Combining Blood-Based Biomarkers and Structural MRI Measurements to Distinguish Persons with and without Significant Amyloid Plaques

Abstract

Background:

Amyloid-β (Aβ) plaques play a pivotal role in Alzheimer’s disease. The current positron emission tomography (PET) is expensive and limited in availability. In contrast, blood-based biomarkers (BBBMs) show potential for characterizing Aβ plaques more affordably. We have previously proposed an MRI-based hippocampal morphometry measure to be an indicator of Aβ plaques.

Objective:

To develop and validate an integrated model to predict brain amyloid PET positivity combining MRI feature and plasma Aβ_42/40 ratio.

Methods:

We extracted hippocampal multivariate morphometry statistics from MR images and together with plasma Aβ_42/40 trained a random forest classifier to perform a binary classification of participant brain amyloid PET positivity. We evaluated the model performance using two distinct cohorts, one from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) and the other from the Banner Alzheimer’s Institute (BAI), including prediction accuracy, precision, recall rate, F1 score, and AUC score.

Results:

Results from ADNI (mean age 72.6, Aβ+ rate 49.5%) and BAI (mean age 66.2, Aβ+ rate 36.9%) datasets revealed the integrated multimodal (IMM) model’s superior performance over unimodal models. The IMM model achieved prediction accuracies of 0.86 in ADNI and 0.92 in BAI, surpassing unimodal models based solely on structural MRI (0.81 and 0.87) or plasma Aβ_42/40 (0.73 and 0.81) predictors.

CONCLUSIONS:

Our IMM model, combining MRI and BBBM data, offers a highly accurate approach to predict brain amyloid PET positivity. This innovative multiplex biomarker strategy presents an accessible and cost-effective avenue for advancing Alzheimer’s disease diagnostics, leveraging diverse pathologic features related to Aβ plaques and structural MRI.

Keywords

Aβ positivity amyloid PET blood-based biomarkers image features MRI

INTRODUCTION

Clinical psychiatrist and neuroanatomist Alois Alzheimer first reported the symptoms of a “pre-senile” and thought to be relatively rare dementia associated with abnormal brain plaque and neurofibrillary emergence and aggregation in November 1906 [1]. However, over that last century, Alzheimer’s disease (AD) has been recognized as a major global health concern that is primarily causing and contributing to cognitive impairment and dementia worldwide with increasing burden and costs [2, 3]. While no cure for AD exists, recent therapeutic advances allow for utilization of both symptomatic and disease-modifying treatments across the AD clinical spectrum to slow clinical progression, thus making the timely diagnosis of AD more important than ever [4 –9].

Brain imaging modalities currently serve as primary methods for assessment of dementia and AD-related brain abnormalities, including for the presence of amyloid and tau pathology [10 –12]. Among them, magnetic resonance imaging (MRI) detects brain shrinkage [13], while positron emission tomography (PET) tracers can measure amyloid-plaque and tau neurofibrillary tangle burden, respectively [12 , 15]. Although PET scans have demonstrated high accuracy in detecting AD-related pathological changes, their costs, radiation exposure, and insurance coverage status restrict their widespread clinical use and can also limit their use in clinical trials [8, 12]. As a less invasive, less costly, and more readily available modality that is considered part of the standard clinical diagnostic pathway [6, 16], MRI has the potential to characterize brain atrophy patterns and changes over time [13]; and can also characterize other useful structural (and functional) features beyond basic volumetric measures.

We have previously developed a series of advanced structural MRI methods and demonstrated their potential advantages in AD diagnostics including multivariate morphometry statistics (MMS) [17] which uses multivariate tensor-based morphometry (mTBM) to encode surface morphometry along the surface tangent direction and radial distance (RD) to encode morphometry along the surface normal direction. This approach outperformed volumetric measures in detecting hippocampal deformation and also exhibited superior performance in determining group differences between AD diagnosis groups [18, 19]. To address the high dimensionality problem of this method in classification tasks, we developed patch analysis-based surface sparse-coding and max-pooling (PASS-MP) system to extract patch-based image features and learn a low-dimensional representation of hippocampal MMS that are predictive of longitudinal cognitive decline [20]. Furthermore, the objective function was modified to improve the sparse coding system by integrating the correntropy based method [21].

Some studies have suggested that it may be feasible to impute amyloid positivity from MRI, potentially providing a lower cost and more accessible alternative compared to more accurate but costly PET scans [22]. For example, Duygu et al. used combined structural MRI and cerebral blood flow patterns as predictors of amyloid positivity status and achieved 83% classification accuracy in early mild cognitive impairment (MCI) individuals [22]. Leveraging the advanced MR analysis techniques our team has previously developed [17 –21], we also demonstrated the feasibility of potentially predicting amyloid burden based on MR data [23].

Recently, there is an increased interest and success in the development of blood-based biomarkers (BBBMs) as research and diagnostic tools for the characterization of amyloid and tau pathologies [24 –26]. In a recent review of using BBBMs in detecting amyloid positivity [27], the highest prediction accuracy using Aβ_42/40 as the predictor was 0.77 [28]. Another study of using plasma Aβ_42/40 ratio as a biomarker for amyloidosis reported an area under the ROC curve (AUC-ROC) of 0.81 (accuracy 75%). The performance was further improved after controlling for cohort heterogeneity (AUC-ROC 0.86 and accuracy 81%) and adding APOE4 copy number and participant age (AUC-ROC 0.90 and accuracy 86%) indicating a high diagnostic accuracy in detection of brain amyloidosis [29]. Also, a recently proposed method combining MRI with BBBMs for longitudinal structural atrophy and cognitive decline prediction demonstrated reliable performance in distinguishing fast vs slow progressors in MCI [30].

In this study, we propose a novel pipeline that integrates imaging features from MRI with an AD BBBM to predict PET-based brain Aβ positivity. We hypothesized that by combining information from structural MRI and an Aβ BBBM in an integrated multimodal (and multiplex) model prediction of amyloid PET positivity would be improved compared to performance by each unimodal model.

MATERIALS AND METHODS

Data used in the preparation of this article was obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (http://adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private partnership led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial MRI, PET, other biological markers, and clinical and neuropsychological assessments can be combined to measure the progression of MCI and early AD. For up-to-date information, see http://www.adni-info.org.

Data acquisition and preprocessing

In this study, we identified and downloaded data in 198 subjects reported in a previous study (Table 1) [31] from the ADNI database with matching plasma Aβ_42/40 measures, florbetapir PET, and T1-MRI scans. In addition, we obtained 111 subjects from Banner Alzheimer’s Institute (BAI) with [¹¹C]-Pittsburgh compound B (PiB) PET, T1-MRI, and plasma Aβ_42/40 measures [29]. Plasma Aβ_42/40 ratios in ADNI and BAI datasets were both quantified using a LC-MS based assay [29, 31]. AUC scores of 0.89 in the ADNI cohort and 0.81 in the BAI cohort were reported respectively by the previous studies [29, 31], indicating promising and consistent performance of this assay to detect brain amyloid positivity. Image analysis was done using previously established methods [32, 33]. Briefly, T1-MRI were processed with FreeSurfer, PET images were motion corrected, summed, and co-registered with MRI, regional standard uptake value ratio (SUVR) with and without partial volume correction was calculated using cerebellar cortex as the reference region, and target region for global index of amyloid burden (mean cortical SUVR, MCSUVR) included frontal, parietal, temporal, and precuneus regions [32, 33]. The Centiloid procedure was adopted to harmonize PET derived amyloid burden measure between the two studies using the two different tracers [34]. According to this procedure, as we previously established, the equivalent amyloid positivity thresholds adopted were MCSUVR > 1.42 for PiB and MCSUVR > 1.19 for florbetapir, with the positivity threshold translated to approximately 16 Centiloid unit [35, 36]. Based on this criterion, 91 subjects were amyloid positive and 107 were amyloid negative for the ADNI cohort, and 41 were Aβ positive and 70 were Aβ negative for the BAI cohort. An overview of the participant summary statistics can be seen in Table 1.

Table 1

Summary statistics of the patients

	ADNI (N = 198)	BAI (N = 111)
Age
Min	55.7	47.0
Max	88.4	85.0
Mean±SD	72.62±6.92	66.21±7.95
Sex
Male	95 (48%)	27 (24.3%)
Female	103 (52%)	84 (75.7%)
Diagnosis
CN	106 (53.5%)	111 (100%)
MCI	92 (46.5%)	0 (0%)
APOE4
NC	122 (61.6%)	55 (49.5%)
HT	68 (34.3%)	35 (31.5%)
HM	8 (4.04%)	21 (18.9%)
Aβ positivity
Aβ+	91 (49.5%)	41 (36.9%)
Aβ–	107 (50.5%)	70 (63.1%)

System overview

Our prediction pipeline can be divided into two parts: 1) Hippocampal structure segmentation and registration: The hippocampal structures were segmented and registered from structural MRI (sMRI). Vertex-wise MMS features were computed for both hemispheres of the brain. These MMS features consist of Radial Distance (RD) and multivariate tensor-based morphometry (mTBM), measuring hippocampal size and the deformation within the surface, respectively; 2) Patch-based feature extraction and classification: Patches were selected from the hippocampal surfaces and sparse codes were generated using the correntropy-induced sparse coding method, followed by max-pooling to reduce the dimensionality. The resulting feature vectors were input into a random forest classifier both by themselves and combining with plasma Aβ_42/40 features.

By employing this pipeline, we aimed to leverage the unique information provided by the hippocampal MMS features and the discriminative power of the correntropy-induced sparse coding method to enhance the predictive performance. Additionally, we investigated the potential improvement achieved by integrating the image-based features with the plasma Aβ_42/40 levels.

Hippocampus segmentation and surface registration

We took advantage of our previously established pipeline protocol for hippocampus segmentation, image registration and hippocampal surface reconstruction [14 , 38]. We also performed surface smoothing to reduce the noise, followed by mesh simplification and refinement using the same method as our previous work [39 –41]. A conformal grid of size 150 * 100 was computed on each surface using a holomorphic 1-form basis [42, 43], followed by the computation of conformal factor and mean curvature on each point and a surface fluid registration method to register the hippocampal surfaces to a template surface [44 –46] For more details, refer to [44].

Surface multivariate morphometry statistics

After the surface registration was completed, we computed RD and mTBM as surface features. The RD measures the hippocampal size by the surface differences along the surface normal direction, while mTBM measures the local surface deformation and has a relatively stronger signal detection power [39 , 48], resulting in a feature vector of size 4 * 1 for each of the 15000 vertices on each hemisphere (Fig. 1).

Fig. 1

Framework for predicting brain amyloid positivity with blood-based marker plasma Aβ_42/40 and MMS-based PASCS-MP representation. Penal (1) shows MMS are extracted from MRI scans. In Penal (2), we generated representations with PASCS-MP and used them with Plasma Aβ_42/40 as predictors to classify the patients with different brain amyloid positivity based on a random forest classifier.

Surface feature dimensionality reduction

To resolve the high-dimension small sample problem, we took advantage of our previously developed Patch Analysis-based Surface Correntropy-induced Sparce Coding and Max-pooling (PASCS-MP) method to generate a representation of the subject with reasonable feature size [23]. From each hemisphere, we selected 504 patches of sizes 10 * 10, resulting in 1008 patches from each subject [39, 49]. Then, we learned a sparse coding representation for each patch by stochastic coordinate coding algorithm, resulting in a 1008(patches) * 1800(features) dimension representation for each subject as the input to max-pooling process.

We borrowed the Correntropy-induced Sparse-coding (SC) from our previous work, which improves the Stochastic Coordinate Coding (SCC) method [50]. In traditional dictionary learning approach, we learn a dictionary D ∈ R^p×m that reconstructs the inputs X ∈ R^p×n from the sparce codes Z ∈ R^m×n, such that the distance between the original input X and the learned representation Z is minimized, subject to regularization on the sparsity of D: $min_{D, z_{i}} \frac{1}{2} \sum_{i = 1}^{n} ∥ {Dz}_{i} - x_{i} ∥_{2}^{2} + λ \sum_{i = 1}^{n} ∥ z_{i} ∥_{1}$ In correntropy measure, however, the similarity function can be replaced by the expectation of Gaussian kernel: $K_{σ} (u, v) = \exp {- {(u - v)}^{2} / σ^{2}},$ resulting in a correntropy-based loss function: $min_{D, z_{i}} \frac{1}{2} \sum_{i = 1}^{n} \exp (- \frac{∥ {Dz}_{i} - x_{i} ∥_{2}^{2}}{σ^{2}}) + λ \sum_{i = 1}^{n} ∥ z_{i} ∥_{1}$ Furthermore, our recent study reformulated the first part of the equation by the half-quadratic technique and introduced an auxiliary variable h_i for each input x_i that can be customized according to different requirements. Intuitively, h_i is a Gaussian kernel of the distance between the original and the reconstructed input. The final objective function becomes: $min_{D, z_{i}} \frac{1}{2} \sum_{i = 1}^{n} h_{i} ∥ {Dz}_{i} - x_{i} ∥_{2}^{2} + λ \sum_{i = 1}^{n} ∥ z_{i} ∥_{1}$ Where $h_{i} = \exp (- \frac{∥ {Dz}_{i} - x_{i} ∥_{2}^{2}}{σ^{2}})$ .

To solve the minimization problem, for each image patch, we update the dictionary D, the latent representation Z and the auxiliary variable h in turn, while keeping the other variables fixed (Algorithm 1). For details of the mathematical induction and proofs, refer to our previous study [23].

Max-pooling and classification

With the computed 1008 * 1800-dimension representation from PASCS process, we performed max-pooling [20, 51] on each feature across all the 1008 patches to keep the maximum feature values. By this process, we scaled down the dimensions of the feature vectors from 1008 * 1800 to 1 * 1800, which is a more manageable size to apply traditional machine learning algorithms.

After getting the pooled image features, we separately trained three classifiers to predict the positivity of brain Aβ. First, we fit a logistic regression model of brain Aβ vs plasma Aβ_42/40. The reason we used logistic regression here instead of random forest for other predictors is that plasma Aβ_42/40 is a sole predictor, which is unsuitable for decision tree-based methods. Second, we trained a random forest classifier on the MMS-based features extracted by PASCS-MP pipeline described earlier. As a third step, another random forest classifier was trained using the joint features of plasma Aβ_42/40 and extracted MMS. For each experiment, we performed a 10-fold cross-validation five times, and evaluated the performance based on the average accuracy, precision, recall rate and F1 score. Both logistic regression and random forest classifiers were implemented by scikit-learn package [52]. In addition, due to the imbalanced property of the BAI data, we adopt the SMOTE upsampling method to balance the positive and negative sample numbers [53].

Performance evaluation protocol

Based on our previous study, the performance of PASCS-MP pipeline relies on four key parameters: the patch size, the dimensionality of the learned representation, the regularization parameter for the l1-norm (λ) and the kernel size σ. We have previously determined that the optimal parameter values for patch size and dimensionality are 10 * 10 and 1800, respectively, which was verified using both ADNI and OASIS datasets [23]. In this study, we verify that the same patch size and dimensionality are still optimal for the new dataset, but the best λ and σ values for our dataset shifted to 0.4 and 2.0, respectively.

To evaluate the model performance using different feature combinations, we adopt 10-fold cross-validation by random shuffling the data samples and splitting them into ten sub-sets. We use one subset as a testing set and the other 9 subsets as training set. After classification, we combine the classification results to form a result set that contains all samples and generate the confusion matrix and calculate the accuracy, precision, recall rate and F1 score. We also compute the area-under-the-curve (AUC) of the receiver operating characteristics (ROC). We applied the same approach and parameter settings on both datasets.

RESULTS

Parameter settings

As determined in our prior work, the best parameters for our model were sparse code dimensionality = 1800 and patch size = 10 * 10, regularization parameter λ= 0.22, kernel size σ= 3.6. In this work, we adopted the same dimensionality and patch size, but we sought to identify the best combination of λ and σ for our datasets, which respectively control the sparsity of the learned representation and the properties of correntropy. We examined the classification accuracy of the learned random forest classifier with 1000 nodes based solely on MMS features. 5-fold cross-validation was applied when searching for each of the optimal parameters. The optimal parameter values were λ = 0.2, σ = 2.0 in ADNI dataset, while the best combination became λ = 0.4, σ = 2.0 in BAI dataset (Table 2). For the sake of generalization and consistency across the two datasets, we decided to use λ = 0.4, σ = 2.0 for both datasets in the subsequent experiments. Furthermore, we performed an additional search among different number of nodes in the random forest classifier and determined that the optimal number of nodes is 1000 (Table 3).

Table 2

Performance comparison of different combinations of λ and σ

Dataset	λ	σ	Accuracy
ADNI	0.2	2	0.87
	0.2	4	0.78
	0.22	3.6	0.77
	0.3	2	0.85
	0.3	4	0.77
	0.4	2	0.83
	0.4	4	0.82
BAI	0.2	2	0.75
	0.2	4	0.75
	0.22	3.6	0.81
	0.3	2	0.77
	0.3	4	0.73
	0.4	2	0.88
	0.4	4	0.79

Table 3

Performance comparison of different numbers of random forest nodes in combined predictor

Dataset	Nodes	Accuracy	Precision	Recall	F1 score
ADNI	100	0.77±0.01	0.77±0.02	0.76±0.03	0.76±0.02
	500	0.79±0.01	0.80±0.02	0.80±0.02	0.80±0.01
	1000	0.82± 0.02	0.82± 0.02	0.81± 0.02	0.81± 0.01
	2000	0.81±0.01	0.81±0.01	0.80±0.01	0.80±0.01
BAI	100	0.89±0.02	0.91±0.02	0.86±0.02	0.88±0.02
	500	0.90±0.01	0.92±0.01	0.86±0.01	0.88±0.01
	1000	0.91± 0.01	0.93± 0.01	0.89± 0.02	0.90± 0.01
	2000	0.88±0.02	0.92±0.02	0.85±0.03	0.87±0.03

Performance comparison result

Accuracy, precision, recall rate, F1 score and ROC-AUC

We evaluated the performance of the three models on amyloid positivity prediction: the logistic regression model using blood Aβ_42/40, the random forest model using surface MMS features, and the random forest model using the combined blood Aβ_42/40 and surface MMS features. The classification accuracies using only plasma Aβ_42/40 are 0.73 in ADNI dataset and 0.81 in BAI dataset, respectively. The accuracies using only surface MMS features are 0.81 in ADNI dataset and 0.87 in BAI dataset, respectively. By integrating MMS features and plasma Aβ_42/40 measurements, the model accuracy improved to 0.85 and 0.90 in ADNI and BAI datasets, respectively, outperforming either baseline models (Table 4). We also calculated the AUC scores of different feature combinations and plotted the ROC curve. As shown in Fig. 2, the multimodal models outperform the individual models, yielding AUC scores > 0.94 in ADNI dataset and > 0.98 in BAI dataset respectively. We performed Delong test to evaluate the improvement of AUC by incorporating imaging features, APOE genotype, age, and sex into our model. As a result, the integrated model achieved a significant improvement compared with plasma Aβ_42/40 predictor (p-value = 0.0024 in ADNI and 0.0021 in BAI), but the improvement was not statistically significant over the image-only model (p = 0.59 in ADNI and 0.99 in BAI). In addition, the models with or without covariates yield similar performance (p = 0.28–0.95 in both datasets). Moreover, we evaluated the model with highest classification accuracy, i.e., MMS + Aβ_42/40 + APOE mode, on different diagnosis groups in ADNI dataset (106 CN with 57 Aβ–and 49 Aβ+; 92 MCI with 43 Aβ–and 49 Aβ+). The model performance was stable across diagnosis groups (Accuracy = 0.90 in CN group and 0.89 in MCI group).

Fig. 2

ROC curves of classification models with different feature combinations. a) ROC curves of ADNI dataset. b) ROC curves of BAI dataset. Incorporated models with image MMS and plasma Aβ_42/40 features outperforms the individual models using only image MMS features or plasma Aβ_42/40 as predictors. Adding covariates slightly improves the AUC score.

Table 4

Performance comparison of different predictors

Dataset	Feature	Accuracy	Precision	Recall	F1 score
ADNI	Plasma Aβ_42/40 (Logistic)	0.73±0.01	0.73±0.01	0.73±0.01	0.73±0.01
	MMS (RF)	0.81±0.01	0.81±0.01	0.80±0.01	0.81±0.01
	MMS + Aβ_42/40 (RF)	0.85±0.01	0.84±0.01	0.84±0.01	0.84±0.01
	MMS + Aβ4_2/40 + Age (RF)	0.84±0.01	0.84±0.01	0.83±0.02	0.83±0.01
	MMS + Aβ _42/40 + APOE (RF)	0.86±0.01	0.85±0.02	0.84±0.02	0.85±0.01
	MMS + Aβ_42/40 + Age + APOE (RF)	0.83±0.01	0.85±0.02	0.85±0.02	0.85±0.02
	MMS + Aβ_42/40 + Age + Sex + APOE (RF)	0.85±0.01	0.85±0.01	0.85±0.01	0.85±0.01
BAI	Plasma Aβ_42/40 (Logistic)	0.81±0.01	0.80±0.01	0.81±0.01	0.80±0.01
	MMS (RF)	0.87±0.01	0.90±0.01	0.83±0.02	0.85±0.02
	MMS + Aβ_42/40 (RF)	0.90±0.02	0.92±0.01	0.87±0.02	0.89±0.02
	MMS + Aβ_42/40 + Age (RF)	0.90±0.01	0.92±0.01	0.87±0.01	0.89±0.01
	MMS + Aβ_42/40 + APOE (RF)	0.92±0.01	0.91±0.02	0.86±0.03	0.87±0.02
	MMS + Aβ_42/40 + Age + APOE (RF)	0.88±0.01	0.92±0.01	0.87±0.02	0.88±0.02
	MMS + Aβ_42/40 + Age + Sex + APOE (RF)	0.89±0.01	0.92±0.01	0.87±0.01	0.89±0.02

We also evaluated the influence of individual demographic features, including age and sex, by incorporating them into the logistic regression model and testing for significance. As a result, both age and sex showed significance (p < 0.05) in the univariate models in ADNI, while sex was not significant in BAI dataset (Table 5). Therefore, we evaluated age and sex as covariates in our combined models. A complete list of all combinations and corresponding performance are listed in Table 4. Overall, we observed that incorporating age slightly reduced the standard deviation of the performance statistics, but the classification accuracies were not substantially influenced.

Table 5

Univariate logistic regression of potential covariates

	p (ADNI)	p (BAI)
Age	0.004^**	0.002^**
Sex	0.036^*	0.168

Significance Code: 0: ‘^***’; 0.001: ‘^**’; 0.01: ‘^*’; 0.05: ‘.’.

DISCUSSION

In this study, we assessed a novel approach for detecting brain amyloid-PET positivity using an integrated multimodal model that combines an MR image feature and a plasma Aβ_42/40 BBBM. To the best of our knowledge, this is the first study that combines this type of advanced morphometric brain imaging information and amyloid BBBMs to assess amyloid-PET positivity. Our previously developed PASCS-MP method for geometric feature extraction successfully generated features that achieved accuracy of 0.81±0.01 in ADNI dataset and 0.87±0.01 in BAI dataset with a random forest classifier based on MR data alone. With the integration of plasma Aβ_42/40 measures, we further improved the accuracy to 0.86±0.01 and 0.92±0.01 in the two datasets respectively, surpassing both baseline unimodal models that used only a single AD-related feature. This result verified our hypothesis that the MMI classification model that integrated MRI image features and plasma biomarkers can achieve superior performance compared to each individual unimodal model. The slight difference in accuracy for plasma Aβ_42/40 assay alone observed in this study compared to the original studies [29, 31] is likely due to the differences in imaging analysis methods and the determination of PET based amyloid positivity.

In our analysis, although univariate logistic regression showed a significant p-value for age, adding age or sex alone as covariates to the IMM model did not significantly improve accuracy. Adding APOE genotype information to the IMM model improved performance, suggesting that APOE genotype can provide additional predictive power beyond our structural MR feature and plasma Aβ_42/40. Further investigations with larger and more diverse datasets would provide additional insights into the potential utility of including these covariates in predicting brain amyloid-PET positivity.

A study limitation involved the moderate sample size (198 for the ADNI dataset and 111 for the BAI dataset) which has the potential to introduce dataset-specific parameters and result in overfitted models. To mitigate this limitation, we made an effort to avoid selecting separate λ and σ model parameter values for each dataset. Instead, we picked a shared set of λ and σ that achieved high prediction accuracy across the two datasets in our preliminary parameter search (Table 2). Even so, the parameter values differ from the determined optimal values in our previous study using larger datasets [23]. In addition, the model has limited generalizability across cohorts, as reflected by low testing AUC (<0.65) when trained using either cohort and tested with the other and distinct dictionaries for the two cohorts learned by the algorithm. We assume the main cause of this issue is the diversity of clinical features of the composite of the two cohorts. For example, diagnosis status is usually strongly correlated with brain amyloid status. However, no MCI or AD sample is available in the BAI cohort. In addition, the effect of distinct sex distribution in the two cohorts has not been thoroughly evaluated in this study due to time limitation and data availability. With an increased cohort size in future, we will be able to conduct an in-depth investigation into this issue through population-based analysis and causal inference techniques, and a more stable and generalizable model can be expected. Another limitation of the study is that we used a specific plasma Aβ-related assay while there are now a range of other BBBMs, including plasma tau phosphorylated at different sites (pTau181, pTau217, pTau231), that show promise for predicting AD-cerebral amyloidosis [28 , 55]. For example, for AD progression prediction, Tau217 and pTau181, combined with memory, executive function, and APOE, achieved high prediction accuracy in BioFINDER (AUC = 0.91) and ADNI (AUC = 0.90) cohorts, respectively [56]. Further studies are warranted to examine the combined performance of our MR based approach together with other BBBMs in the detection of amyloid status across the AD continuum.

There were active researches that used MRI biomarkers to predict brain amyloid burden [22 , 57–63]. Among different brain morphometry features, hippocampal features were adopted in a number of research [22 , 61], including ours [23, 64]. The prior research demonstrated hippocampal measures are reasonable MRI biomarkers for brain amyloid identification. However, we agree that other cortical measures will also be useful for this line of research, and it is our ongoing commitment to study other cortical measures for brain amyloid prediction.

In summary, these results suggest the potential of advanced structural MRI features providing similar accuracy to plasma Aβ_42/40 in classifying amyloid-PET positivity, and support the promise of integrating such advanced structural MR features with blood-based biomarkers along with demographic information and APOE4 status can further improve amyloid-PET predictive accuracy. These multimodal and multiplex integrated biomarker approaches hold great promise and merit further investigation to aid improvements in developing and utilizing more efficient and accurate AD biomarkers for clinical and research use, and to advance therapeutic and preventative strategies.

AUTHOR CONTRIBUTIONS

Yanxi Chen (Conceptualization; Formal analysis; Methodology; Software; Validation; Visualization; Writing – original draft; Writing – review & editing); Yi Su (Conceptualization; Formal analysis; Investigation; Methodology; Project administration; Resources; Writing – review & editing); Jianfeng Wu (Methodology; Software); Kewei Chen (Conceptualization; Formal analysis; Methodology; Validation; Writing – review & editing); Alireza Atri (Conceptualization; Writing – review & editing); Richard J. Caselli (Conceptualization; Funding acquisition; Resources; Validation; Writing – original draft); Eric M. Reiman (Conceptualization; Funding acquisition; Resources; Supervision; Writing – review & editing); Yalin Wang (Conceptualization; Formal analysis; Funding acquisition; Investigation; Methodology; Project administration; Resources; Software; Supervision; Validation; Writing – original draft; Writing – review & editing).

Footnotes

ACKNOWLEDGMENTS

The last author would like to thank Dr. Duygu Tosun-Turgut of UC San Francisco for helpful discussion during the early phase of this project. 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: Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; BioClinica, Inc.; Biogen Idec Inc.; Bristol-Myers Squibb Company; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; GE Healthcare; Innogenetics, N.V.; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Medpace, Inc.; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Synarc Inc.; and Takeda Pharmaceutical Company. 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 (). The grantee organization is the Northern California Institute for Research and Education, and the study is coordinated by the Alzheimer’s Disease Cooperative Study at the University of California, San Diego. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.

FUNDING

Algorithm development and image analysis for this study were partially supported by the National Institute on Aging (R21AG065942, R01AG069453, RF1AG073424, and P30AG072980), the National Eye Institute (R01EY032125), the National Institute of Dental & Craniofacial Research (R01DE030286), and the State of Arizona via the Arizona Alzheimer Consortium.

CONFLICT OF INTEREST

Yalin Wang and Richard J. Caselli are both Editorial Board Members of this journal but were not involved in the peer-review process of this article nor had access to any information regarding its peer-review. Alireza Atri has received consulting-related honoraria or travel in the past 10 years from AbbVie, Acadia, Allergan, Axovant, AZ Therapies, Biogen, Eisai, Grifols, JOMDD, Lundbeck, Merck, Roche/Genentech, Novo Nordisk, Prothena, Qynapse, Sunovion, Suven, and Synexus. Eric M. Reiman is a co-founder and advisor for ALZPath, a company involved in the development of blood-based biomarkers for Alzheimer’s disease, including plasma PTau217. The remaining authors have no conflict of interest to report.

DATA AVAILABILITY

Source code of this paper is available on github: .

Data used in this study are available from the corresponding author, Yalin Wang, upon reasonable request.

The supplementary material is available in the electronic version of this article: .

References

Hippius

, Neundörfer

(2023) The discovery of Alzheimer’s diseas. Dialogues Clin Neurosci 5, 101–108.

Qiu

, Kivipelto

, Von

Strauss E

(2009) Epidemiology of Alzheimer’s disease: Occurrence, determinants, and strategies toward intervention . Dialogues Clin Neurosci 11 ,111–128.

El-Hayek

, Wiley

, Khoury

, Daya

, Ballard

, Evans

, Karran

, Molinuevo

, Norton

, Atri

(2019) Tip of the iceberg: Assessing the global socioeconomic costs of Alzheimer’s disease and related dementias and strategic implications for stakeholders . J Alzheimers Dis 70,323–341.

Rasmussen

, Langerman

(2019) Alzheimer’s disease–why we need early diagnosis. Degener Neurol Neuromuscul Dis 9, 123.

Atri

(2019) Current and future treatments in Alzheimer’s disease. . Semin Neurol 39, 227–240.

Liss

, Seleri Assunção

, Cummings

, Atri

, Geldmacher

, Candela

, Devanand

, Fillit

, Susman

, Mintzer

, Bittner

, Brunton

, Kerwin

, Jackson

, Small

, Grossberg

, Clevenger

, Cotter

, Stefanacci

, Wise-Brown

, Sabbagh

(2021) Practical recommendations for timely, accurate diagnosis of symptomatic Alzheimer’s disease (MCI and dementia) in primary care: A review and synthesis. J Intern Med 290 , 310–334.

Angioni

, Delrieu

, Hansson

, Fillit

, Aisen

, Cummings

, Sims

, Braunstein

, Sabbagh

, Bittner

, Pontecorvo

, Bozeat

, Dage

, Largent

, Mattke

, Correa

, Gutierrez Robledo

, Baldivieso

, Willis

, Atri

, Bateman

, Ousset

, Vellas

, Weiner

(2022) Blood biomarkers from research use to clinical practice: What must be done? A report from the EU/US CTAD Task Force. J Prev Alzheimers Dis 9, 569–579.

Schindler

, Atri

(2023) The role of cerebrospinal fluid and other biomarker modalities in the Alzheimer’s disease diagnostic revolution. Nat Aging 3, 460–462.

Reiman

(2023) Drug trial for Alzheimer’s disease is a game changer. Nature 615, 42–43

10.

Reiman

, Jagust

(2012) Brain imaging in the study of Alzheimer’s disease. Neuroimage 61, 505–516.

11.

Montagne

, Nation

, Pa

, Sweeney

, Toga

, Zlokovic

(2016) Brain imaging of neurovascular dysfunction in Alzheimer’s disease. Acta Neuropathol 131, 687–707.

12.

Dickerson

, Atri

(2020) Molecular imaging biomarkers in dementia: Amyloid and tau PET imaging aids evaluation of patients suspected of having Alzheimer disease or other dementias . Pract Neurol (Fort Wash Pa) 19, 34.

13.

Dickerson

, Bakkour

, Salat

, Feczko

, Pacheco

, Greve

, Grodstein

, Wright

, Blacker

, Rosas

, Sperling

, Atri

, Growdon

, Hyman

, Morris

, Fischl

, Buckner

(2009) The cortical signature of Alzheimer’s disease: Regionally specific cortical thinning relates to symptom severity in very mild to mild AD dementia and is detectable in asymptomatic amyloid-positive individuals . Cereb Cortex 19, 497–510.

14.

Chételat

, Arbizu

, Barthel

, Garibotto

, Law

, Morbelli

, van de Giessen

, Agosta

, Barkhof

, Brooks

, Carrillo

, Dubois

, Fjell

, Frisoni

, Hansson

, Herholz

, Hutton

, Jack

Jr , Lammertsma

, Landau

, Minoshima

, Nobili

, Nordberg

, Ossenkoppele

, Oyen

WJG

, Perani

, Rabinovici

, Scheltens

, Villemagne

, Zetterberg

, Drzezga

(2020) Amyloid-PET and 18F-FDG-PET in the diagnostic investigation of Alzheimer’s disease and other dementias . Lancet Neurol 19, 951–962.

15.

Ossenkoppele

, Schonhaut

, Schöll

, Lockhart

, Ayakta

, Baker

, O’Neil

, Janabi

, Lazaris

, Cantwell

, Vogel

, Santos

, Miller

, Bettcher

, Vossel

, Kramer

, Gorno-Tempini

, Miller

, Jagust

, Rabinovici

(2016) Tau PET patterns mirror clinical and neuroanatomical variability in Alzheimer’s disease . Brain 139, 1551–1567.

16.

Atri

(2019) The Alzheimer’s disease clinical: Diagnosis and management . Med Clin 103, 263–293 spectrum.

17.

Wang

, Chan

, Toga

, Thompson

(2009) Multivariate tensor-based brain anatomical surface morphometry via holomorphic one-forms. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2009:12th International Conference, London, UK, September 20–24, 2009, Proceedings, Part I 12, pp. 337-344.

18.

Dong

, Zhang

, Wu

, Li

, Schron

, McMahon

, Shi

, Gutman

, Chen

, Baxter

, Thompson

, Reiman

, Caselli

, Wang

(2019) Applying surface-based hippocampal morphometry to study APOE-E4 allele dose effects in cognitively unimpaired subjects. Neuroimage Clin 22 , 101744.

19.

Dong

, Zhang

, Stonnington

, Wu

, Gutman

, Chen

, Su

, Baxter

, Thompson

, Reiman

, Caselli

, Wang

(2020) Applying surface-based morphometry to study ventricular abnormalities of cognitively unimpaired subjects prior to clinically significant memory decline. Neuroimage Clin 27 , 102338.

20.

Dong

, Zhang

, Li

, Wang

, Leporé

, Thompson

, Caselli

, Ye

, Wang

; Alzheimer’s Disease Neuroimaging Initiative (2020) Integrating convolutional neural networks and multi-task dictionary learning for cognitive decline prediction with longitudinal images. J Alzheimers Dis 75, 971–992.

21.

Gui

, Sun

, Ji

, Tao

, Tan

(2016) Feature selection based on structured sparsity: A comprehensive study. IEEE Trans Neural Netw Learn Syst 28, 1490–1507.

22.

Tosun

, Joshi

, Weiner

; the Alzheimer’s Disease Neuroimaging Initiative (2014) Multimodal MRI-based imputation of the Aβ+ in early mild cognitive impairment . Ann Clin Transl Neurol 1, 160–170.

23.

, Dong

, Gui

, Zhang

, Su

, Chen

, Thompson

, Caselli

, Reiman

, Ye

, Wang

(2021) Predicting brain amyloid using multivariate morphometry statistics, sparse coding, and correntropy: Validation in 1,101 individuals from the ADNI and OASIS databases. Front Neurosci 15, 669595.

24.

Teunissen

, Verberk

IMW

, Thijssen

, Vermunt

, Hansson

, Zetterberg

, van der Flier

, Mielke

, Del Campo

(2022) Blood-based biomarkers for Alzheimer’s disease: Towards clinical implementation. Lancet Neurol 21, 66–77.

25.

de Wolf

, Ghanbari

, Licher

, McRae-McKee

, Gras

, Weverling

, Wermeling

, Sedaghat

, Ikram

, Waziry

, Koudstaal

, Klap

, Kostense

, Hofman

, Anderson

, Goudsmit

, Ikram

(2020) Plasma tau, neurofilament light chain and amyloid-β levels and risk of dementia; a population-based cohort study. Brain 143, 1220–1232.

26.

Pereira

, Janelidze

, Smith

, Mattsson-Carlgren

, Palmqvist

, Teunissen

, Zetterberg

, Stomrud

, Ashton

, Blennow

, Hansson

(2021) Plasma GFAP is an early marker of amyloid-β but not tau pathology in Alzheimer’s disease. Brain 144, 3505–3516.

27.

Cianflone

, Coppola

, Mirabelli

, Salvatore

(2021) Predictive accuracy of blood-derived biomarkers for amyloid-β brain deposition along with the Alzheimer’s disease continuum: A systematic review . J Alzheimers Dis 84, 393–407.

28.

Janelidze

, Mattsson

, Palmqvist

, Smith

, Beach

, Serrano

, Chai

, Proctor

, Eichenlaub

, Zetterberg

, Blennow

, Reiman

, Stomrud

, Dage

, Hansson

(2020) Plasma P-tau181 in Alzheimer’s disease: Relationship to other biomarkers, differential diagnosis, neuropathology and longitudinal progression to Alzheimer’s dementia. Nat Med 26, 379–386.

29.

West

, Kirmess

, Meyer

, Holubasch

, Knapik

, Hu

, Contois

, Jackson

, Harpstrite

, Bateman

, Holtzman

, Verghese

, Fogelman

, Braunstein

, Yarasheski

(2021) A blood-based diagnostic test incorporating plasma Aβ42/40 ratio, ApoE proteotype, and age accurately identifies brain amyloid status: Findings from a multi cohort validity analysis. Mol Neurodegener 16, 30.

30.

Xie

, Das

, Wisse

LEM

, Ittyerah

, de Flores

, Shaw

, Yushkevich

, Wolk

; Alzheimer’s Disease Neuroimaging Initiative (2023) Baseline structural MRI and plasma biomarkers predict longitudinal structural atrophy and cognitive decline in early Alzheimer’s disease. Alzheimers Res Ther 15, 79.

31.

Ovod

, Ramsey

, Mawuenyega

, Bollinger

, Hicks

, Schneider

, Sullivan

, Paumier

, Holtzman

, Morris

, Benzinger

, Fagan

, Patterson

, Bateman

(2017) Amyloid β concentrations and stable isotope labeling kinetics of human plasma specific to central nervous system amyloidosis. Alzheimers Dement 13, 841–849.

32.

, D’Angelo

, Vlassenko

, Zhou

, Snyder

, Marcus

, Blazey

, Christensen

, Vora

, Morris

, Mintun

, Benzinger

(2013) Quantitative analysis of PiB-PET with freesurfer ROIs. PLoS One 8, e73377.

33.

, Blazey

, Snyder

, Raichle

, Hornbeck

, Aldea

, Morris

, Benzinger

TLS

(2015) Quantitative amyloid imaging using image-derived arterial input function. PLoS One 10, e0122920.

34.

Klunk

, Koeppe

, Price

, Benzinger

, Devous

Sr , Jagust

, Johnson

, Mathis

, Minhas

, Pontecorvo

, Rowe

, Skovronsky

, Mintun

(2015) The Centiloid Project: Standardizing quantitative amyloid plaque estimation by PET. Alzheimers Dement 11, 115.e14–4.

35.

, Flores

, Wang

, Hornbeck

, Speidel

, Joseph-Mathurin

, Vlassenko

, Gordon

, Koeppe

, Klunk

, Jack

Jr , Farlow

, Salloway

, Snider

, Berman

, Roberson

, Brosch

, Jimenez-Velazques

, van Dyc

, Galasko

, Yuan

, Jayadev

, Honig

, Gauthier

, Hsiung

, Masellis

, Brooks

, Fulham

, Clarnette

, Masters

, Wallon

, Hannequin

, Dubois

, Pariente

, Sanchez-Valle

, Mummery

, Ringman

, Bottlaender

, Klein

, Milosavljevic-Ristic

, McDade

, Xiong

, Morris

, Bateman

, Benzinger

TLS

(2019) Comparison of Pittsburgh compound B and florbetapir in cross-sectional and longitudinal studies. Alzheimers Dement (Amst) 11, 180–190.

36.

, Flores

, Hornbeck

, Speidel

, Vlassenko

, Gordon

, Koeppe

, Klunk

, Xiong

, Morris

, Benzinger

TLS

(2018) Utilizing the Centiloid scale in cross-sectional and longitudinal PiB PET studies. Neuroimage Clin 19, 406–416.

37.

Patenaude

, Smith

, Kennedy

, Jenkinson

(2011) A Bayesian model of shape and appearance for subcortical brain segmentation. Neuroimage 56, 907–922.

38.

Lorensen

, Cline

(1987) Marching cubes: A high resolution 3D surface construction algorithm. ACM SIGGRAPH Comput Graph 21, 163–169.

39.

Stonnington

, Wu

, Zhang

, Shi

, Bauer Iii

, Devadas

, Su

, Locke

DEC

, Reiman

, Caselli

, Chen

, Wang

; Alzheimer’s Disease Neuroimaging Initiative (2021) Improved prediction of imminent progression to clinically significant memory decline using surface multivariate morphometry statistics and sparse coding. J Alzheimers Dis 81, 209–220.

40.

Hoppe

(1996) Progressive meshes. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pp. 99-108.

41.

Loop

(1987) Smooth subdivision surfaces based on triangles. Master’s Thesis, Department of Mathematics, University of Utah, Salt Lake City.

42.

Wang

, Lui

, Gu

, Hayashi

, Chan

, Toga

, Thompson

, Yau

S-T

(2007) Brain surface conformal parameterization using Riemann surface structure. IEEE Trans Med Imaging 26, 853–865.

43.

Wang

, Zhang

, Gutman

, Chan

, Becker

, Aizenstein

, Lopez

, Tamburo

, Toga

, Thompson

(2010) Multivariate tensor-based morphometry on surfaces: Application to mapping ventricular abnormalities in HIV/AIDS. Neuroimage 49, 2141–2157.

44.

Shi

, Thompson

, Gutman

, Wang

; Alzheimer’s Disease Neuroimaging Initiative (2013) Surface fluid registration of conformal representation: Application to detect disease burden and genetic influence on hippocampus. Neuroimage 78, 111–134.

45.

Shi

, Stonnington

, Thompson

, Chen

, Gutman

, Reschke

, Baxter

, Reiman

, Caselli

, Wang

;Alzheimer’s Disease Neuroimaging Initiative (2015) Studying ventricular abnormalities in mild cognitive impairment with hyperbolic Ricci flow and tensor-based morphometry. Neuroimage 104, 1–20.

46.

Leow

, Huang

S-C

, Geng

, Becker

, Davis

, Toga

, Thompson

(2005) Inverse consistent mapping in 3D deformable image registration: Its construction and statistical properties. In Information Processing in Medical Imaging: 19th International Conference, IPMI 2005, Glenwood Springs, CO, USA, July 10–15, 2005. Proceedings 19, pp. 493-503.

47.

Pizer

, Fritsch

, Yushkevich

, Johnson

, Chaney

(1999) Segmentation, registration, and measurement of shape variation via image object shape. IEEE Trans Med Imaging 18, 851–865.

48.

Thompson

, Hayashi

, De Zubicaray

, Janke

, Rose

, Semple

, Hong

, Herman

, Gravano

, Doddrell

, Toga

(2004) Mapping hippocampal and ventricular change in Alzheimer disease. Neuroimage 22, 1754–1766.

49.

Zhang

, Shi

, Stonnington

, Li

, Gutman

, Chen

, Reiman

, Caselli

, Thompson

, Ye

, Wang

(2016) Hyperbolic space sparse coding with its application on prediction of Alzheimer’s disease in mild cognitive impairment. Med Image Comput Comput Assist Interv 9900, 326–334.

50.

Lin

, Li

, Sun

, Lai

M-J

, Davidson

, Fan

, Ye

(2014) Stochastic coordinate coding and its application for drosophila gene expression pattern annotation. arXiv preprint arXiv:14078147.

51.

Boureau

Y-L

, Ponce

, LeCun

(2010) A theoretical analysis of feature pooling in visual recognition. In Proceedings of the 27th international conference on machine learning (ICML-10), pp. 111-118.

52.

Pedregosa

, Varoquaux

, Gramfort

, Michel

, Thirion

, Grisel

, Blondel

, Prettenhofer

, Weiss

, Dubourg

, Vanderplas

, Passos

, Cournapeau

, Brucher

, Perrot

, Duchesnay

(2011) Scikit-learn: Machine learning in Python. J Mach Learn Res 12, 2825–2830.

53.

Chawla

N V

, Bowyer

, Hall

, Kegelmeyer

(2002) SMOTE: Synthetic minority over-sampling technique. J Artif Intell Res 16, 321–357.

54.

Janelidze

, Bali

, Ashton

, Barthélemy

, Vanbrabant

, Stoops

, Vanmechelen

, He

, Dolado

, Triana-Baltzer

, Pontecorvo

, Zetterberg

, Kolb

, Vandijck

, Blennow

, Bateman

, Hansson

(2023) Head-to-head comparison of 10 plasma phospho-tau assays in prodromal Alzheimer’s disease. Brain 146, 1592–1601.

55.

Milà-Alomà

, Ashton

, Shekari

, Salvadó

, Ortiz-Romero

, Montoliu-Gaya

, Benedet

, Karikari

, Lantero-Rodriguez

, Vanmechelen

, Day

, González-Escalante

, Sánchez-Benavides

, Minguillon

, Fauria

, Molinuevo

, Dage

, Zetterberg

, Gispert

, Suárez-Calvet

, Blennow

(2022) Plasma p-tau231 and p-tau217 as state markers of amyloid-β pathology in preclinical Alzheimer’s disease. Nat Med 28, 1797–1801.

56.

Palmqvist

, Tideman

, Cullen

, Zetterberg

, Blennow

; Alzheimer’s Disease Neuroimaging Initiative; Dage JL, Stomrud E, Janelidze S, Mattsson-Carlgren N, Hansson O (2021) Prediction of future Alzheimer’s disease dementia using plasma phospho-tau combined with other accessible measures.Nat Med 27, 1034–1042.

57.

Tosun

, Chen

, Yu

, Sundell

, Suhy

, Siemers

, Schwarz

, Weiner

,; Alzheimer’s Disease Neuroimaging Initiative (2016) Amyloid status imputed from a multimodal classifier including structural MRI distinguishes progressors from nonprogressors in a mild Alzheimer’s disease clinical trial cohort. Alzheimers Dement 12, 977–986.

58.

Trzepacz

, Hochstetler

, Yu

, Castelluccio

, Witte

, Dell’Agnello

, Degenhardt

; Alzheimer’s Disease Neuroimaging Initiative (2016) Relationship of hippocampal volume to amyloid burden across diagnostic stages of Alzheimer’s disease. Dement Geriatr Cogn Disord 41, 68–79.

59.

Petrone

, Casamitjana

, Falcon

, Artigues

, Operto

, Cacciaglia

, Molinuevo

, Vilaplana

, Gispert

; Alzheimer’s Disease Neuroimaging Initiative (2019) Prediction of amyloid pathology in cognitively unimpaired individuals using voxel-wise analysis of longitudinal structural brain MRI. Alzheimers Res Ther 11, 72.

60.

Ansart

, Epelbaum

, Gagliardi

, Colliot

, Dormont

, Dubois

, Hampel

, Durrleman

; Alzheimer’s Disease Neuroimaging Initiative* and the INSIGHT-preAD study (2020) Reduction of recruitment costs in preclinical AD trials: Validation of automatic pre-screening algorithm for brain amyloidosis. Stat Methods Med Res 29, 151–164.

61.

Pekkala

, Hall

, Ngandu

, van Gils

, Helisalmi

, Hänninen

, Kemppainen

, Liu

, Lötjönen

, Paajanen

, Rinne

, Soininen

, Kivipelto

, Solomon

(2020) Detecting amyloid positivity in elderly with increased risk of cognitive decline. Front Aging Neurosci 12, 228.

62.

Ott

, Cohen

, Gongvatana

, Okonkwo

, Johanson

, Stopa

, Donahue

, Silverberg

;Alzheimer’s Disease Neuroimaging Initiative (2010) Brain ventricular volume and cerebrospinal fluid biomarkers of Alzheimer’s disease. J Alzheimers Dis 20, 647–657.

63.

Ten Kate

, Redolfi

, Peira

, Bos

, Vos

, Vandenberghe

, Gabel

, Schaeverbeke

, Scheltens

, Blin

, Richardson

, Bordet

, Wallin

, Eckerstrom

, Molinuevo

, Engelborghs

, Van Broeckhoven

, Martinez-Lage

, Popp

, Tsolaki

, Verhey

FRJ

, Baird

, Legido-Quigley

, Bertram

, Dobricic

, Zetterberg

, Lovestone

, Streffer

, Bianchetti

, Novak

, Revillard

, Gordon

, Xie

, Wottschel

, Frisoni

, Visser

, Barkhof

(2018) MRI predictors of amyloid pathology: Results from the EMIF-AD Multimodal Biomarker Discovery study. Alzheimers Res Ther 10, 100.

64.

, Su

, Zhu

, Jalili Mallak

, Lepore

, Reiman

, Caselli

, Thompson

, Chen

, Wang

; Alzheimer’s Disease Neuroimaging Initiative (2023) Improved prediction of amyloid-β and tau burden using hippocampal surface multivariate morphometry statistics and sparse coding. J Alzheimers Dis 91, 637–651.