Disease aggressiveness signatures of amyotrophic lateral sclerosis in white matter tracts revealed by the D50 disease progression model

Abstract Numerous neuroimaging studies in amyotrophic lateral sclerosis (ALS) have reported links between structural changes and clinical data; however phenotypic and disease course heterogeneity have occluded robust associations. The present study used the novel D50 model, which distinguishes between disease accumulation and aggressiveness, to probe correlations with measures of diffusion tensor imaging (DTI). DTI scans of 145 ALS patients and 69 controls were analyzed using tract‐based‐spatial‐statistics of fractional anisotropy (FA), mean‐ (MD), radial (RD), and axial diffusivity (AD) maps. Intergroup contrasts were calculated between patients and controls, and between ALS subgroups: based on (a) the individual disease covered (Phase I vs. II) or b) patients' disease aggressiveness (D50 value). Regression analyses were used to probe correlations with model‐derived parameters. Case–control comparisons revealed widespread ALS‐related white matter pathology with decreased FA and increased MD/RD. These affected pathways showed also correlations with the accumulated disease for increased MD/RD, driven by the subgroup of Phase I patients. No significant differences were noted between patients in Phase I and II for any of the contrasts. Patients with high disease aggressiveness (D50 < 30 months) displayed increased AD/MD in bifrontal and biparietal pathways, which was corroborated by significant voxel‐wise regressions with D50. Application of the D50 model revealed associations between DTI measures and ALS pathology in Phase I, representing individual disease accumulation early in disease. Patients' overall disease aggressiveness correlated robustly with the extent of DTI changes. We recommend the D50 model for studies developing/validating neuroimaging or other biomarkers for ALS.


| INTRODUCTION
Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disease predominantly characterized by the loss of motor-neuron function; however, given its well documented-effects on extra-motor domains of the central nervous system, it is now considered a multisystem disease. Prognoses for this condition remain poor; the median survival is 3 years after symptom onset (Dorst, Ludolph, & Huebers, 2018;Kiernan et al., 2011). Despite considerable ongoing efforts to develop curative therapies, Riluzole is still the only globally approved disease-modifying therapy (Miller, Mitchell, Lyon, & Moore, 2007;Petrov, Mansfield, Moussy, & Hermine, 2017). In principle, ALS care and research face one major challenge: the disease is characterized by substantial intraindividual and interindividual phenotypic heterogeneity, with highly variable progression and survival rates (Simon et al., 2014;Westeneng et al., 2018).
Reliable biomarkers that can capture patients' individual disease aggressiveness are therefore urgently needed. Magnetic resonance imaging (MRI) has already been recognized as a promising noninvasive technique that provides insights into the neurodegenerative mechanisms underlying ALS Steinbach, Gaur, Stubendorff, Witte, & Grosskreutz, 2018;Turner et al., 2011). In particular, the use of diffusion tensor imaging (DTI) for capturing ALS associatedpathology in white-matter (WM) fibertracts and classifying different disease subtypes has been established Muller et al., 2016). Several DTI-based studies have demonstrated ALSassociated loss of WM structural integrity on a case-control-level (e.g., in the corticospinal tract [CST] and corpus callosum), albeit with inconsistent and controversial results, for example, regarding the extent and significance of affected regions (Bede & Hardiman, 2014;Floeter, Danielian, Braun, & Wu, 2018;Li et al., 2012;Zhang et al., 2018). Most important, former DTI studies frequently failed to show correlations of derived structural abnormalities with patients' clinical disease severity (Canu et al., 2011;Geraldo et al., 2018;Keller et al., 2011;Rose et al., 2012;Senda et al., 2011). Within the ALS and wider neurodegenerative research community, this discrepancy is often referred to as the "correlation gap" .
It has been postulated, that longitudinal MRI data would offer the best insight into neuroimaging alterations across the disease course, thus having the potential to close this gap. However, the nature of ALS causes relevant obstacles for serial MRI assessments during patients' course of the disease, for example, due to cognitive impairment, bulbar and respiratory symptoms and poor longevity (Chio et al., 2014;Filippi et al., 2015). It is therefore not surprising, that former longitudinal studies captured rather small sample sizes with assumable selection bias and/or short disease periods with often not more than two to three time points (Agosta et al., 2009;Alruwaili et al., 2019;Bede & Hardiman, 2018;Keil et al., 2012;Menke, Proudfoot, Talbot, & Turner, 2018;Steinbach et al., 2015;van der Graaff et al., 2011). Alternatively, large-scale cross-sectional datasets can provide valuable pseudo-longitudinal substitutes, but require a quantitative understanding of individual disease state and trajectory at the time of MRI acquisition.
The recently developed D50 model provides a novel way to quantify disease aggressiveness separately from disease accumulation even in highly heterogenous ALS cohorts. Briefly, the model incorporates a sigmoidal decline of the ALS Functional Rating Scale (ALSFRS-R) throughout the disease course because a curvilinear progression was suggested before (Franchignoni, Mora, Giordano, Volanti, & Chio, 2013;Gordon et al., 2010;Senda et al., 2017). The D50 model addresses the difficulty in characterizing functional loss at the individual and population level using traditional clinical indices like the progression rate (PR) (Poesen et al., 2017;Prell, Gaur, Steinbach, Witte, & Grosskreutz, 2020;Steinbach et al., 2020). As such, the model quantitatively (a) captures phenotypic complexity, (b) reduces noise associated with traditional disease progression parameters, and (c) provides distinct descriptors of patients' aggressivity and individual disease covered.
We hypothesize that these parameters close the correlation gap between clinical characteristics of ALS and in vivo measures of cerebral structural integrity. The present study applied tract-based spatial statistics (TBSS) in a large cohort of patients with detailed D50 model-based descriptors of the ALS disease process to quantify the tract damage ALS causes depending on disease aggressiveness and phases of individual disease covered.

| Subjects
This study was approved by the local Ethics committee ) and all experimental procedures were performed in accordance with the 1964 Declaration of Helsinki; written informed consent was obtained from individual participants prior to enrolment.
All participants were consecutively recruited from the Department of Neurology at Jena University Hospital. Available neuroimaging of participants, based on a harmonized MRI protocol, were inspected (FLAIR and T1 weighted structural images) by a trained analyst (R. S.) and neuroradiologist (T. M.) and excluded if any relevant intracranial pathologies (e.g., tumors, cysts, stroke, or bleedings) were present.
All patients with ALS were examined by a trained neurologist at enrollment and follow-up visits (minimum of two assessments for each patient) and met the criteria of definite, probable or laboratory-supported probable ALS (Brooks, Miller, Swash, Munsat,, & World Federation of Neurology Research Group on Motor Neuron, 2000). Exclusion criteria included the presence of (a) juvenile ALS, (b) primary lateral sclerosis, (c) clinically relevant dementia symptoms, (d) any comorbidities that could affect motor performance and (e) a D50 value above 100 months. Finally, 145 patients and 69 healthy controls were included in the analyses (for a CONSORT diagram refer to Supplementary Figure S1).

| The D50 disease progression model
The D50 model provides parameters of overall disease aggressiveness, local disease activity, and individual disease covered (Poesen et al., 2017;Prell et al., 2020;Steinbach et al., 2020). Figure 1 provides a pictorial overview of how the model estimation was performed based on all the regularly collected ALSFRS-R scores that were available per patient ( Figure 1a). Briefly, the D50 model describes the disease course of individual ALS patients as a sigmoidal state transition from full health to functional loss. The curve is calculated using iterative fitting of available ALSFRS-R scores ( Figure 1b). The value dx describes the time constant of ALSFRS-R decline and the value D50 is defined as the estimated time taken in months for a patient to lose 50% of his/her functionality (equivalent to an ALSFRS-R score of 24). Given that dx and D50 correlate linearly in ALS cohorts (Poesen et al., 2017;Prell et al., 2020), the D50 value provides a unified descriptive measure of individual patients' overall disease aggressiveness. This allowed us to classify patients as having either low (D50 ≥ 30 months) or high (D50 < 30 months) disease aggressiveness. The cutoff value of 30 months corresponds to the median of D50 values (here: 28.8 months), that is typically observed in comparable cohorts of patients with ALS treated at our center (Prell et al., 2020;Steinbach et al., 2020). We excluded patients with a D50 value above 100 months and only one ALSFRS-R score, to ensure a high level of reliability for the calculation of the D50 model.
Normalizing an individual's real-time disease trajectory to D50 yields the parameter relative D50 (rD50), an open-ended linear reference scale where 0 signifies symptom onset and 0.5 indicates the time point of halved functionality. Patients can be categorized into at least 3 phases: an early semistable Phase I (0 ≤ rD50 < 0.25), an early progressive Phase II (0.25 ≤ rD50 < 0.5), and late progressive and late stable Phases III/IV (rD50 ≥ 0.5; Figure 1c).
The model also yields two descriptors of local disease activity that can be calculated at any given time point across the patient's disease course (here at MRI); namely the calculated functional state (cFS) and the calculated functional loss (cFL) (Figure 1b). These provide a measure of functional deterioration which greatly reduce the noise inherent to the ALSFRS-R score (Bakker et al., 2020) and provide adjacent averaging. This was particularly useful in 41 patients who did not receive ALSFRS-R scoring at the time of MRI. The other 114 patients received an ALSFRS-R assessment within 10 days prior to or after MRI acquisition (see Supplementary Table S1 for the clinical data of this subcohort, comparing traditional, and D50 disease metrics). Thus, the D50 model enabled the inclusion of all ALS patients in the subsequent neuroimaging analyses. Figure 1d presents histograms of relevant variables used for the calculation of the D50 model and resulting parameters. The distributions of these variables indicate that the ALS cohort of this study (n = 145) well represents the regional ALS population available at our center (n = 420). The original DICOM images were converted into the Nifti format using the Dcm2Nii (MRIcroN, version 4/2010) script. All subsequent pre-processing steps were performed using the FMRIB Software Library (FSL, version 5.0, http://fsl.fmrib.ox.ac.uk/fsl/). Initial steps included brain-tissue extraction and correction for possible eddycurrent induced distortions. A diffusion tensor model was applied to each voxel using DTIFIT to calculate maps for fractional anisotropy (FA); mean diffusivity (MD); axial diffusivity (AD, corresponding to the first eigenvector L1); and radial diffusivity (RD, calculated as average of the L2 and L3 maps). All FA images were visually inspected (by R. S.) and excluded if any image artifacts were observed.

| MRI data acquisition and processing of DTI images
Additional processing followed the standardized pipeline for TBSS analyses. First, each subject's FA image was nonlinearly registered to a standard FA template (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FMRIB58_FA), and then averaged to a study-specific template to which each subject's FA image was subsequently nonlinearly registered. Thereafter, a mean FA image was created, thresholded at 0.2 and thinned to obtain a mean FA skeleton. These projection-steps upon the FA skeleton were subsequently also applied to the other non-FA data images (MD, AD, and RD). Note: Continuous data are summarized for † as mean ± SD or for ‡ as median ± interquartile range (each with the total range in brackets). For § categorial data, the number of cases and percentages are given. Some variables are time-point dependent throughout patients' individual disease course and refer to the day of MRI acquisition (as labeled with "at MRI"), the others depict constant characteristics of patients' overall disease course. Riluzole intake is reported in the context of (a) at MRI acquisition (b) a relevant intake (>50%) of 100 mg/day throughout the entire disease course (from symptom-onset until end of study/time of death). Abbreviations: ALS, amyotrophic lateral sclerosis; cFL, calculated functional loss; cFS, calculated functional state; D50, estimated time in months for an individual to lose 50% of functionality; MRI, magnetic resonance imaging; rD50, relative D50.

| Statistical analyses
were older, had a lower cFS and presented more frequently with bulbar onset than those with low aggressiveness (D50 > 30 months).
All of the above-mentioned potential confounding variables were subsequently included as nuisance covariates for the intergroup TBSS analyses. For an overview of the subgroups and reasoning for the chosen covariates, see Supplementary Table S2. 3.2 | Widespread WM changes in ALS patients relative to controls At the case-control level, patients with ALS exhibited widespread decreases of FA in several WM tracts (Figure 2; p < .001; corrected for age and gender), including the bilateral CST, body/genu of the F I G U R E 2 Between-group differences for patients and healthy controls. Tract-based spatial statistics (TBSS) between-group comparisons of patients (n = 145) and healthy controls (n = 69) revealed widespread ALS-related WM changes. These were most evident in decreased FA values, mainly located at bilateral CSTs, body/genu of the corpus callosum, adjacent corona radiata, and WM of bilateral frontal, parietal, temporal, occipital lobes, brainstem, and right-cerebellar pathways. RD increases for the ALS cohort were colocalized with the FA changes (although less widespread), while the MD increases affected mainly the bilateral CSTs and the body of the corpus callosum. When applying the same significant threshold, no differences were noted for AD between patients and controls. No significant between-group differences for AD contrasts, FA increases, or MD/RD decreases were noted.  the ALSFRS-R that anatomically were well colocalized with the contrasts of correlations with the cFS (Figure 5a,b). Again, significant and extensive patterns could be revealed in correlation contrasts with the D50 value, most important for MD and AD contrasts and also to a smaller extent for the RD (Figure 5c). For the PR, no significant correlations could be shown with any DTI-measure for this group of 114 ALS patients (Figure 5d).

| DISCUSSION
The present study is the first to comprehensively evaluate tractrelated structural integrity in patients with ALS by distinguishing between individual accumulated disease/disease covered and disease aggressiveness. These disease characteristics were defined and assessed using the novel D50 model. Figure 6 gives a pictorial overview of the  reported in TBSS case-control analyses (Alshikho et al., 2016;Cardenas-Blanco et al., 2016;Trojsi et al., 2019;Zhang et al., 2018 (Cirillo et al., 2012;de Albuquerque et al., 2017;Prudlo et al., 2012;Sage et al., 2009;Trojsi et al., 2015). Conversely, studies with smaller cohorts (n = 12-15) reported no significant correlations with the ALSFRS-R (Geraldo et al., 2018;Metwalli et al., 2010;Rose et al., 2012), which likely demonstrates a sample size effect. Given that our regression analyses revealed no additional changes in the supratentorial tracts (no changes for FA/RD noted either) for patients in Phase II, these results suggest that disease accumulation-driven changes primarily occur in Phase I (Table 2; Supplementary Figure S3a,b). In keeping with this, direct subgroup comparisons between Phase I and II patients did not reveal any difference in DTI metrics. This may potentially explain why Keller et al. (2011) reported no correlations between FA values and ALSFRS-R scores for their cohort of 33 patients; it is possible that only patients in Phase II were analyzed given that the highest reported ALSFRS-R score was 40 points (compare to Figure 1c). Here we show that in Phase II, the only significant correlations noted are within bilateral cerebellum between the cFS and MD/AD elevations and rD50 and AD increases.
Cerebellar involvement in ALS has been previously reported by DTI studies (Keil et al., 2012;Menke et al., 2018;Prell et al., 2013;Trojsi et al., 2015); our data indicate that the dynamic structural changes in the cerebellum are a feature of ongoing disease in Phase II.
Of particular interest is the behavior of measurable FA changes across the disease course: although significant and widespread FA decreases reflect ALS-driven pathology at the case-control level (Figures 2 and 6a We demonstrated that the D50 model is well suited for crosssectional analyses and allows comparisons between highly heterogeneous patients. One of the advantages of this approach is the ability to include patients who did not receive ALSFRS-R assessment at the time of MRI acquisition. Traditionally, these patients would have been excluded from further analyses, thus bearing the potential of selection bias. We here used the cFS instead, which showed stable associations with DTI-signals (Figures 4 and 5b). We further show, that the correlations revealed with the cFS are comparable to those with the original ALSFRS-R scores in the respective subgroup analyses (Figure 5a,b).
F I G U R E 6 Legend on next page.
The reliability of the parameter cFS is further supported by the observation that the averaged differences between model-derived cFS and original ALSFRS-R scores per patient had a median of −0.1 points (interquartile range 0.25; total range: −1.68 to 0.72; Figure 1d). We conclude that the cFS is a usable marker for a patient's remaining functionality, which can be calculated for any given time point and decreases the noise typically associated with ALSFRS-R assessments, for example, due to different raters (Bakker et al., 2020;Franchignoni et al., 2013).
Our approach may also help guide future studies aimed at longitudinally analyzing TBSS correlates of ALS-related functional loss.
These are more

| Stable signatures of ALS disease aggressiveness in WM tracts revealed with DTI
We further demonstrated that individuals with higher overall disease aggressiveness (i.e., lower D50) display DTI changes within established core regions of ALS pathology. We observed substantial MD and AD elevations in fronto-parietal long association tracts. The involvement of these pathways has been previously reported; AD elevations are often concomitant with cognitive and/or behavioral impairment (Alruwaili et al., 2018;Kasper et al., 2014), both of which have been noted in patients with more rapidly progressive disease forms (Bock et al., 2017;Calvo et al., 2017).
In the D50 model, overall disease aggressiveness is captured by parameter D50 whereas disease activity at the time of MRI is represented by cFL; cFL yielded significant correlations for patients in Phase I and not those in Phase II or the entire ALS cohort. This is curious, regarding the relatively low interindividual range of the parameter cFL in Phase I, while cFL variance substantially increases for Phase II patients (Supplementary Table S2). Again, this underscores DTI's sensitivity toward even subtle changes during very early disease.
Here, using cFL (indicator of local disease activity) and D50, we have demonstrated consistent and robust associations between DTIquantified changes and disease aggressiveness in ALS. Similar analyses have been previously performed using the PR parameter that has inherent limitations. As an index, the PR presumes that progression in ALS is linear and thus does not fully capture the patients' progression throughout the whole disease course (Franchignoni et al., 2013;Gordon et al., 2010;Simon et al., 2014). This may explain why in the regression analyses of patients with an available ALSFRS-R no significant correlations with the parameter PR could be revealed, while D50 showed stable correlations (Figure 5c,d).
Unsurprisingly, the majority of prior TBSS studies did also not observe any relevant voxel-wise correlations between the PR and DTI metrics (Alshikho et al., 2016;Geraldo et al., 2018;Kopitzki et al., 2016;Rose et al., 2012;Senda et al., 2017). Cirillo et al. (2012) noted an inverse correlation between FA and the PR within the CSTs F I G U R E 6 Summary of significant tract-based spatial statistics (TBSS) results in association with the D50 model. The plots refer to suprathreshold Voxels of significant results for the respective TBSS analyses. The histograms show the number of voxels for the respective p-value level (with increasing significance from left to right on the x-axis). (a,c-e) Comparing ALS patients to healthy controls revealed the most relevant significant differences for FA (a). However, accumulated disease (rD50) did not correlate with the FA, but with MD, RD and AD (c), for the latter being most evident for the subcohort of Phase I patients (d) and still detectable within Phase II patients (e). No significant differences were noted when comparing patients in Phase I with those in Phase II (data not shown). (b,f-h) The subcohort comparison of patients' D50-derived overall disease aggressiveness revealed the most significant differences for MD and AD (b), which was comparable in regression analyses within the entire ALS cohort (f). For AD, this was mainly driven by the Phase II patients (h), while only minor AD correlations were noted for Phase I patients (g). Again, no associations between FA and D50 were observed. In general, the results showed stable associations between increased AD and higher disease aggressiveness (lower D50), located in long fronto-central associations tracts. (i-k) The regression analyses with the parameter cFS showed in principle comparative patterns as in the concerning analyses with the parameter rD50, but in addition also relevant correlations with FA (i). These were additionally confirmed within the subgroup of Phase I patients (j) and for Phase II patients there were again only minor correlations with AD/MD (in bilateral cerebellum). (l) Significant correlations with the parameter cFL were found only within the subcohort of Phase I patients, not for Phase II patients or the entire ALS-cohort, which again underlines the capability of DTI to detect changes even during the relatively early stable disease Phase I. ALS, amyotrophic lateral sclerosis; D50, estimated time in months for an individual to lose 50% of functionality; cFL, calculated functional loss, cFS, calculated functional state, rD50, relative D50, each at the time of MRI Here, we provide a rational foundation for the use of D50 modelderived parameters, especially since the D50 value itself captures patients' disease aggressiveness based on their individual disease trajectories.

| Limitations
The present study is not without its limitations. Standardized neuropsychological and genetic profiles were unavailable for the entire cohort. Nevertheless, we posit that given its size, the cohort is representative of the variance within the regional patient population.
Accordingly, the distribution of representative disease characteristics mirrored all the ALS patient data available at our neuromuscular center ( Figure 1d).

| CONCLUSION
The present study has demonstrated the potential of the D50 model for characterizing highly heterogenous cross-sectional ALS cohorts and generating informative pseudo-longitudinal data while accounting for strongly varying phenotypic presentation. Within this framework, we show that TBSS analyses are particularly sensitive toward disease accumulation-driven changes in early disease (Phase I). Second, characteristic structural WM changes are strongly associated with higher disease aggressiveness. Taken together, these observations suggest that DTI reflects disease activity, which appears to remain stable over time. Therefore, DTI may constitute a robust readout for the detection of treatment-induced reduction of disease activity prior to obvious clinical improvement.
Finally, we show that longitudinal clinical data capture in ALS should be tailored to individual disease trajectories, especially within mixed patient groups of differing progression types. We recommend using the D50 model to stratify patients in further studies developing or validating neuroimaging or other biomarkers for ALS.