Baseline Multimodal Imaging to Predict Longitudinal Clinical Decline in Atypical Alzheimer’s Disease

Abstract

There are recognized neuroimaging regions of interest in typical Alzheimer’s disease which have been used to track disease progression and aid prognostication. However, there is a need for validated baseline imaging markers to predict clinical decline in atypical Alzheimer’s Disease. We aimed to address this need by producing models from baseline imaging features using penalized regression and evaluating their predictive performance on various clinical measures.

Baseline multimodal imaging data, in combination with clinical testing data at two time points from 46 atypical Alzheimer’s Disease patients with a diagnosis of logopenic progressive aphasia (N = 24) or posterior cortical atrophy (N = 22), were used to generate our models. An additional 15 patients (logopenic progressive aphasia = 7, posterior cortical atrophy = 8), whose data were not used in our original analysis, were used to test our models. Patients underwent MRI, FDG-PET and Tau-PET imaging and a full neurologic battery at two time points. The Schaefer functional atlas was used to extract network-based and regional gray matter volume or PET SUVR values from baseline imaging. Penalized regression (Elastic Net) was used to create models to predict scores on testing at Time 2 while controlling for baseline performance, education, age, and sex. In addition, we created models using clinical or Meta Region of Interested (ROI) data to serve as comparisons.

We found the degree of baseline involvement on neuroimaging was predictive of future performance on cognitive testing while controlling for the above measures on all three imaging modalities. In many cases, model predictability improved with the addition of network-based neuroimaging data to clinical data. We also found our network-based models performed superiorly to the comparison models comprised of only clinical or a Meta ROI score.

Creating predictive models from imaging studies at a baseline time point that are agnostic to clinical diagnosis as we have described could prove invaluable in both the clinical and research setting, particularly in the development and implementation of future disease modifying therapies.

Introduction

Described first by Alois Alzheimer, Alzheimer’s Disease (AD) typically manifests clinically as an insidious, progressive amnestic syndrome with prominent deficits in episodic memory.^1,2 The neuropathologic diagnosis of AD necessitates the presence of neuritic plaques and neurofibrillary tangles, composed of beta amyloid and hyperphosphorylated paired helical filament tau, respectively.³ However with the advent of in vivo biomarkers to aid in diagnosis, AD neuropathology has also been seen in a number of non-amnestic phenotypes, or atypical AD variants, which affect different cognitive domains than typical amnestic AD.⁴ In contrast to typical AD, atypical AD variants can present with dysexecutive, visual, language, and behavioral symptomatology as the predominant features.⁵ In addition, these patients classically present at a younger age than typical AD, with symptoms beginning prior to the age of sixty-five.⁶ There is also significant clinical heterogeneity in clinical course, including among patients with the same clinical syndrome.⁷ Because of this, misdiagnosis is common and time to diagnosis is often prolonged. It is also postulated that the prevalence of atypical AD is lower than typical AD.⁸ Thus, there are relatively few large cohorts of patients with longitudinal data for atypical AD patients.

Two well-studied variants of atypical AD are the visual variant (posterior cortical atrophy, or PCA) and the language variant (logopenic progressive aphasia or, LPA). PCA is a progressive syndrome with patient presentations most notable for higher-order visual deficits in spatial and object perception, aspects of Gerstmann and Balint syndromes, and relative sparing of episodic memory.^9–11 Likewise, LPA is a progressive syndrome defined by prominent difficulty with language, specifically with impaired single-word retrieval and repetition of phrases, phonologic errors, and impaired sentence comprehension with spared object knowledge and single-word comprehension.^12,13

Both variants have available literature that characterize neuroimaging features, however long-term evaluation from a clinicoradiologic perspective is an area of active interest given the relative sparsity of longitudinal studies.¹⁴ On MRI, the most significant reductions in grey matter are in the occipital and parietal lobes in PCA patients, with concomitant hypometabolism on FDG-PET.^15,16 In contrast, LPA patients display atrophy on MRI in the left temporoparietal region, with hypometabolism on FDG-PET in these same areas.^17,18,19 We hypothesized that we could utilize baseline multi-modal imaging data collected from a large cohort of patients to formulate models that would accurately predict longitudinal decline.

Other studies have also explored using baseline neuroimaging with modalities such as MRI, FDG-PET, and Tau-PET to predict clinical changes in typical and atypical AD.^20–30 In comparison, we created models using sparse regression from baseline neuroimaging data of LPA and PCA patients to predict scoring on future neurologic testing. Models were created using one of three imaging modalities (MRI, FDG-PET, and Tau-PET), and compared across modalities. We also compared our imaging-based models against models that utilized only baseline clinical data, as a fair test of whether imaging data improves predictions. Finally, we also compared our more complex imaging-based models, where different regions can be selected for different clinical predictions, to ones that incorporate simple, widely used imaging data from meta-ROIs that are static across models. A predictive model that is agnostic to underlying diagnosis such as we have described could prove valuable in both the clinical and research settings.

Methods

We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study. No part of the study procedures or analysis plans was preregistered prior to the research being conducted.

Participants:

The primary cohort consisted of forty-six atypical AD patients with a diagnosis of LPA (N = 24) or PCA (N = 22), who were recruited by the Neurodegenerative Research Group (NRG) at Mayo Clinic into an ongoing NIH-funded study focused on atypical AD. All patients had two study visits between 2011 and 2020. The median time between visits was 365 days. The PCA patients met published criteria for PCA and the LPA patients met criteria for LPA.^11,12 In addition, data from fifteen additional patients (LPA = 7, PCA = 8) who were not part of the original analyses were used as a test set to evaluate model generalizability. For study inclusion, we required that patients had evidence of beta-amyloid deposition on (11)C-labeled Pittsburgh Compound-B (PiB) PET.^2,11–13,31 Patients were excluded from the study if there were other clinical conditions that better explained their symptoms.

Clinical Assessment:

Patients underwent a comprehensive neurological and neuropsychological evaluation at both timepoints. The details of the neurological and neuropsychological evaluation were reported in a previous study.³² The clinical evaluation included: cognitive assessment with the Montreal Cognitive Assessment (MoCA), the modified Clinical Dementia Rating (CDR) Scale to assess global function, and the Neuropsychiatric Inventory questionnaire (NPI-Q) and Cambridge Behavioral Inventory (CBI) to assess neuropsychiatric and behavioral changes.^33–37 The Movement Disorder Society-sponsored revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) Part III was used to evaluate for Parkinsonism.³⁸ The language evaluation used fifteen item Boston Naming Test (BNT) and the Repetition of Sentences subtest of the Boston Diagnostic Aphasia Examination (BDAE) to evaluate naming and repetition, respectively.³⁹ Surface dyslexia was tested by reading of a list of phonetically challenging words, such as “colonel” or “yacht”, which require knowledge of the word for correct pronunciation.⁴⁰ Additional language testing included rapid single word retrieval on the Letter Fluency task.⁴¹ Apraxia was assessed by the praxis subtest on the Western Aphasia Battery (WAB).^42,43 Oculomotor apraxia and optic ataxia, were assessed on neurologic examination and defined as the inability to voluntarily direct one’s gaze to a particular point, and the impairment of goal-directed hand movements towards visually presented targets, respectively. Gerstmann syndrome was defined as the following: acalculia (three or less on the MoCA calculation), left-right confusion, agraphia, and finger agnosia. Simultanagnosia severity was scored on a 20-point scale.³² Simultanagnosia testing was designed to assess the patient’s ability to perceive the overall context of the presented picture or figure, with examples such as overlapping line drawings, Ishihara plates, and Navon figures.^44–46 Ventral stream tasks, color matching and assessment of apperceptive prosopagnosia by facial recognition, were also tested.^47–49 The Visual Object and Space Perception Battery (VOSP) incomplete letters and cube tests were used to assess visual perception and spatial skills, the Rey-Osterrieth (Rey-O Copy) Complex Figure copy trial was used to assess visual spatial abilities, auditory attention was assessed with the Digit Span subtest from the Wechsler Memory Scale-III (WMS-III-III).^50,51 Verbal and visual memory were assessed with the AVLT and WMS-III-III Visual Reproduction subtest.^52–55 Legal copyright restrictions prevent public archiving of the above tests, which can be obtained from the copyright holders in the cited references. The study was approved by the Mayo Clinic IRB, with patient consent obtained prior to inclusion.

Image acquisition:

Participants completed brain MRI and flortaucipir tau PET at the enrollment visit. All participants had 3T MRI performed from a single vendor (GE Healthcare) using a standardized protocol that included a 3D magnetization prepared rapid acquisition gradient echo (MPRAGE) sequence.⁵⁶ All MPRAGE images were corrected for gradient non-linearity and intensity non-uniformity. A subset of participants (44/46 patients in the original cohort, 8/15 in the test patient group) also had fluorine 18 (18F) fluorodeoxyglucose (FDG) PET imaging performed using a standardized protocol on a PET/CT scanner (GE Healthcare) operating in 3D mode. Participants were injected with 18F-FDG in a dimly lit room and an 8-min FDG scan was performed after a 30-min uptake period, consisting of four 2-min dynamic frames following a low dose CT transmission scan. Standard corrections were applied to address attenuation, scatter, random coincidences, and decay. The four-frame sequences were averaged to create a single static image.

For Tau-PET, an intravenous bolus injection of approximately 370 MBq of flortaucipir was administered, followed by a 20-minute PET acquisition performed 80 min after injection. All patients underwent Tau-PET imaging.

Image Processing:

Details of our fully automated pipeline using Advanced Normalization Tools and SPM12 including in-house modifications and tools, have been described previously.^31,57,58 MPRAGE scans were normalized to the Mayo Clinic Adult Lifespan Template (MCALT; https://www.nitrc.org/projects/mcalt/) and segmented with MCALT priors/settings.^57–61 The Schaefer Functional Atlas and MCALT ADIR122 (based on AAL 120) atlas were then propagated to patient space.^62–64 The Schaefer atlas includes 100 regions of interest (ROIs) divided into eight networks (Visual, Somatomotor, Dorsal Attention, Salience/Ventral Attention, Limbic, Control, Default, Temporal Parietal) each of which is further divided into subnetworks (e.g. Visual Central, Visual Peripheral, Somatomotor, etc) and hemisphere, yielding a total of seventeen subnetworks per hemisphere (Supplemental Figure 1). As the Schaefer atlas does not contain the amygdala or hippocampus, we added those from the ADIR122 atlas and combined them with the two limbic subnetworks from the Schaefer atlas. We also added the left and right deep gray nuclei, the left and right cerebellar hemisphere, and the cerebellar midline structures. In total this resulted in eighteen ‘network-ROIs’ (sixteen Schaefer, one deep gray, one cerebellar hemisphere) for the left and the right hemibrains and one midline cerebellar one. To contrast with the ‘network based’ ROI approach, which involves an ROI selection phase outlined in the statistical analyses section below, we also used a standard meta-ROI validated in prior typical AD research and introduced by Landau and Jagust (Supplemental Figure 2).^31,65

For each of the network-ROIs and the meta-ROI, we calculated the gray matter volume by taking the sum over the voxels in the ROI from the gray matter probabilities produced during segmentation, and then scaling this sum by the participants total intracranial volume.

For participants with PET available the images were co-registered to the associated MPRAGE. The template-to-participant mapping from the MRI segmentation was then used to propagate the atlases to participant space. The median uptake value in the pons (for FDG) and cerebellar crus (for flortaucipir) was calculated and every voxel in the PET was divided by this value to obtain standardized uptake value ratio (SUVR) images. The SUVR within a network-ROI or meta-ROI was then calculated by taking the median within each constituent ROI and then calculating the weighted sum, with the weight determined by the size of the constituent ROI. This is standard practice, and accounts for the fact that differently sized ROIs should contribute different amounts to an overall meta or composite ROI³¹.

Statistical Analyses:

We used elastic net regression to generate linear models that could be used to predict clinical testing scores at Time 2 using baseline clinical and neuroimaging. The most common penalized regression methods, lasso and ridge regression, introduce a penalty to ordinary least squares (OLS) estimates to reduce variance in generated models.⁶⁶ Elastic net models combine these two and are controlled by two parameters, alpha and lambda. The alpha parameter sets the degree of mixing between ridge and lasso regressions, with values approaching 1 favoring an increasing lasso ratio. We set alpha at 0.99, strongly favoring the lasso penalty and variable selection. However, the lasso penalty, when “deciding” between correlated variables, will arbitrarily choose one variable to keep. Ridge regression will reduce correlated variables, but not remove them from the model. To avoid arbitrary removal of correlated variables but allow for model complexity to be sufficiently reduced given the high number of predictors, the alpha value above was chosen, as recommended by the authors of the “glmnet” package (https://glmnet.stanford.edu/articles/glmnet.html). Lambda defines the regularization strength and was obtained through cross validation.^67,68 Elastic net regression maintains several advantages, such as sparing of highly correlated variables and prevention of saturation when the number of predictors exceeds the number of subjects, while allowing for sparse selection by reducing coefficients of poor predictors to 0, effectively reducing model complexity.⁶⁹

We used the same modeling strategy for each clinical variable of interest. Only participants with baseline and follow up data available for that variable were included. First, we created models using only baseline clinical data, including the variable of interest, as predictors and set the clinical variable at follow up as the outcome (e.g. MoCA_Time2 ~ Intercept + Age _Time1 + MoCA _Time1 + Sex _Time1 + Education_Time1 +…). These models serve as the minimal baseline to which models with imaging can be compared. It is worth noting that this practice is not common – investigators rarely use a ‘clinical only’ or ‘baseline’ model such as this, which could lead to overestimation of the predictive value of imaging data or the need for complex models. Second, we augmented the clinical-only models by adding the baseline imaging value on the respective meta-ROI. Separate models were fit for each imaging modality. Again, this serves as an important imaging-baseline set of models to which we can compare our more complex, network-ROI based ones. Third, we included the volume or uptake for each network-ROI instead of the meta-ROI. These models allow for specific network-ROIs to be ‘selected’ depending on the clinical variable that is being predicted, in contrast to the static meta-ROI models. For all models, we included sex, age, and education at baseline. We also ran models including baseline phenotype designation (PCA vs LPA) but present these in Supplementary Table 1 as they were nearly identical in performance to those without baseline phenotype (this is discussed in detail later). After establishing the models above, we also used the baseline clinical and imaging data from the fifteen hold-out participants to predict the clinical variables at Time 2.

In order to compare models with the baseline (clinical only) and meta-ROI models, we computed the error for each observation (predicted/fitted value vs true value) for each model. We then used the Wilcoxon paired rank sum test, as was done in prior work to determine if a model was statistically significantly better.⁷⁰ A similar strategy was used to compare across modalities, with the only difference being that we only compared models across modalities if the models were significantly better than baseline. For example, if MRI and FDG based models predicted MoCA better than baseline, but tau PET did not, we would compare MRI vs FDG only.

False discovery rate correction for multiple comparisons was used with the overall alpha set at 0.05. Analyses were performed with RStudio (Version 1.4.1103. RStudio: Integrated development environment for R. Boston, MA. http://www.rstudio.org/) using the “glmnet” package (version 4.1–1) for elastic net regression.^68,71,72

Analysis code using test patient data is publicly available on the platform Kaggle (Coburn et al-Cortex 2024 (kaggle.com). Raw patient data are not publicly available due to ethical barriers (no consent from participants).

Results

Patient baseline demographics and clinical characteristics are found in Table 1. There were no statistically significant differences between the LPA and PCA groups utilizing a two-sample t-test. Additional patient characteristics are found in Supplementary Table 3.

Table 1 :

Baseline Patient Characteristics

	All (46)	LPA (24)	PCA (22)	p-value
Female (%)	29 (63%)	15 (62.5%)	14 (63.6%)	0.938
Age, Mean (SD)	66.2 (7.43)	67.6 (7.41)	64.7 (7.31)	0.187
Education, Mean (SD)	15.9 (2.45)	16 (2.34)	15.8 (2.62)	0.757
CDR SB	4.1 (3.24)	3.5 (2.11)	4.73 (4.06)	0.216
MoCA	18.5	19.2 (5.04)	17.9 (5.91)	0.428

Clinical Only Models:

Results for the models using only clinical data from the baseline visit to predict test performance at follow up are shown in Table 2. We also included the coefficients for the clinical instrument at baseline and demographic covariates. Within-sample R-squared and mean absolute error (MAE), as well as out of sample R-squared and mean absolute error (MAE), are shown as predictive performance indicators. The R-squared was good (>0.5) in the training set for most measures, emphasizing the importance of considering baseline clinical data in predictive models. Performance in the test cases dropped in nearly all cases but remained good in most. A comparison between these models and ones that include baseline diagnosis is shown in Supplemental Table 1. Only models predicting NPI were improved by adding baseline diagnosis.

Table 2:

Clinical Only Model Results

	Train								Test
Instrument	N	Δ	Baseline	Male	Age	Education	R-sq	MAE	N	R-sq	MAE
MoCA (/30)	46	−4.91	0.853	0	0	0	0.719	1.755	13	0.787	2.185
CBI (/180)	42	11.50	0.914	0	2.896	0	0.717	11.168	12	0.357	9.714
AVLT Delayed Recall (/15)	46	−0.91	0.679	0	0	0	0.583	1.255	10	0.575	1.482
WMS-III-III Visual Reproduction II (/104)	41	−2.07	0.197	0	0	0	0.065	25.991	8	0.242	20.729
Rey-O Copy (/36)	44	−3.26	0.692	0.500	1.096	0	0.805	2.798	10	0.869	1.794
VOSP Letters (/20)	42	−1.79	0.958	0	0.333	−0.203	0.726	1.951	11	0.690	2.788
VOSP Cubes (/10)	41	−0.81	0.840	0.774	−0.152	−0.162	0.747	1.173	10	0.581	1.258
BNT (/15)	44	−1.82	0.928	1.324	0.030	−0.102	0.660	1.207	8	0.858	0.835
BDAE Repetition (/10)	45	−0.79	1.170	0.665	0	0	0.775	0.594	9	0.420	1.086
WMS-III-III Digit Span	46	−1.54	0.878	0	0.333	0	0.601	1.557	10	0.670	2.998
MoCA Calculation (/3)	45	−1.27	0.433	0.482	0.480	0	0.585	0.483	7	−0.165	0.443
Modified CDR	42	1.83	0.751	0.506	−0.210	0	0.606	1.414	12	F	2.574
Letter Fluency	44	−6.27	1.091	0.405	2.046	0.593	0.850	3.334	12	0.699	6.889
NPI	44	1.11	0.812	0	0	−0.154	0.549	1.631	13	0.102	1.760
Gerstmann Score (/7)	29	−1.66	0	0	0	0	0	1.345	1	-	-
Color Matching (/5)	41	−0.20	0.932	0.490	0.140	0.091	0.583	0.280	7	0.516	0.457
Ishihara (/6)	41	−0.39	0.823	0	0	0	0.805	0.687	11	0.952	0.218
Surface Dyslexia (/6)	40	−0.60	0.931	0	0	−0.037	0.406	0.677	11	0.006	0.728
Overlapping Drawings (/5)	41	−0.71	0.863	0	0	0	0.710	0.858	11	0.814	0.446
Navon (/6)	41	−0.59	0.812	0	0	0	0.684	0.865	11	0.430	0.865
Simultanagnosia (/20)	41	−2.24	0.964	0	0.339	0	0.830	1.792	11	0.902	1.613
Facial Recognition (/10)	42	0.17	0.384	0	0	0	0.326	0.406	12	0.263	0.500
UPDRS III	46	3.87	1.630	0	−1.920	0.506	0.736	2.877	13	−0.012	2.873
WAB Praxis (/60)	46	−1.59	1.013	0	0	0	0.749	1.547	12	0.563	1.547

Results for the models using only clinical data from the baseline visit to predict test performance at follow up are shown above in Table 2. Columns indicate the number of cases included in each model (N), the mean difference from baseline to follow up in the training set (Δ), the model coefficients (Baseline - Education), and finally the R-squared (R-sq) and mean absolute error (MAE). The R-squared was (>0.5) for 20/24 tasks in the training set. Performance in the test cases dropped in many cases, but R-squared values remained (>.5) in 13/23 tasks.

MRI Models

There were no measures where models with clinical and MRI meta-ROI values outperformed models with only clinical measures at baseline. Models incorporating network-based GM volume were statistically better than baseline or MRI meta-ROI models for predicting Time 2 performance on MoCA, Rey-Osterreith Figure Copy, BDAE Repetition, BNT, Digit Span, MoCA Calculation, and Overlapping Drawings. Images of the regions contributing to the network-based models that were significantly better than baseline clinical models while controlling for multiple comparisons are shown in Figure 1. However, predictive performance in the test set dropped considerably in the case of BDAE Repetition and MoCA Calculation.

Figure 1: Results for tests where network-based MRI prediction was superior after correction for multiple comparisons.

Shown above, regions in red are indicative of a positive association, such that atrophy in these regions resulted in predictions of more impaired clinical scores, whereas blue areas indicated a negative association where atrophy predicted less impaired scores. Language predominant tasks such as BDAE Repetition and BNT highlighted a predominantly positive correlation with Time 2 performance in relation to gray matter volume in diffuse left greater than right temporo-parietal and attention network regions. Interestingly, on these language-based tasks, atrophy in visual network regions was associated with better language scores, which may suggest that our PCA cohort scored comparatively better. The visual task Overlapping Drawings highlighted a positive correlation with left and right occipital regions. Working memory tasks, such as MoCA calculation and global tasks such as the MoCA, revealed associations with diffuse brain regions, with the MoCA calculation highlighting a positive association with temporo-parietal and limbic network areas. On the MoCA, the strongly positive association between peri-Sylvian structures on the right was unexpected but may reflect asymmetry and severity. Please see Supplemental Figure 4 for a complete list of network regions that were associated with a positive or negative association.

FDG PET Models

For the CBI and WMS-III Visual Reproduction II, models with the FDG-meta-ROI and clinical measures were statistically significantly better than clinical models without imaging. Network-based FDG models were not better than the meta-ROI models for these two measures, and in the case of the CBI, the meta-ROI model was statistically significantly better than the network-based FDG model. Models incorporating network-based FDG uptake were statistically better than baseline or FDG meta-ROI models for predicting Time 2 performance on MoCA Calculation, Rey-O Copy, BNT, Overlapping Drawings, NPI, Modified CDR, Gerstmann Score, Letter Fluency. Images of the regions contributing to the network-based models are shown in Figure 2. However, predictive performance in the test set dropped considerably in the case of BDAE, Letter Fluency, and MoCA Calculation.

Figure 2: Results for tests where network-based FDG PET prediction was superior after correction for multiple comparisons.

Shown above, areas in red indicate a positive association, such that hypometabolism in these network regions resulted in predictions of more impaired clinical scores, whereas blue areas indicated a negative association where hypometabolism predicted less impaired scores. Similar to MRI data, occipital network regions showed a positive association with the visual task Overlapping Drawings. The language associated tasks BNT and Letter Fluency revealed a positive association with left sided attention and temporal-parietal network regions. The modified CDR, a measure of global function, showed a positive association with visual, temporal-parietal, and limbic network regions. Interestingly, cerebellar network regions revealed a predominantly negative association with MoCA Calculation, Letter Fluency, Overlapping Drawings, and Rey-O Figure Drawing. This was unexpected, but perhaps serves as a marker of overall disease severity. Please see Supplemental Figure 4 for a complete list of network regions that were associated with a positive or negative association.

Tau PET Models

For NPI, models with the tau-meta-ROI and clinical measures were statistically significantly better than clinical models without imaging. Models incorporating network-based tau PET uptake were statistically better than baseline or baseline and tau PET meta-ROI models for predicting Time 2 performance on Rey-Osterrieth Copy, Letter Fluency, VOSP Letters, BDAE, Ishihara, Overlapping Drawings, BNT, MoCA Calculation, NPI, CDR, CBI. Images of the regions contributing to the network-based model are shown in Figure 3. However, predictive performance in the test set dropped considerably in the case of CBI and NPI.

Figure 3: Results for tests where network-based Tau PET prediction was superior after correction for multiple comparisons.

In the above figure, areas in red indicate a positive association, such that increased tau PET uptake in these regions resulted in predictions of more impaired clinical scores, whereas blue areas indicated a negative association where increased tau PET signal predicted less impaired scores. Note that the direction of effects was reversed in order to make the figure more easily comparable to Figures 1 and 2. The language associated task BNT revealed a positive association with left sided somatomotor, default and temporal-parietal network regions, as well as right limbic network regions. Unexpectedly, there was a positive association with left, and negative association with right cerebellar network regions. A negative association with cerebellar network regions was also seen on the CBI, Modified CDR, and MoCA Calculation. As previously suggested, this may be indicative of overall disease severity, rather than suggestive of cerebellar network involvement. For the tasks Modified CDR and MocA Calculation, visual and limbic network regions are strongly represented. Visual network region involvement may suggest that the PCA patients in our cohort had disproportionately impacted scores on these clinical tasks. Please see Supplemental Figure 4 for a complete list of network regions that were associated with a positive or negative association.

Cross Modality Comparisons

For each modality, we examined the meta-ROI vs clinical baseline and network-ROI vs clinical baseline models and identified clinical tests where there was a statistically significant (p<0.05, corrected) difference in predictions for models with imaging variables compared to baseline clinical models. For example, for the MoCA Calculation task, network-based MRI, FDG PET and tau PET models all significantly outperformed clinical models. A visual representation of the statistically significant network-ROI or meta-ROI models for each imaging modality is shown in Figure 4, as well as a comparison between the network-ROI and meta-ROI models. For multiple clinical tests, including modified CDR, NPI, MoCA Calculation and BNT, network-ROI models for at least two neuroimaging modalities were significantly better than clinical models and meta-ROI models. This was also true for a single neuroimaging modality with MoCA, CBI, Gerstmann Score, BDAE, Overlapping Drawings and Rey-O figure drawing. Only for the CBI was the meta-ROI model statistically superior to the clinical model, and only for VOSP Cubes was the clinical model statistically significantly superior to the network-ROI model.

Figure 4: Modality Comparison.

In the figure above, a subset of clinical tests are represented on the x axis and the change in R-squared value (comparing network-ROI or meta-ROI models to clinical models) is represented on the y axis. The network-ROI models are represented with a solid line and meta-ROI models with a dashed line, color coded by each neuroimaging modality. The clinical tests where the respective imaging-based model is statistically significantly better than the clinical model are represented by a diamond, or by a solid circle if not significantly better, with statistical significance based on mean absolute error (MAE). As shown above, 10 network-ROI models based on at least one neuroimaging modality were associated with statistically significant improvement over clinical models. Only one meta-ROI model was statistically superior to the clinical model, and only one clinical model was statistically significantly superior to the network-ROI model.

Discussion

There exists a large amount of clinical heterogeneity in Alzheimer’s disease, not only between typical and atypical subtypes, but within subtypes, and nowhere is this more noticeable or important than in the variable rate of progression. This has implications for patients and families, who depend on prognostication to plan life with the disease. Similarly, this heterogeneity impacts clinical trial design where it may mask treatment effects or lead to spurious conclusions.⁷³ It is no surprise that more accurate prediction of the presence, nature, and severity of cognitive impairment at a future timepoint is an active area of research. Most prior studies have focused on the typical manifestation of Alzheimer’s disease, i.e. amnestic mild cognitive impairment or amnestic dementia, and have relied on global cognitive measures such as the Mini-Mental Status Examination (MMSE) or the CDR. In this study, we explored the ability of imaging measures to aid prediction of domain-specific impairment in atypical Alzheimer’s disease phenotypes. We did so using sparse regression on multimodal baseline imaging data sampled from brain networks, in combination with clinical data from LPA and PCA patients to predict decline on tests of global and selective cognitive measures. Importantly, we compared these to models comprised of solely clinical data or models using clinical data and existing of Meta ROIs, a step rarely done in previous studies. As shown above, in many cases model predictability improved with the addition of network-based neuroimaging data in addition to clinical data. Interestingly, we found that the clinical only models also performed well, as referenced above. In a small hold-out set used to test our models, predictive performance remained generally strong among imaging modalities. Our findings support the incorporation of baseline imaging measures along with baseline clinical and demographic data in the prediction of future clinical impairment in atypical Alzheimer’s disease.

Early studies used machine learning based on MRI brain scans to predict current or future clinical status, such as Alzheimer’s Disease dementia vs Mild Cognitive Impairment.^74–78 Building upon this work, in an attempt at creating a predictive model for prediction of decline on cognitive testing from baseline neuroimaging data, Duchesne et al used high dimensional linear regression of baseline MRI features of the medial temporal lobes from patients with mild cognitive impairment to predict MMSE changes over a one-year period. They compared models of solely clinical data and mixed clinical-imaging models and found that the mixed clinical-imaging models were most predictive.²⁹ Other studies have created models that predict typical and atypical AD patient decline on cognitive testing over time from baseline imaging using a variety of methods, such as ordinary least squares (OLS), sparse regularization, random forest, and support vector machines (SVM).^20–30,79 Consistently, the authors have concluded that baseline imaging measures could be used to predict longitudinal clinical decline.

One could argue that imaging may be irrelevant in predictions, and that clinical only models are suitable for prognostication. Few studies have compared baseline clinical models to imaging based or mixed imaging-clinical models, potentially leading to an overestimation of the importance of neuroimaging when creating predictive models. In another study, Zhang et al compared baseline imaging-based, clinically based, and mixed clinical-imaging based models, and found mixed clinical-imaging models created from baseline imaging across multiple neuroimaging modalities were most predictive.²⁸ To account for potential overestimation, we also compared our mixed neuroimaging-clinical based models to models comprised of only clinical data to ensure there was not overestimation of neuroimaging importance. As alluded to above, clinical only models generally performed well, although in our study very few of the clinical models performed superiorly to mixed clinical-neuroimaging based models, in keeping with the findings of the above studies. However, there were many clinical measures where the addition of imaging data did not meaningfully improve the predictive performance. One important caveat to consider regarding this finding, is that there may also be overestimation of the predictive ability of the clinical models. As the follow up time point is one year after the initial encounter, it is reasonable to think that on many global measures, such as CDR or MoCA, scores will not have changed drastically. If the prediction horizon was further from baseline, clinical-only models may suffer, and baseline imaging data may prove to be increasingly valuable over time in predictive models.

Similarly, one could consider models that incorporate simple imaging measures such as temporal meta-ROI or Alzheimer’s Disease signature cortical thickness from Tau-PET or MRI, respectively, rather than complex regional or network based measures or complicated machine learning features.^80,81 This could serve to limit model complexity, as standardized brain regions are summated into a composite term. However, our data suggest that these imaging measures add very little over baseline clinical measures. Few models were improved by adding meta-ROI imaging measures, as shown in Supplemental Table 4. Furthermore, in the majority of these cases the meta-ROI models were outperformed by imaging models that used network based data.

Consideration could be given to a combined model, utilizing all three imaging modalities. However, while the possibility exists this would provide a superior model, an alternative possibility is that one imaging modality simply dominates the model. Additionally, all three modalities may not be available for clinicians outside of specialty centers. The most readily available imaging modality outside of specialized centers is MRI, which performed well in regard to number of statistically significant models in comparison to both FDG-PET and Tau-Pet. Significant MRI based models included both global and domain specific tasks such as the MoCA and BNT, respectively. However, given the paucity of significant overlap for statistically significant models of the different clinical tasks between modalities, it is possible that separate regressions for each imaging modality may provide a unique perspective for prognostication on different clinical tasks. Although, no imaging modality appeared to necessarily excel in a specific cognitive domain, with both global and domain specific models significant for each modality. In the future, additional studies will be needed to address these important questions.

In many cases, the networks selected by our models aligned with the clinical task. As examples, atrophy of left sided temporal-parietal and attention networks predicted more impaired scores on the BNT and BDAE tasks for MRI based models, and hypometabolism in the bilateral occipital network regions predicted more impaired scores on the Overlapping Drawings task for FDG based models. However, interestingly, a small proportion of the networks ‘selected’ by our models do not intuitively appear to subserve the relevant cognitive or behavioral function (see Supplemental Figure 4 for a complete list of networks selected by imaging modality). We propose that since baseline clinical measures were included and likely provided similar information as baseline imaging findings in the regions subserving a cognitive or behavioral function, it is in fact not surprising that the selected networks were, in some instances, in an unexpected location. As an example, for digit span one would have expected frontal-parietal working memory regions. However, if baseline clinical performance is strongly correlated with imaging in these areas, they may not be selected by the model, whereas regions that add unique predictive information – perhaps denoting overall disease severity or disease stage – would be selected. This may be why, in the case of digit span for MRI based models, right hemisphere regions, such as somatosensory motor areas, were more predictive in the models. This may also explain the strong representation of cerebellar networks in Tau PET and some FDG PET models. Also unexpected was atrophy in visual network regions predicting better scores on the BNT and BDAE for MRI based models, and hypometabolism in similar regions predicting better scores on Letter Fluency for the FDG based model. An explanation for this may be that our PCA cohort scored comparatively better on these tasks.

In an attempt to prove our models were agnostic to the clinical diagnoses presented in this study, we also ran models that included baseline phenotype designation (PCA vs LPA, see Supplemental Table 1), and as shown these models have nearly identical performance to those without baseline phenotype. This is likely driven by the fact that the baseline clinical and imaging findings adequately capture the ‘effect’ of phenotype such that adding it doesn’t offer any predictive improvement. There are obvious merits in not relying on baseline phenotype, but our dataset still consisted of well phenotyped PCA and LPA patients and how our findings would generalize to mixed phenotypes remains to be seen. A truly agnostic model would provide generalizability to a larger patient base, including in cases with phenotypic overlap or unknown phenotype at baseline.

Creating such models as we have described could be utilized in both the research and clinical settings. Disease modifying agents are slowly being added to the armamentarium of AD therapies, and many are currently in development. Anti-amyloid monoclonal antibodies have recently been approved by the United States Food and Drug Administration (FDA).^82–84 As future pharmaceutical trials commence, the ability to predict decline from baseline imaging will be of vital importance for assessing treatment efficacy. Such models would also prove invaluable if agnostic to clinical phenotype. As our models are not tied to a specific phenotype, this versatility could allow use across a wide range of clinical phenotypes under a single neuropathologic entity such as AD. For a clinical trial, this would allow for assessment of treatment effects among a potentially diverse patient population. As only two phenotypes were used in our study, additional studies with a multitude of phenotypes will be needed. Additionally, from a clinical standpoint, the ability to accurately prognosticate on various clinical measures would offer both the patient and family the ability to prepare for various stages of the clinical course and ameliorate distress that comes with disease uncertainty. This is particularly true in cases where the syndrome may not neatly fit under a specific diagnosis, as there may be significant heterogeneity amongst patients. Various measures to ease symptomatologic burden could be implemented prophylactically.

A strength of our study involves the volume of clinical data acquired. The data acquired are from one of the largest prospective atypical AD cohorts described, with a variety of longitudinal clinical and imaging modalities. This allowed us to create a plethora of unique models. Limitations include a demographic which is predominately Caucasian with an average education level achieved that surpasses the national average, thus potentially limiting generalizability. Only LPA and PCA patients were included in our study due to the other atypical AD phenotypes being less common variants and not being well represented in our database. Additionally, in our holdout set of participants used to test our models, very few underwent FDG PET, limiting the conclusions we can reliably draw on this dataset, although the training data appear strong. In the future, models incorporating data from both typical and other atypical AD phenotypes will be necessary for truly agnostic models. Another limitation is that only two time points were used, one year apart. Models where predictability spans multiple time points, over a multi-year period will be the objective. In addition, as addressed above, clinical-only models may be overly predictive, given the short time span. In future studies, whether clinical only models continue to show high predictability, or whether there is a divergence over time in comparison to mixed clinical-imaging models, will be a point of interest. Finally, our meta-ROI models were created as has previously been described for typical AD. In the future, as the ability to calculate a baseline imaging value on the respective atypical AD meta-ROI, these may perhaps serve as better comparisons. These limitations will need to be addressed in future studies in the anticipation of producing increasingly predictive models.

In summary, we investigated the use of baseline multimodal imaging utilizing MRI, FDG-PET, and Tau-PET to predict clinical decline in patients with atypical Alzheimer’s Disease (AD), specifically logopenic variant primary progressive aphasia or posterior cortical atrophy. We created predictive models using penalized regression and found that baseline imaging data in combination with baseline clinical data significantly forecasted future cognitive performance, more so than those using only clinical data or a meta-region of interest (ROI). This merits potential for management of patients with heterogeneous clinical syndromes and assessment of efficacy of future disease-modifying therapies.

Table 3:

MRI Model Results

								Train			Test
					vs clinical			vs meta-ROI
Instrument	N	Network-ROIs	R-sq	MAE	Δ R-sq	Δ MAE	p	Δ R-sq	Δ MAE	p	N	R-sq	MAE
MoCA	46	24	0.939	1.068	0.221	−0.687	<0.001**	0.198	−0.671	<0.001**	13	0.799	1.393
CBI	42	1	0.690	10.573	−0.027	−0.595	0.285	-	-	-	12	0.280	13.677
AVLT Delayed Recall	46	3	0.577	1.283	−0.006	0.027	0.302	-	-	-	10	0.628	0.808
WMS-III Visual Reproduction II	41	1	0.000	27.220	−0.065	1.228	0.125	-	-	-	8	0.195	20.042
Rey-O Copy	44	11	0.862	2.593	0.056	−0.206	0.040	0.037	0.233	0.104	10	0.837	1.741
VOSP Letters	42	1	0.724	1.959	−0.002	0.008	0.620	-	-	-	11	0.711	2.456
VOSP Cubes	41	0	-	-	-	-	-	-	-	-	-	-	-
BNT	44	7	0.793	0.930	0.133	−0.277	0.008**	0.117	−0.400	0.020*	8	0.572	2.319
BDAE	45	14	0.936	0.382	0.162	−0.212	<0.001**			-	9	0.420	0.616
WMS-III Digit Span	46	8	0.717	1.496	0.116	−0.061	0.151	0.117	−0.015	0.047*	10	0.660	2.296
MoCA Calculation	45	19	0.897	0.301	0.312	−0.182	<0.001**	0.316	−0.151	<0.001**	7	−0.357	0.605
Modified CDR	42	1	0.553	1.480	−0.053	0.066	0.171	−0.016	0.000	0.339	12	0.726	2.797
Letter Fluency	44	0	-	-	-	-	-	-	-	-	-	-	-
NPI	44	0	-	-	-	-	-	-	-	-	-	-	-
Gerstmann Score	29	0	-	-	-	-	-	-	-	-	-	-	-
Color Matching	41	12	0.763	0.360	0.180	0.080	0.335	0.161	0.013	0.062	7	0.466	0.631
Ishihara	41	1	0.803	0.695	−0.002	0.008	0.175	-	-	-	11	0.941	0.303
Surface Dyslexia	40	3	0.486	0.954	0.080	0.278	0.216	0.016	0.056	0.900	11	0.198	0.983
Overlapping Drawings	41	3	0.779	0.607	0.069	−0.251	0.048**	-	-	-	11	0.823	0.556
Navon	41	7	0.754	0.756	0.070	−0.109	0.062	0.055	−0.150	0.368	11	0.518	0.884
Simultanagnosia	41	1	0.819	2.087	−0.011	0.295	0.280	-	-	-	11	0.922	1.779
Facial Recognition	42	0	-	-	-	-	-	-	-	-	-	-	-
UPDRS III	46	8	0.822	2.866	0.086	−0.012	0.119	0.100	0.163	0.064	13	0.329	3.450
WAB Praxis	46	2	0.749	1.574	0.000	0.027	0.940	−0.007	−0.039	0.957	12	0.767	1.222

^*p<0.05 (unadjusted for multiple comparisons)

^**p<0.05 (adjusted for multiple comparisons)

^-omitted since one/both imaging modalities had zero imaging coefficients included in the selected model

Results for the models using clinical data from the baseline visit in addition to MRI data at the network level from baseline visit. Columns indicate the number of cases included in each model (N), the number of networks with non-zero coefficients (Networks), the R-squared (R-sq) and median absolute error (MAE), and the p-values for the Wilcoxon paired rank sum tests comparing network-based models to clinical-only and clinical and meta-ROI models. As shown above, 6/19 models were statistically significantly improved by the addition of MRI network data in addition to baseline. Additionally, 4/11 of the Network based models were statistically significantly superior to meta-ROI based models. Some models were omitted due to zero imaging coefficients in the selected model.

Table 4:

FDG PET Model Results

	Train										Test
					vs clinical			vs meta ROI
Instrument	N	Network-ROIs	R-sq	MAE	Δ R-sq	Δ MAE	p	Δ R-sq	Δ MAE	p	N	R-sq	MAE
MoCA	43	8	0.824	1.455	0.103	−0.271	0.131	0.084	−0.196	0.110	7	0.812	1.631
CBI	39	5	0.780	9.403	0.069	−1.915	0.028*	−0.027	0.724	0.046*	7	0.340	14.749
AVLT Delayed Recall	43	5	0.643	1.091	0.092	−0.004	0.359	-	-	-	5	0.357	0.467
WMS-III Visual Reproduction II	38	2	0.204	22.339	0.134	−3.924	0.268	−0.012	2.371	0.243	4	0.277	16.190
Rey-O Copy	41	13	0.923	1.551	0.115	−1.105	0.001**	-	-	-	6	0.861	1.123
VOSP Letters	39	2	0.800	1.469	0.033	−0.367	0.139	-	-	-	6	0.670	2.516
VOSP Cubes	39	1	0.721	1.225	−0.050	0.178	0.007**	-	-	-	6	0.325	1.941
BNT	41	12	0.926	0.807	0.266	−0.422	<0.001**	-	-	-	5	0.803	2.358
BDAE	42	4	0.834	0.882	0.052	0.295	0.102	-	-	-	5	0.045	0.875
WMS-III Digit Span	43	10	0.802	1.357	0.193	−0.169	0.079	0.180	−0.034	0.122	5	0.360	3.493
MoCA Calculation	42	17	0.910	0.290	0.327	−0.222	<0.001**	0.328	−0.217	<0.001**	3	−1.575	0.803
Modified CDR	39	7	0.803	1.152	0.200	−0.250	0.013**	0.113	−0.077	0.008**	7	0.787	1.854
Letter Fluency	41	13	0.926	2.324	0.082	−1.210	0.003**	-	-	-	6	0.401	5.586
NPI	41	6	0.698	1.459	0.146	−0.172	0.002**	0.112	−0.023	0.090	7	0.273	1.411
Gerstmann Score	27	1	0.124	1.375	0.124	−0.281	0.008**	−0.111	0.110	0.279	1	NA	NA
Color Matching	38	7	0.736	0.217	0.125	−0.062	0.129	0.119	−0.055	0.220	3	0.620	0.221
Ishihara	38	4	0.837	0.541	0.020	−0.224	0.137	-	-	-	6	0.946	0.466
Surface Dyslexia	37	5	0.616	0.808	0.171	0.154	0.070	-	-	-	6	−0.065	0.825
Overlapping Drawings	38	7	0.871	0.373	0.164	−0.485	0.005**	0.160	−0.485	0.009**	6	0.917	0.400
Navon	38	2	0.692	0.994	0.033	0.129	0.397	-	-	-	6	0.675	0.776
Simultanagnosia	38	5	0.856	2.083	0.019	0.281	0.287	0.016	0.308	0.556	6	0.863	1.469
Facial Recognition	39	4	0.405	0.452	0.091	0.047	0.241	0.099	−0.021	0.167	7	0.295	0.542
UPDRS III	43	7	0.791	2.626	0.055	−0.139	0.278	0.042	−0.198	0.267	8	0.270	3.301
WAB Praxis	43	4	0.815	1.596	0.069	0.049	0.218	0.043	0.079	0.262	7	0.840	1.229

^*p<0.05 (unadjusted for multiple comparisons)

^**p<0.05 (adjusted for multiple comparisons)

^-omitted since one/both imaging modalities had zero imaging coefficients included in the selected model

As shown above, 9/24 models were statistically significantly improved by the addition of Network data. Additionally, 3/14 Network based models were statistically significantly superior to meta ROI based models. Columns indicate the number of cases included in each model (N), the number of networks with non-zero coefficients (Networks), the R-squared (R-sq) and mean absolute error (MAE), and the p-values for the Wilcoxon paired rank sum tests comparing network-based models to clinical-only and clinical and meta-ROI models. Some models were omitted due to zero imaging coefficients in the selected model.

Table 5:

Tau PET Model Results

	Train										Test
						vs clinical		vs meta-ROI
Instrument	N	Network ROIs	R-sq	MAE	Δ R-sq	Δ MAE	p	Δ R-sq	Δ MAE	p	N	R-sq	MAE
MoCA	46	4	0.820	1.917	0.101	0.162	0.368	0.088	−0.075	0.254	13	0.799	1.393
CBI	42	9	0.872	5.537	0.155	−5.631	<0.001**	0.136	−4.999	<0.001**	12	0.280	13.677
AVLT Delayed Recall	46	10	0.731	0.874	0.148	−0.381	0.069	0.128	−0.217	0.141	10	0.628	0.808
WMS-III Visual Reproduction II	41	1	0.175	21.863	0.110	−4.128	0.134	−0.086	−0.246	0.599	8	0.195	20.042
Rey-O Copy	44	10	0.875	1.844	0.069	−0.954	0.047*	0.037	−0.776	0.199	10	0.837	1.741
VOSP Letters	42	4	0.782	1.718	0.056	−0.233	0.049*	-	-	-	11	0.711	2.456
VOSP Cubes	41	0	-	-	-	-	-	-	-	-	-	-	-
BNT	44	19	0.941	0.697	0.280	−0.510	<0.001**	0.268	−0.828	<0.001**	8	0.572	2.319
BDAE	45	7	0.886	0.674	0.111	0.080	0.024*	0.134	−0.249	0.002**	9	0.420	0.616
WMS-III Digit Span II	46	5	0.701	1.353	0.100	−0.204	0.203	0.084	−0.253	0.133	10	0.660	2.296
MoCA Calculation	45	11	0.726	0.400	0.141	−0.083	0.012**	0.144	−0.041	0.009**	7	−0.357	0.605
Modified CDR	42	11	0.897	0.577	0.291	−0.837	<0.001**	0.252	−0.493	<0.001**	12	0.726	2.797
Letter Fluency	44	5	0.880	2.498	0.031	−0.837	0.042*	0.022	−0.725	0.057	12	0.792	4.133
NPI	44	5	0.594	1.385	0.046	−0.245	0.003**	0.048	−0.168	0.001**	13	0.109	2.072
Gerstmann Score	29	0	-	-	-	-	-	-	-	-	1	-	-
Color Matching	41	6	0.575	0.287	−0.008	0.007	0.555	-	-	-	7	0.466	0.631
Ishihara	41	5	0.857	0.627	0.052	−0.061	0.026*	-	-	-	11	0.941	0.303
Surface Dyslexia	40	3	0.487	0.839	0.081	0.162	0.150	0.029	0.121	0.361	11	0.198	0.983
Overlapping Drawings	41	3	0.795	0.484	0.085	−0.374	0.015*	-	-	-	11	0.823	0.556
Navon	41	2	0.696	0.912	0.012	0.047	0.316	-	-	-	11	0.518	0.884
Simultanagnosia	41	2	0.845	1.858	0.015	0.065	1.000	-	-	-	11	0.922	1.779
Facial Recognition	42	0	-	-	-	-	-	-	-	-	-	-	-
UPDRS III	46	4	0.792	1.884	0.056	−0.993	0.053	0.041	−0.887	0.053	13	0.329	3.450
WAB Praxis	46	3	0.828	1.460	0.080	−0.087	0.084	-	-	-	12	0.767	1.222

^*p<0.05 (unadjusted for multiple comparisons)

^**p<0.05 (adjusted for multiple comparisons)

^-omitted since one/both imaging modalities had zero imaging coefficients included in the selected model

As shown above, 11/21 models were statistically significantly improved by the addition of Network data. Additionally, 6/14 Network based models were statistically significantly superior to meta ROI based models. Columns indicate the number of cases included in each model (N), the number of networks with non-zero coefficients (Networks), the R-squared (R-sq) and mean absolute error (MAE), and the p-values for the Wilcoxon paired rank sum tests comparing network-based models to clinical-only and clinical and meta-ROI models. Some models were omitted due to zero imaging coefficients in the selected model.

Data Availability:

Raw data are not publicly available due to ethical barriers (no consent from participants), but can be requested from the senior author at botha.hugo@mayo.edu under the condition of a data sharing agreement.