸F] Florbetaben (FBB) amyloid-PET imaging

• [¹⁸F] Florbetaben (FBB)

  Florbetaben, a fluorine-18 (18F)-labeled stilbene derivative (formerly known as BAY-949172), trade name NeuraCeq, is a diagnostic radiotracer developed for routine clinical application to visualize β-amyloid plaques in the brain. [reference]

MRI

Acquisition - 3D T1 weighted MR images (T1 WI)

• MRI scanner: Discovery MR750 3.0 T scanner (GE Healthcare)
• Coil: 32-channel head coil
• Imaging parameters: field of view (FOV) 230mm, matrix size 256$\times$256, slice thickness 1.0mm, voxel size 0.9$\times$0.9$\times$1.0mm

Pre-processing

1. Segmentation: Segmented the MR images into different tissue types: gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) using Statistical Parametric Mapping toolbox CAT12.

2. Normalization: The Segmented GM images are then normalized to the Montreal Neurological Institute (MNI) template, which is a standard anatomical template commonly used in neuroimaging research.

   Standard anatomical templates are widely used in human neuroimaging processing pipelines to facilitate group level analyses and comparisons across different subjects and populations. The MNI-ICBM152 template is the most commonly used standard template, representing an average of 152 healthy young adult brains. [reference]

3. Resampling and Smoothing: Resampled the images to an isotropic voxel size of 2$\times$2$\times$2mm³ and smoothed using a 5mm full-width-at-half-maximum Gaussian kernel.
   • This step helps to standardize the voxel size an reduce noise in the images.

4. Source-based morphometry (SBM): Incorporates independent component analysis (ICA) to automatically decompose the anatomical brain images into independent spatial maps characterizing different modes of anatomical variability accorss all individuals.

   In signal processing, independent component analysis (ICA) is a computational method for separating a multivariate signal into additive subcomponents. This is done by assuming that at most one subcomponent is Gaussian and that the subcomponents are statistically independent from each other. [reference]

   

5. ICA processing: The 3D GM images (91$\times$109$\times$91 voxels) are loaded and converted into a 1D array format (1$\times$902,629) for processing.
   • A brain mask is created to select relevant (208,082) voxels for ICA using FastICA.
   • The number of extracted independent components (ICs) is a hyperparameter that is tuned for subsequent model building.

6. Spatial Regression: The extracted ICs are used as spatial regressors for each participant’s GM images, with weighting coefficients ($\beta$) determining the effect of each IC on the GM image.

   \[I_{GM}=\beta_1 IC_1 + \beta_2 IC_2 + ... + \beta_K IC_K\]

Machine learning

• Input features: ICA’s $\beta$-values, demographic characteristics (age and sex), cognitive assessments, APOE genotype
• Input conduction: The model is trained and tested using various combinations of input features.
  1. All input features together
  2. Each combination of features: brain images alone, brain images + cognitive assessments, etc.
  3. Different combination of diagnoses: AD+HC, AD+MCI+HC
• Model: Gaussian kernel support vector machine (SVM)
  • Training involves classification using 5-vold cross-validation.
  • Testing is performed over all splits (5 times), ensuring robust evaluation.
• Interpretability: SHaply Additive exPlanations (SHAP)
  • SHAP values, based on game theory, indicate the influence of features on predictions.
  • Features with large absolute SHAP values have a strong influence on predictions.
  • Clinical features with positive and negative SHAP values were associated with A$\beta$+ and A$\beta$- conditions, respectively

Statistical analysis

Explores relationships between variables, identifying associations with diagnoses, and testing hypotheses in the context of Alzheimer’s disease research.

• Two-tailed t-test / Chi-square test
  • Two-tailed t-test: used to compare the means of two groups to determine if there is a significant difference between them.
  • Chi-square test: used to test the independence between categorical variables.
• Relationships among features: Pearson’s correlation analysis for continous variables
  • Measures the strength and direction of linear relationships between pairs of continuous variables.
  • Provides insights into how variables are related to each other.
• Associations with diagnoses: Analysis of variance (ANOVA)
  • Used to anaylze the difference among group menas in a sample.
  • Useful when there’re more than two groups being compared, as it determines whether there are statistically significant differences among the group means.

Results

118 cases used for the final model building

Model performance

A$\beta$ positivity prediction

• The final model: the model trained with brain images + cognition + APOE as input
• The highest accuracy (89.8%) and AUC (0.888) with brain images + cognition + APOE
• The lowest accuracy (84.7%) and AUC (0.830) with brain images alone

The final model’s performance for predicting A$\beta$ positivity in each diagnosis

• The highest accuracy (89.8%) when including all the paticipants
• The lowest accuarcy (75.9%) based solely on MCI

SBM

7 independent components (ICs) were derived from the final SBM model

• Each component showed spatially maximally independent GM volum patters.
• IC 1 showed a significant correlation with cognitive measures an A$\beta$ positivity.
• Only AD and IC 1 showed a significant association.
• Other diagnoses were not associated with any ICs.

Discussion

The proposed model predicted A$\beta$ positivity successfully. (accuracy 89.8%, AUC 0.888)

• With 118 participants’ data consisting of the features: brain MRI, cognitive info., genetic info.
• Predicted correctly in non-AD subjects, such as those with FTLD syndrome and psychiatric disorders.
• Among covariants in the final model, IC 1 had the strongest impact realted to A$\beta$ positivity prediction.

Performance

1. Informative heterogeneity of features among non-AD participants
   • The performance of the model based only on AD continuum achieved slightly lower (88.4%) than on all cases.
2. Advantages of SBM
   • The model based on diverse clinical populations may be better suited for application in clinical settings.
     • Patients visiting physicians’ would have various neurocongitive disorders beyond the AD continuum.
   • The proposed model based only on brain images (accuracy 84.7%) may assist for screening of potential candidates for AD-related clinical trials.
   • SBM detects subtle morphological changes and unknown patterns in brain structures associated with ND diseases without relying on existing atlases.
3. Comparable prediction performance in MCI patiences
   • Surpassed the accuracy of the physician’s clinical diagnosis of AD (75.9% > 70%)

Feature Importance of the model - SHAP

All ICs demonstrated greater importance compared to demoghrapic and cognitive features such as MMSE. The three most influential features in the model were identified as follows: IC 1, logical memory (LM) I, and LM II.

• IC 1 exhibited a significantly correlation with A$\beta$ positivity and cognivite measures.
  • Its spacial pattern of the loading coefficients closely resembled the cortical pattern observed in neurodegeneration (ND) in AD, particularly in the parietal lobe.
  • No Medial temporal lobe (MTL) atrophy was observed in any IC, which is the typical AD pattern.
    • This discrepance suggests a potential indication of tau pathodology rather than A$\beta$ pathology.
• LM scores reflected memory impairments, a cardinal symptom of AD.
• The presence of APOE -$\epsilon#4 also emerged as a significant factor.

Furthermore, the model revealed distinct associations between IC 1 and A$\beta$ positivity, as well as IC 4 and age.

• This indicates the model’s ability to discriminate between AD-related ND from normal aging in brain imaging, suggesting that the pathological process of AD is not strictly age-dependent. → Brain atrophy patterns in normal aging processes can be distinguished from those in neurodegeneartive disease.

Limitation

1. A$\beta$ positivity was determined only by amyloid-PET scan: CSF A$\beta$ would be a more sensitive marker in the pre-clinical status.
2. a limited number of samples: could be affect accuracy of a machine learning model.
3. Longitudinal follow-up data might improve model performance, rather than a cross-sectional approach.