Can MRI screen for CSF biomarkers in neurodegenerative disease?

Abstract

Objective:

Alzheimer disease (AD) and frontotemporal lobar degeneration (FTLD) may have overlapping clinical presentations despite distinct underlying neuropathologies, thus making in vivo diagnosis challenging. In this study, we evaluate the utility of MRI as a noninvasive screening procedure for the differential diagnosis of AD and FTLD.

Methods:

We recruited 185 patients with a clinically diagnosed neurodegenerative disease consistent with AD or FTLD who had a lumbar puncture and a volumetric MRI. A subset of 32 patients had genetic or autopsy-confirmed AD or FTLD. We used singular value decomposition to decompose MRI volumes and linear regression and cross-validation to predict CSF total tau (tt) and β-amyloid (Aβ_1-42) ratio (tt/Aβ) in patients with AD and patients with FTLD. We then evaluated accuracy of MRI-based predicted tt/Aβ using 4 converging sources including neuroanatomic visualization and categorization of a subset of patients with genetic or autopsy-confirmed AD or FTLD.

Results:

Regression analyses showed that MRI-predicted tt/Aβ is highly related to actual CSF tt/Aβ. In each group, both predicted and actual CSF tt/Aβ have extensively overlapping neuroanatomic correlates: low tt/Aβ consistent with FTLD is related to ventromedial prefrontal regions while high tt/Aβ consistent with AD is related to posterior cortical regions. MRI-predicted tt/Aβ is 75% accurate at identifying underlying diagnosis in patients with known pathology and in clinically diagnosed patients with known CSF tt/Aβ levels.

Conclusion:

MRI may serve as a noninvasive procedure that can screen for AD and FTLD pathology as a surrogate for CSF biomarkers.

There is urgent need to improve in vivo diagnosis of neurodegenerative conditions like Alzheimer disease (AD) and frontotemporal lobar degeneration (FTLD) because of potential treatments targeting the underlying abnormal proteinopathies. Despite distinct biochemical abnormalities, AD and FTLD often share similar clinical features and thus clinical diagnosis alone is unreliable.¹ CSF biomarkers based on a ratio of total tau (tt) and β-amyloid (Aβ_1-42) (tt/Aβ) provide a highly sensitive and specific in vivo diagnostic tool for discriminating between AD and FTLD.^2–4 In our autopsy series, tt/Aβ yields 90% sensitivity and 97% specificity.⁴ Despite reasonable diagnostic accuracy, a lumbar puncture is often viewed by patients as invasive. In the setting of a clinical trial where repeated monitoring of endpoints such as CSF biomarkers may be important, an alternative biomarker that provides equivalent diagnostic accuracy, but is more appealing to patients, would be ideal. Volumetric MRI may serve this role since patients with autopsy- or CSF-defined AD and FTLD have relatively distinct neuroanatomic profiles of gray matter (GM) neurodegeneration.^5–7

We present a novel screening method for discriminating between AD and FTLD based on MRI-predicted tt/Aβ CSF values in individuals. This approach yields a single, biologically meaningful value that can screen individuals with a noninvasive MRI and identify borderline cases that would benefit from a more invasive CSF follow-up analysis. Additionally, our approach, in contrast to radioligand PET imaging, is widely available and considerably less expensive. Based on 4 converging analyses, our findings suggest that MRI can yield a noninvasive biomarker screening for neurodegenerative disease.

METHODS

Protocol approval, registration, and patient consent

Patients were recruited from University of Pennsylvania Department of Neurology. Written informed consent was obtained using a University of Pennsylvania Institutional Review Board–approved protocol.

Participants

We investigated 185 patients clinically diagnosed with a neurodegenerative disease, who had a CSF profile (see below) consistent with AD (n = 88) or non-AD (n = 97). Based on their phenotypic characteristics, we inferred that non-AD is consistent with FTLD. All patients were clinically diagnosed by a board-certified neurologist (M.G.), an expert in neurodegenerative diseases, using published criteria.^8–13 Demographic information is summarized in table 1. Clinical phenotypes included in the cohort are detailed in table e-1 on the Neurology® Web site at www.neurology.org. All participants had a lumbar puncture and volumetric MRI, acquired on average 4.9 months (SEM = 1.4) apart, and 58.9% were acquired the same day. Patients with AD and patients with FTLD only had mild dementia and were matched for disease severity using the Mini-Mental State Examination [t(177) = 1.81; p = 0.07], age at MRI [t(183) = 0.22; p > 0.1], and education [t(183) = 1.05; p > 0.1]. Patients with AD and patients with FTLD were also matched for disease duration (U = 3756; p > 0.1). Patient groups differed in gender (χ² = 7.98; p = 0.005), but an analysis of variance comparing linear regression models (see below) that did not include gender and that included gender as a nuisance covariate revealed no significant difference (F = 2.63; p > 0.1).

Table 1.

Mean (SEM) demographic characteristics of CSF-defined Alzheimer disease and frontotemporal lobar degeneration cohort

Validation cohort

A subset of patients (n = 32) additionally had either a pathogenic genetic mutation consistent with AD (PSEN1) or FTLD (GRN, C9orf72) or a detailed neuropathologic investigation, as described in appendix e-1. This includes 11 with AD and 21 with FTLD.

CSF analysis

CSF analytes of tt and Aβ_1-42 were obtained using previously reported procedures and evaluated with either a sandwich ELISA (INNOTEST, Innogenetics, Ghent, Belgium)² or a LUMINEX xMAP platform (INNO-BIA AlzBio3, Innogenetics).¹⁴ tt/Aβ values were combined across platforms using an autopsy-validated conversion factor.⁴

Structural MRI methods

A detailed description of MRI methods is provided in appendix e-1, A–G, and figure e-1. Briefly, all participants underwent a high-resolution structural T1-weighted magnetization-prepared rapid gradient echo MRI acquired using a 3.0 T Siemens Trio scanner (appendix e-1A). MRI images were registered into a common stereotactic space and segmented into GM cortical density maps (appendix e-1B). To manage the large number of voxels within each GM volume, we employed a dimensionality reduction technique that identifies spatial patterns that account for the variance within a dataset. Specifically, we used singular value decomposition (SVD) to decompose the images into eigenvectors and eigenvalues that account for the variance in the data (appendix e-1C). SVD is a data-driven region-of-interest (ROI) method that is similar to a traditional principal components statistical approach. This identifies sets of voxels that are correlated with one another, and reduces dimensionality from thousands of voxels to a smaller number of eigenvectors or intercorrelated clusters that capture the variance across a given dataset.

To identify the optimal number of eigenvalues, we first used a k-folds cross-validation procedure that involved randomly dividing the entire cohort of MRI images into 20 folds, each comprised of 9–10 images. Cross-validation was used in an effort to not only identify significant relationships, but to evaluate the predictive role of this algorithm for application in novel datasets. To do this we performed linear regressions using 19 folds for training, tested on the remaining fold, and then repeated this process a total of 20 times until each fold was tested. For each of 20 splits of the data, we used the training set to identify an optimal linear model by iteratively increasing the number of eigenvectors in the model (tt/Aβ ∼ 1 + EIG₁; tt/Aβ ∼ 1 + EIG₁ + EIG₂ + … + EIG_N-1). We selected the model with the lowest mean prediction residual, and found that the first 15 eigenvectors most accurately predicted tt/Aβ (appendix e-1D). We then applied the training model to the test set to identify prediction error on the held out data. We repeated the procedure across each of 20 folds such that tt/Aβ was predicted for each individual patient (appendix e-1E). Finally, since our SVD procedure does not yield neuroanatomic regions, we performed whole-brain voxelwise regressions in SPM8 (http://www.fil.ion.ucl.ac.uk/spm/) to relate actual and predicted tt/Aβ values to the anatomic distribution of MRI cortical density (appendix e-1F).

Classification evaluation

To evaluate the classification accuracy of our MRI-based prediction algorithm, we first generated a receiver operating characteristic curve using actual CSF tt/Aβ values to determine sensitivity and specificity in our validation cohort with known AD and FTLD. We selected the cutoff that mutually maximized sensitivity and specificity (appendix e-1G).

RESULTS

tt/Aβ prediction results

We computed a linear regression to examine the relationship between MRI-predicted tt/Aβ values on the basis of GM atrophy and actual tt/Aβ values measured in CSF. This was accomplished by entering the 15 optimal eigenvectors identified during cross-validation and feature selection as independent factors and the actual CSF tt/Aβ values as a dependent variable. This revealed a highly significant fit between MRI-predicted tt/Aβ values and actual CSF tt/Aβ values (F_15,169 = 9.61; p < 0.0001). The mean (SEM) prediction error for this regression was 0.61 (0.04). This suggests that MRI can be used to predict tt/Aβ obtained from CSF using a noninvasive procedure.

Neuroanatomic visualization

To illustrate the neuroanatomic regions related to CSF- and MRI-predicted tt/Aβ values, we performed 2 voxelwise whole-brain regressions. First, the regression relating CSF tt/Aβ to GM revealed that lower actual tt/Aβ values, which are associated with FTLD (see below), were related to reduced GM density in ventromedial prefrontal cortex, orbital frontal cortex, insula, thalamus, and anterior temporal cortex (figure 1A, yellow and green). Higher tt/Aβ values associated with AD were related to reduced GM density in posterior regions including superior parietal cortex, precuneus, and occipital association cortex (figure 1B, yellow and green).

Figure 1. Univariate regression results relating actual tt/Aβ and predicted tt/Aβ to GM density in MRI.

(A) Decreasing tt/Aβ is related to decreased density in anterior GM regions consistent with frontotemporal lobar degeneration. (B) Increasing tt/Aβ is related to decreased density in posterior GM regions consistent with Alzheimer disease. Green reflects overlap of both actual and predicted tt/Aβ, yellow reflects only actual tt/Aβ, blue reflects only predicted tt/Aβ. Markers are y-axis of coronal slices in Montreal Neurological Institute coordinate space. Aβ = β-amyloid; GM = gray matter; tt = total tau.

A regression revealed a very similar distribution of reduced GM density associated with tt/Aβ levels predicted by MRI. Lower predicted tt/Aβ values, consistent with FTLD, were related to reduced GM density in frontal regions including ventromedial prefrontal cortex, orbital frontal cortex, insula, caudate, and anterior temporal cortex (figure 1A, blue and green). By comparison, higher predicted tt/Aβ values, consistent with AD, were related to reduced density in posterior GM regions, including superior parietal cortex, precuneus, angular gyrus, and occipital cortex (figure 1B, blue and green).

In figure 1, overlapping regions identified in the regression analyses relating both predicted and actual tt/Aβ to GM density are illustrated in green. We found considerable overlap between neuroanatomic regions associated with tt/Aβ. The anatomic peaks of all clusters observed in the neuroanatomic validation analyses are summarized in table 2.

Table 2.

Neuroanatomic regions with reduced cortical density related to actual tt/Aβ and predicted tt/Aβ

Measured and MRI-predicted CSF tt/Aβ classification in patients with known pathology

To determine how well actual CSF tt/Aβ classified individual patients as having AD or FTLD in our genetic and autopsy validation cohort, we computed a receiver operating characteristic curve. Actual CSF tt/Aβ values provided a strong classifier of neuropathologic diagnosis (area under the curve = 0.92; p < 0.001). Using a cutoff of −1.38, we obtained 91% sensitivity and 81% specificity for correctly classifying AD, with an overall classification accuracy of 84%. Seventeen of 21 patients were classified as having FTLD, and 10 out of 11 patients were correctly classified as having AD.

Using CSF tt/Aβ predicted from MRI and the −1.38 cutoff identified in the actual CSF analysis, we found 75.0% overall classification accuracy. In patients with genetic or autopsy-confirmed FTLD, 17 out of 21 (81%) were accurately predicted as having FTLD using CSF tt/Aβ predicted from MRI. The 4 misclassified cases were identified using genetic mutations (2 C9orf72; 2 GRN) as a proxy for underlying neuropathology and it is possible that these patients have comorbid AD in addition to FTLD.¹ Alternatively, recent neuroimaging evidence suggests that patients with C9orf72 expansion and GRN mutations may have a more posterior distribution of disease,¹⁵ which can partially overlap with the distribution of disease observed in AD and therefore lead to misclassification. For patients with genetic or autopsy-confirmed AD, 7 out of 11 (64%) were accurately classified as having AD using MRI-predicted CSF tt/Aβ. A χ² analysis confirmed significantly accurate classification of AD and FTLD in this validation cohort on the basis of predicted tt/Aβ (χ² = 6.36; p < 0.05). Overall classification accuracy based on MRI-predicted tt/Aβ and actual CSF values did not statistically differ (χ² = 0.36; p > 0.1).

To evaluate whether misclassification was consistent across CSF and MRI modalities, we evaluated the underlying pathology for each misclassified case, and these are summarized in figure e-2. One case of autopsy-confirmed FTLD had CSF and MRI consistent with AD. Three cases of autopsy-confirmed FTLD had CSF consistent with FTLD but were predicted on the basis of MRI as AD. Four additional cases had autopsy-confirmed AD with CSF consistent with AD, but were classified on the basis of MRI as FTLD.

Measured and MRI-predicted CSF tt/Aβ classification in clinical patients with CSF

To evaluate the overall accuracy of MRI for screening clinical patients for underlying pathology, we compared the accuracy of predicted tt/Aβ relative to actual tt/Aβ for the entire cohort of patients included in our study. A χ² analysis confirmed significant screening accuracy (χ² = 46.65; p < 0.001) and a comparison of the distribution of patients classified as having AD or FTLD using actual CSF tt/Aβ does not differ statistically from classification using MRI-predicted tt/Aβ (χ² = 0.02; p > 0.1). Of 97 patients with a CSF tt/Aβ profile consistent with FTLD, 73 also demonstrated a predicted tt/Aβ consistent with FTLD (75.3% accuracy). Of 88 patients with a CSF tt/Aβ profile consistent with AD, 66 also achieved a predicted tt/Aβ consistent with FTLD. Together, we achieved 75.1% overall classification accuracy based on MRI-predicted CSF tt/Aβ relative to actual tt/Aβ levels (see figure 2).

Figure 2. Scatterplot illustrating actual CSF tt/Aβ relative to predicted tt/Aβ derived from MRI screening procedure.

Dotted lines represent CSF-defined total tau (tt)/β-amyloid (Aβ) cutoff values identified in receiver operating characteristic curve analysis of autopsy and genetic validation cohort. Blue indicates patients with actual CSF tt/Aβ consistent with frontotemporal lobar degeneration (FTLD) and red indicates patients with actual CSF tt/Aβ consistent with Alzheimer disease (AD).

Since our cohort was clinically heterogeneous, we evaluated whether MRI classification accuracy differed across clinical phenotypes. A χ² analysis confirmed that classification was equally accurate across clinical phenotypes (χ² = 3.36; p > 0.1) (see table e-1).

DISCUSSION

GM density measured noninvasively with MRI can be used to predict tt/Aβ values that distinguish between AD and FTLD with reasonable accuracy. Specifically, MRI-predicted and actual CSF tt/Aβ values are highly correlated, predicted tt/Aβ accurately defines the anatomic distribution of atrophy in AD and FTLD, and predicted tt/Aβ values are reasonably accurate at classifying individual patients as having AD or FTLD pathology. This study establishes empirical evidence that an MRI-based technique can predict a single, biologically valid level of CSF tt/Aβ. This may contribute to diagnosis and treatment trials of neurodegenerative conditions by screening for individuals requiring a more invasive diagnostic lumbar puncture. In comparison to nonspecific MRI approaches, this MRI-based technique also potentially minimizes power demands of studies by reducing multifocal regions associated with neurodegenerative disease to a single value and thus provides a candidate marker to monitor in clinical treatment trials.

Structural MRI has been proposed to contribute to clinically relevant studies such as meeting inclusion criteria and ascertaining response during treatment trials.¹⁶ This is valuable in neurodegenerative conditions like AD and FTLD, where clinical criteria are less than optimal markers of disease. In some work, the anatomic distribution of GM atrophy using MRI has been suggested to help identify patients meeting diagnostic criteria for entry into a trial. This is more straightforward in AD, where a single region—the hippocampus—is often the focus of anatomic diagnosis.^17–19 However, potential difficulties include hippocampal sparing in some clinical variants of AD.²⁰ In FTLD, there may be several different anatomic foci of atrophy,⁵ or lobar atrophy may not be observable.^21,22 With 75% overall diagnostic accuracy, MRI-based classification using predicted CSF is intended to serve as a screen to identify individuals requiring additional diagnostic studies. This algorithm may reduce the number of more invasive diagnostic procedures such as lumbar punctures.

Likewise, the advantages of using predicted CSF tt/Aβ values derived from individual MRI are immense when considering MRI as an outcome measure in a treatment trial.¹⁶ Nonspecific MRI markers such as analyses of whole brain and ventricular volumes have been proposed to minimize missing longitudinal changes in a heterogeneous group of patients with different anatomic distributions of focal atrophy. This approach is associated with a large cost because of reduced sensitivity and limited power—anatomic ascertainment is nonspecific in order to minimize the error to detect change that can occur in several different brain regions. The SVD-mediated analysis of MRI GM atrophy used in the present study does not depend on a multiplicity of anatomic locations and instead integrates brain-wide atrophic voxels into a single biologically meaningful cluster. Since the outcome measure reflects the actual CSF tt/Aβ value in individual subjects with reasonable accuracy, this approach additionally provides an opportunity to measure direct molecular effects of intervention trials designed to target tau or Aβ_1-42.^23–25

Comparative studies are very important because of the frequent difficulty distinguishing clinically similar phenotypes with distinct underlying histopathologic features. Inappropriate patients may be entered into etiologically specific trials based on sensitive data that identify members of a group relative to healthy controls, but these data are not specific. Beyond the obvious biological problems, there are increased power demands associated with diagnostic inaccuracy. Previous automated MRI-based classification studies for discriminating between AD and FTLD have yielded relatively high accuracy, but these studies were restricted to clinically diagnosed patients, and patient groups were not defined on the basis of CSF or autopsy studies,^26,27 and thus it is difficult to infer true accuracy. One additional comparative MRI classification study has been conducted to discriminate between neuropathology-confirmed AD and FTLD, and this yielded high sensitivity (94.7%) and specificity (83.3%) for diagnosis.²⁸ However, the report was restricted to smaller cohorts with many different scanners and relied on a highly complex support vector machine approach. A major advantage of the SVD approach is that the only parameter is the amount of variance that should be retained for the model (the number of eigenvectors), and this single parameter may easily be set by cross-validation. Our approach provides information on individual patients as well.

To implement this screening procedure for individual patients, the current eigenvalues and corresponding regression weights would require application to each individual's MRI volume. Since optimal eigenvalues were defined using 20-fold cross-validation, they should be reproducible in an independent cohort. More recent SVD methods may facilitate extending this procedure to individuals by implementing a “sparseness” parameter that constrains eigenvalues to clusters of voxels that exceed an a priori extent threshold and thus can support a more traditional ROI-based approach.²⁹ This is consistent with research suggesting that “signature regions” are useful for evaluating neurodegenerative disease.³⁰

In the current study, we used whole-brain voxelwise regression analyses to visualize GM regions associated with increased or decreased tt/Aβ ratios. Few previous studies have evaluated the direct relationship between CSF biomarkers and MRI.^3,31–33 In one study, a relationship was observed between Aβ_1-42 and whole brain volume in healthy adults, and between tt and whole brain volume in patients with AD.³¹ However, these relationships involved only nonspecific global brain volume. In a study of a clinically defined cohort of patients with frontotemporal dementia, reduced tt levels correlated with insula, anterior cingulate, ventromedial and orbital frontal, and superior temporal cortical density.³ In the present study, we observed that higher predicted tt/Aβ values are related to reduced GM density in posterior cortical regions, consistent with imaging studies of AD; lower predicted tt/Aβ values were related to reduced GM density in anterior cortical regions, consistent with imaging studies of FTLD.^5–7 We restricted our analysis to the ratio of tt/Aβ since this combination of analytes is most sensitive and specific for the diagnosis of FTLD and AD.^2,3,34 Future studies independently evaluating the relationship between other CSF biomarkers and MRI may help elucidate the complex relationship between CSF analytes and cortical atrophy.

One caveat to consider in using MRI as a CSF screening procedure is related to the trajectory of disease. While we demonstrate that the distribution of GM atrophy in AD and FTLD is highly related to a distinct range of CSF tt/Aβ values, these biological changes may occur at different stages in the disease course.³⁵ Empirical data regarding longitudinal changes in CSF analyte values is scant, and results have been controversial.^36,37 In this context, one reason for the paucity of longitudinal CSF studies is patients' perception of invasiveness and their unwillingness to undergo multiple lumbar punctures. The lack of longitudinal CSF studies highlights the importance of establishing an alternative, less invasive screening procedure such as MRI.

Our results suggest that MRI provides a reasonably accurate screening tool discriminating between AD and FTLD. MRI-based screening yields an acceptable anatomic distribution of disease and predicted CSF values that are similar to actual CSF values from a procedure that is perceived as less invasive and more easily repeated. Since this method yields a single biological value, it is possible to use MRI to screen patients for inclusion in clinical trials in a cost-effective manner and to provide a biologically plausible outcome measure that optimizes power in drug treatment trials.

Glossary

Aβ

β-amyloid

AD

Alzheimer disease

FTLD

frontotemporal lobar degeneration

GM

gray matter

ROI

region of interest

SVD

singular value decomposition

tt

total tau

AUTHOR CONTRIBUTIONS

Corey T. McMillan and Brian Avants drafted/revised manuscript for content, contributed to study concept/design, performed analysis/interpretation of the data, and performed statistical analysis. David Irwin, Jon Toledo, David Wolk, and Vivianna Van Deerlin contributed to analysis/interpretation of the data and acquisition of the data. Leslie Shaw, John Trojanowski, and Murray Grossman drafted/revised manuscript for content, contributed to study concept/design, performed analysis/interpretation of the data, obtained funding, and provided supervision.

STUDY FUNDING

Supported by the Wyncote Foundation and the NIH (HD060406 to C.T.M., NS44266 to M.G., AG17586 to M.G., J.Q.T., and V.V., AG15116 to M.G., AG32953 to M.G. and J.Q.T., NS53488 to M.G., J.Q.T., and V.V., and AG10124 to J.Q.T and V.V.). D.J.I. is supported by the NIH T32-AG000255. J.B.T. was supported by a grant of the Alfonso Martín Escudero Foundation.

DISCLOSURE

The authors report no disclosures relevant to the manuscript. Go to Neurology.org for full disclosures.