Multimodal Neuroimaging‐Guided Stratification in Amyotrophic Lateral Sclerosis Reveals Three Disease Subtypes: A Multi‐Cohort Analysis

ABSTRACT

Amyotrophic lateral sclerosis (ALS) is a multisystem disease with marked pathophysiological and clinical heterogeneity, making individual and objective characterization of the degree of disease progression and disease‐related subtrajectories challenging. Here, we use in vivo multimodal neuroimaging data and computational models to generate personalized indices of ALS progression and subtrajectory. We used structural and diffusion weighted imaging of 691 participants (58% ALS) from two independent ALS data sets (North American and Utrecht cohorts) to extract regional values of grey matter (DM) density and white matter (WM) microstructural integrity. Contrastive trajectory inference (cTI) allowed us to identify and separate latent, multivariate patterns in neuroimaging features highlighting ALS‐associated pathological processes, which were used to generate subject‐specific indices of disease progression and subtrajectory. Disease subtrajectories were based on distinct patterns of alterations in neuroimaging data considering subjects at different disease progression levels. The neuroimaging‐based, personalized index of disease progression is indicative of clinical symptom severity (North American: p < 0.01 and Utrecht: p < 0.01) and displays alignment with the King's College staging system (p = 0.001 and p = 0.002). Three ALS subtrajectories were identified that displayed distinct alterations in the motor, limbic system, and widespread cortical and subcortical changes that also differed in clinical symptom manifestation. Our analysis has shown that neuroimaging data encodes subject‐specific, disease‐related patterns that can be leveraged to obtain an in vivo proxy of disease progression and putative disease subtype.

Amyotrophic lateral sclerosis (ALS) is recognized as a multifaceted disease with varying symptom profiles calling for a more comprehensive understanding of ALS. We demonstrate the feasibility of extracting personalized disease progression markers as well as distinct disease subtrajectories from in vivo multimodal neuroimaging data in a heterogeneous ALS population.

Summary.

• We used data‐driven and unsupervised methods to extract patient‐specific indices of disease progression and subtrajectory based on distinct alterations of neuroimaging data in ALS subjects.

• Results were largely reproduced in a secondary large‐scale ALS data set.

• Findings suggest that individual pathological patterns are encoded in structural and microstructural MRI data that can be recovered and leveraged to provide an objective index of disease progression on a subject‐specific level.

1. Introduction

Amyotrophic lateral sclerosis (ALS) is a fatal neurological disorder that predominately affects the upper and lower motor neurons (UMN and LMN) in the brain and spinal cord, causing progressive weakness in limb and bulbar muscles as well as axial and respiratory (Kalra et al. 2020). Additionally, it is recognized that the pathology extends beyond the motor system of the brain, presenting in cognitive and/or behavioral impairments that share commonalities with frontotemporal dementia (FTD) (Ferrea et al. 2021). The clinical presentation of ALS patients is heterogeneous in both symptom manifestation and prognosis, leading to a considerable degree of variability in disability profiles, disease progression rates, and survival times. Thus, providing an accurate prognosis of a patient's disease course remains challenging for clinicians.

It has become clear that individualized therapeutic strategies are needed for the optimal management of ALS. Clinically, the disease stage as well as the disease phenotype of an individual are important determinants for optimal patient‐tailored disease management. Several staging systems have been proposed to estimate a patient's disease stage and to find patterns in similarly impaired patients. These methods are often univariate and clinical in nature, for example, the King's College (Roche et al. 2012) or MITOS (Chiò et al. 2015) staging systems. While these methods showed promise in clinical trials, (Balendra et al. 2015; Anab et al. 2024) they rely on careful consideration of observed parameters and rely on the interpretation of medical cues (Bede et al. 2022). Furthermore, while the ALSFRS‐R is the most common index of measuring disease severity, it only assesses a subject's motor performance, thus ignoring potential cognitive and/or behavioral impairments. Objective, multivariate, and quantitative indicators of an individual's disease stage have the potential to alleviate these drawbacks and could offer an unbiased way to capture and evaluate the multifaceted disease stage of individual patients.

Similarly, disease phenotypes are typically established through univariate clinical characterization such as site of onset, degree of UMN and LMN involvement, progression rates, or genetic mutation (Dukic et al. 2022). A study that investigated neuroimaging‐based clustering of ALS subjects reported two clusters. Although these clusters did not show any differences in clinical symptom presentation, they did report a predominance of C9orf72 mutations in one group (Bede et al. 2022). Others have shown imaging‐based clusters with distinct brain pathology and differences in cognitive impairment (Tan et al. 2022). Clinically defined cognitively impaired subjects were also reported to show distinct trajectories of cortical thinning across multiple King's College stages, which differ from ALS subjects without cognitive impairment (Consonni et al. 2020). A recent computational study reported staging indices and two distinctive ALS sub‐groups based on TDP‐43 pathology (Alexandra et al. 2023). Despite recent advancements in ALS stratification, a unifying framework for characterizing distinctive disease progression from in vivo multimodal patterns is still lacking.

In this work, we hypothesized that in vivo multimodal neuroimaging encodes patient‐specific disease heterogeneity that can be leveraged to obtain objective, individualized indices of disease progression and disease subtrajectory in ALS. The underlying assumptions are that patterns of neurodegeneration, observable with MRI and underpinning disease heterogeneity, can be used to generate an index of disease progression based on the degree of degeneration, and that disease subtrajectories can be identified by grouping subjects based on similar patterns of degeneration across the disease course. Our specific objectives were (i) to establish a robust and personalized disease progression index based on latent patterns in morphological and microstructural neuroimaging data; (ii) identify neuroimaging‐driven, putative ALS subtypes by grouping subjects with common latent patterns and likely following biologically distinctive disease subtrajectories; (iii) assess the clinical utility of the individual disease progression index and putative subtypes; and (iv) examine the generalizability of this approach by validating the methodology in a secondary large‐scale ALS data set.

2. Materials and Methods

In this study, two large‐scale, multimodal neuroimaging ALS data sets were used: the Canadian ALS Neuroimaging Consortium (CALSNIC) and the Utrecht datasets. CALSNIC is a multicenter consortium across Canada and the USA, prospectively collecting longitudinal neuroimaging, neurological, and clinical characteristics using harmonized imaging protocols across sites. The Utrecht dataset is a single center dataset from the University Medical Center Utrecht in The Netherlands; likewise, it prospectively collects longitudinal neuroimaging, neurological, and clinical characteristics. Here, only cross‐sectional neuroimaging and clinical characteristics were investigated. Data collection was approved by local participating ethics committees, and all subjects gave written informed consent according to the Declaration of Helsinki.

2.1. Inclusion Criteria

All ALS patients fulfilled the revised El Escorial criteria (Brooks et al. 2000) of possible, probable lab‐supported, probable, or definite ALS and had symptoms duration of less than or equal to 5 years.

2.2. Exclusion Criteria

Subjects were excluded if they had any history of neurological trauma. Subjects who did not have structural and diffusion‐weighted data acquired were also excluded. To avoid potential non‐linear effects of age, subjects younger than 35 years of age were excluded. Healthy control subjects who fulfilled the criteria of being cognitively impaired (< 2 standard deviations below the mean) according to their ECAS performance were excluded (McMillan et al. 2022).

2.3. Demographics

A total of 691 (CALSNIC: 115 healthy controls (HC), 119 ALS, Utrecht: 174 HC, 283 ALS) subjects were included with cross‐sectional multimodal MRI as well as neurological and cognitive assessments. Details of the demographics can be found in Table 1.

TABLE 1.

Study sample demographics and clinical characteristics. HC were combined from CALSNIC and Utrecht data sets (115 CALSNIC and 174 Utrecht). Shown are median [range], p values comparing ALS to HC, a indicates only CALSNIC HC were used for comparison. An additional Table S1 has been provided in the supplementary document, listing percent of missing subjects per measure.

	HC	ALS (CALSNIC)	p	ALS (Utrecht)	p
N	289	119	—	283	—
Age (years); median [range]	61 [37 80]	60 [37 77]	0.85	63 [37 81]	0.17
Sex (f/m)	114/175	50/69	0.63	95/188	0.14
Symptom duration (months)	—	19.83 [2.57 59.93]	—	15.34 [1.74 56.12]	—
ALSFRS‐R	—	39 [18 47]	—	39 [18 48]	—
UMN	—	4 [0 12]	—	2 [0 15]	—
Tapping Finger a	58 [34 89]	43.75 [0 83]	< 0.01	—	—
Tapping Foot a	41.25, [18 75]	24.5 [0 63]	< 0.01	—	—
ECAS total	115 [100134]	108 [61127]	< 0.01	111 [55135]	< 0.01
ECAS ALS‐spec	86 [71100]	81 [45 95]	< 0.01	83 [39100]	< 0.01
Disease progression rate		0.45 [0.05 4.68]	—	0.53 [0 6.89]	—

2.4. Neuroimage Acquisition

2.4.1. CALSNIC

Neuroimaging acquisitions were standardized across 3T scanner types/vendors, with minor differences in scanning parameters due to vendor specifications. Details about individual sequences for each vendor can be found in¹. For the sites with Siemens scanners, the parameters were as follows:

High resolution T1 weighted images were acquired at 1 × 1 × 1 mm voxel size.

Diffusion weighted imaging (DWI) was acquired as follows: voxel size = 2 × 2 × 2 mm, diffusion directions = 30, b value = 1000 s/mm², and five b0 images.

2.4.2. Utrecht

High‐resolution T1‐weighted images were acquired using two 3T Philips Achieve Medical Scanners. Acquisition parameters were: three‐dimensional fast field echo using parallel imaging; repetition time/echo time = 10/4.6 ms, flip angle 88, slice orientation sagittal, 0.75 × 0.75 × 0.8 mm voxel size, field of view (FOV) = 160 × 240 × 240 mm, and reconstruction matrix = 200 × 320 × 320 covering the whole brain. From all subjects, two sets of 30 weighted diffusion scans and five unweighted b0 scans each were acquired as follows: DWI‐MR using parallel imaging SENSE p‐reduction 3; high angular gradient set of 30 different weighted directions, TR/TE 57035/68 ms, 2 × 2 × 2 mm voxel size, 75 slices, b = 1000 s/mm², second set with reversed k‐space read‐out.

2.5. Image Processing

T1 weighted images were analysed using cat12 (Gaser et al. 2022), including denoising and bias correction, non‐linear registration to the ICBM152 MNI template, and tissue segmentation and spatial smoothing with a 6 mm Gaussian kernel. Regional GM densities were extracted in 83 cortical, sub‐cortical, and cerebellar regions utilizing areas from the DKT (Desikan et al. 2006) atlas and AAL (Tzourio‐Mazoyer et al. 2002) atlas. DWI data were processed using Tractoflow (Theaud et al. 2020), which uses a combination of FSL (Jenkinson et al. 2012), MRtrix3 (Tournier et al. 2012), ANTs (Avants et al. 2008), and DIPY tools to provide a reproducible pipeline for diffusion weighted images. It includes steps of denoising, eddy current correction, brain extraction, N4 bias correction, normalization, and resampling, prior to calculating the diffusion tensor imaging (DTI) metrics. The JHU labels atlas was used to extract regional fractional anisotropy (FA) values in 48 WM areas.

2.6. Data Harmonization

Due to the multi‐site nature of the CALSNIC project, all neuroimaging data were harmonized using ComBat (Johnson et al. 2007) to reduce the effect of different scanners. ComBat has been shown to effectively reduce the effects of different scanner platforms in various multi‐site neuroimaging studies (Fortin et al. 2018, 2017). During harmonization, the effects of age, sex, and diagnosis were preserved. Similarly, as the Utrecht data was acquired using two separate MRI scanners, the neuroimaging data was also harmonized, retaining the effects of age, sex, and diagnosis.

2.7. Clinical Data Imputation

Missing data in the clinical assessments were imputed using trimmed scores regression with internal principal component analysis (PCA) (Folch‐Fortuny et al. 2016).

2.8. Data Aggregation

To increase the sample size of the HC and to establish a common reference data, the CALSNIC and Utrecht HC data were combined to use as a single reference group; consequently, harmonization was done again to incorporate the different sources of the data. This time, we used a study index (0 for CALSNIC and 1 for Utrecht) to remove possible cohort effects, while retaining age, sex, and diagnosis. Following this, outliers in the regional neuroimaging data, in each data set and diagnosis group, were identified if they were more than five scaled median absolute deviations (MAD) away from the median. Outlier correction was done by imputing the values using trimmed scores regression with internal PCA (Folch‐Fortuny et al. 2016). Table S2 lists the number of subjects with at least five regions as outliers for all diagnosis groups and cohorts.

2.9. Indexing Disease Progression and Subtrajectories

We used cTI (Iturria‐Medina, Khan, et al. 2020; Iturria‐Medina, Carbonell, et al. 2021) to do the joint modeling of disease progression and subtrajectories, based on population‐level neuroimaging data. The two main outputs of cTI are (i) an individualized continuous marker of disease progression (hereon, ALS imaging‐based progression score [ALS‐IPS]), and (ii) a sub‐group membership for each participant following a distinct ALS subtrajectory. As a first step, the high‐dimensional, multi‐modal neuroimaging data is reduced by identifying the low‐dimensional pattern enriched in subjects diagnosed with ALS that is relative to normative healthy controls. This is achieved by applying contrastive PCA (cPCA) (Abid et al. 2018) to the whole data set, reducing it to a low dimensional space that is free from naturally occurring variations in a healthy population and thus capturing ALS‐associated patterns. Then, the personalized disease progression score is calculated using a Minimum Spanning Tree as the distance from each individual patient to the normative controls. This distance is then normalized to 0 and 1 to indicate the relative disease progression of all subjects within the investigated population. In other words, the more pathological alterations found in a subject's neuroimaging data, and thus likely to be more severely affected, the greater the distance from that subject to healthy controls, and thus a higher ALS‐IPS. Lastly, the subjects are assigned a group membership index that is based on an Expectation Maximization algorithm (Iturria‐Medina et al. 2022) that maximizes the alignment of the subjects within each specific subtrajectory based on their neuroimaging data. Finally, a statistical subtypes stability and significance analysis is performed via randomized permutations. Each disease subtype's stability is calculated as the rate at which pairs of subjects group together into the same cluster upon repeated clustering on random subsets of the input data. The comparison of each subtype's intrinsic stability with a generated null distribution allows testing its significance.

Here the input for cTI was the multi‐modal neuroimaging features (GM density of 83 ROIs and FA of 48 WM ROIs) and all healthy controls of both datasets were selected as the normative background for the cPCA. Additionally, we selected the most severely affected patients as a “target” group for the cPCA, which will help to identify and isolate ALS specific alterations more robustly, since in a neuroimaging study patients with a low disease severity are typically overrepresented and thus may obscure robust patterns of aberrant brain changes. How the “target” population was defined is explained in section “Clinical Disease composite score.” cTI is freely available to the research community as part of the NeuroPM‐box (Iturria‐Medina, Carbonell, et al. 2021), available here: https://www.neuropm‐lab.com/neuropm‐box‐download.html.

Prior to submitting the regional multi‐modal imaging values to cTI, the effects of age and sex were regressed out using linear regression. The regression was trained on the control cohort to obtain estimates of healthy aging and subsequently applied to the ALS group. An overview of the analysis steps is depicted in Figure 1.

FIGURE 1.

Analysis pipeline for each of the two ALS data sets. Multimodal neuroimaging features were regionally extracted and adjusted for potential scanner differences in each ALS data set. In order to compare both sets, the data were harmonized across studies, followed by outlier correction, and adjustment for age and sex. The combined CALSNIC and Utrecht HC were used as the reference HC sample, while the CALSNIC and Utrecht ALS samples were used separately in cTI to identify ALS subtrajectories and individualized disease progression scores (ALS‐IPS). These were then evaluated for clinical relevancy and finally assessed for similarity across data sets.

2.10. Clinical Disease Composite Score

We generated a clinical disease composite score (DCS) for each ALS patient where a higher score indicates a higher degree of impairments. For this, we employed a PCA on clinical assessments such as ALSFRS‐R, ECAS ALS‐specific score, and UMN burden score, and inverted the scores for ALSFRS‐R and ECAS so that higher scores indicate worse performance. The final score was generated as the average of all components weighted by their associated explained variance. This was done to obtain a score comprised of multiple clinical assessments, evaluating different aspects of the disease (motor as well as cognitive performance) to reflect the multi‐system nature of the disease, rather than defining the severity based on a single assessment of a specific symptom domain. This score also served to identify the most advanced subjects where the “target” population for the cTI was chosen as the subjects falling within the top 30% of the disease composite score. The percentage was chosen as it strikes a good balance of sample size and disease severity for the target population. Figures S1 and S2 show the result from cTI associating the ALS‐IPS to the DCS and the King's stages for both cohorts using different target sizes, where a similar pattern to what is presented in the main text can be observed.

2.11. Statistical Analysis

To test the clinical utility of the ALS‐IPS, we correlated it with the disease composite score as well as individual clinical measures of disease severity in all ALS subjects as well as within each subtrajectory. Additionally, we applied the Jonckheere‐Terpstra test to examine statistical trends of ALS‐IPS across clinically established disease stages according to the King's College staging system (Balendra et al. 2014), again in all ALS subjects and within each subtrajectory.

Subtrajectories were examined for differences in cortical atrophy and WM integrity by performing groupwise comparisons between subtrajectories and HC in each region.

We performed group‐wise comparisons of clinical assessments using ANOVAs with age, sex, and symptom duration as covariates and report the p values of the post hoc Tukey tests for all continuous variables corrected for multiple comparisons. Symptom duration was accounted for in the model to ensure that potential differences between subtrajectories were not due to early and late stages of the disease. Comparisons of proportions were performed using Fisher's exact test. Statistical significance levels were set to p < 0.05.

3. Results

We will first discuss the results obtained from the CALSNIC cohort, followed by a summary of the Utrecht ALS cohort, and lastly examine the similarity of results between cohorts. Demographics and clinical characteristics are listed in Table 1, with Table S1 containing the percentages of missing data per variable.

3.1. cTI Outputs

cTI identified three ALS subtrajectories, ALS_motor, ALS_diffuse, and ALS_limbic, in the CALSNIC data set. The names of the subtrajectories were ascribed according to the most prominent regions that show differences to HC, and the spatial extent is presented in section Neuroimaging based ALS subtrajectories differ in radiological and clinical pathological expression. Figure 2A shows the number of subjects in each of the subtrajectories as well as the total number of HC used. The distribution of ALS‐IPS is shown in Figure 2B, where all three sub‐groups have comparable disease stages. The 10 most impactful imaging features characterizing the disease space, per modality, are displayed in Figure 2D. The most relevant GM features were the left entorhinal and right middle temporal cortex, whereas the right superior corona radiata and the splenium of the corpus callosum were the most contributing features of the WM. The relative contribution of representing the disease space of the top 10 features of each modality is shown in Figure 2C, which shows an even split of importance between GM and WM features. The feature contributions were derived from the cTI weights summarized over all components.

FIGURE 2.

cTI output summary. (A) number of subjects in each ALS subtrajectory as well as the total number of HC used as reference. (B) distribution of ALS‐IPS of subtrajectories (C) relative contribution of the top 10 features of GM and WM to the characterisation of the ALS disease space (D) ranking the top 10 GM and WM features according to their summarized contribution.

3.2. Neuroimaging‐Based Index of Disease Progression Correlates With Clinically Observed Impairments

To assess whether our neuroimaging‐derived index for disease progression, ALS‐IPS, has clinical utility, we compared it to clinical symptom severity as well as to the King's College staging system. As an initial indicator of usefulness, we correlated the ALS‐IPS with the disease composite score in all ALS subjects and observed a significantly positive relation (R = 0.56, p < 0.01) suggesting that the ALS‐IPS can be used as an in vivo proxy for overall symptom severity, as shown in Figure 3A. Comparing the ALS‐IPS with clinically derived stages based on the King's College system in all ALS subjects reveals a significant trend (p = 0.001) of higher ALS‐IPS in progressive King's College stages (Figure 3B).

FIGURE 3.

Correlations of ALS‐IPS and clinical evaluations in the ALS cohort and subtrajectories. (A) significant correlation of ALS‐IPS and DCS in all ALS subjects (B) positive trend of higher ALS‐IPS in later stages of the King's College staging system (C) ALS‐IPS significantly correlates with both motor and cognitive tests, as well as the disease composite score, DCS, in ALS_motor and ALS_limbic. In ALS_diffuse only the cognitive performance scores correlate with ALS‐IPS. (D) positive significant trends of higher ALS‐IPS in later King's College stages in ALS_motor and ALS_limbic. Significance level: P < 0.05.

The ALS‐IPS correlates with clinical motor and cognitive performance differentially across subtrajectories, as shown in Figure 3C. The ALS‐IPS correlates with both motor (ALSFRS‐R and UMN burden score) and cognitive (ECAS total and ECAS ALS‐specific) performance in ALS_motor and ALS_limbic. In ALS_diffuse on the other hand, ALS‐IPS only correlates with cognitive performance. Additionally, ALS‐IPS also significantly correlates with the disease composite score, DCS, in ALS_motor and ALS_limbic and not in ALS_diffuse. Additionally, a positive significant trend between ALS‐IPS and King's College stages can be observed in subtrajectories ALS_motor and ALS_limbic, as seen in Figure 3D.

3.3. Neuroimaging Based ALS Subtrajectories Differ in Radiological and Clinical Pathological Expression

To test whether the identified disease subtrajectories would show spatially distinct alterations of GM and WM microstructure, we performed pairwise regional comparisons between each of the ALS subtrajectories and HC on the demographic adjusted imaging features. Figure 4A displays the regional t‐stats for each GM and WM region in each subtrajectory, where ALS_motor shows the most change in regions encompassing the sensorimotor system, namely the right postcentral and left precentral gyri, as well as in the bilateral corticospinal tract. ALS_diffuse exhibits widespread decreases in GM‐density with the most change in the bilateral thalamus, as well as reduced FA in the bilateral superior corona radiata. The areas with the most change, compared to controls, in ALS_limbic are regions comprising the limbic system, with decreased GM‐density in the left amygdala and hippocampus, as well as a reduction in WM‐FA in the right fornix and body of the corpus callosum.

FIGURE 4.

Imaging and clinical differences in ALS subtrajectories. (A) ALS_motor shows largest differences in the sensorimotor system right postcentral, and left precentral gyri, together with bilateral corticospinal tract contributions. ALS_diffuse exhibits large decreases of GM‐density in the bilateral thalamus as well as reduced FA in the bilateral superior corona radiata. ALS_limbic shows largest alterations in regions constituting to the limbic system, decreased GM‐density in the left amygdala and hippocampus, as well as reductions of WM‐FA in the right fornix and body of corpus callosum. Shown are the t‐stats per ROI comparing each subtrajectory to HC where p < 0.05. (B) group‐wise comparisons of demographic, disease aggressiveness, and overall disease severity, the shading for male and female is indicated by darker and lighter colors (C) clinical comparisons between ALS subtrajectories significance level: P < 0.05.

Additionally, we probed if the neuroimaging‐derived subtrajectories would also differ in demographic or clinical symptom expression (Figure 4B,C). No differences in subject demographics, disease aggressiveness as measured by disease progression rate (DPR), or overall symptom severity were found. However, both ALS_motor and ALS_limbic displayed impaired motor performance when compared to ALS_diffuse as evidenced by reduced finger tapping scores. No differences in cognitive performance could be observed.

3.4. Verifying Neuroimaging‐Derived Disease Progression Indexing and Subtrajectories in a Second ALS Cohort—Utrecht Data Set

As in the CALSNIC cohort, cTI identified three ALS subtrajectories in the Utrecht data as well. Figures 5 and 6 summarize the progression indexing and subtrajectory results using the Utrecht ALS subjects. Figure 5A,B, show the number of subjects as well as the ALS‐IPS distribution per subtrajectory. The most contributing modality is GM‐density with a split of 62%–38% of WM‐FA (Figure 5C), with the most impactful regions describing the ALS disease space ranked in Figure 5D, which shows deviations from the CALSNIC cohort. Overlapping regions of the top 10 contributing WM areas are the superior corona radiata, body, and splenium of the corpus callosum. Overlapping GM areas are right entorhinal, right middle temporal, accumbens, superior frontal, insula, and precentral gyri.

FIGURE 5.

Summary of ALS disease indexing and subtrajectory results in the Utrecht ALS cohort. (A) number of subjects in each ALS subtrajectory and total HC used (B) distribution of ALS‐IPS in subtrajectories and the HC population (C) and (D) relative and top 10 most contribution brain regions per modality in describing the ALS disease space. Here, the split is 62%–38% as opposed to an even split in the CALSNIC data set. (E–H) significant correlations of ALS‐IPS and clinical assessments and King's College stages, all associations in all ALS subtrajectories were significant.

FIGURE 6.

Summary of ALS disease indexing and subtrajectory results in the Utrecht ALS cohort. (A) regional t‐stats of comparisons of ALS subtrajectories and HC per modality (B) pairwise comparisons of demographic variables between all subtrajectories. ALS_motor was generally younger and had the lowest ALS‐IPS. No differences were found in symptom duration and disease progression rate. Disease composite score shows an increase in ALS_motor compared to the other subtrajectories. (C) pairwise comparisons of clinical evaluations where differences in motor performance between the ALS subtrajectories can be observed (D) Comparison of common clinical dichotomizations, ALS_motor has significantly lower ratio of cognitively impaired subjects.

Similar to the CALSNIC data, a correlation of ALS‐IPS and the disease composite score is significant in all subjects, as seen in Figure 5E, and a positive trend of higher ALS‐IPS in advanced King's College stages can be observed as well (Figure 5F). Moreover, the ALS‐IPS was significantly associated with individual motor and cognitive symptom severity as well as with DCS in all subtrajectories, as seen in Figure 5G, with a positive significant trend of ALS‐IPS in King's College stages as well (Figure 5H).

The spatial extent of brain pathology in subtrajectories per modality is displayed in Figure 6A. ALS_motor shows the largest change in areas that are part of the motor system, with decreases in GM‐density in bilateral post‐ and precentral areas, together with a large decrease in WM‐FA in bilateral cerebellar peduncles and superior corona radiata. ALS_diffuse shows widespread changes throughout the brain, including changes in GM‐density in the left hippocampus and left inferior parietal areas and left hippocampus. The largest differences compared to HC in WM‐FA are in bilateral fornix and cerebellar peduncles. Lastly, ALS_limbic exhibits regional changes in areas comprising the limbic system with large decreases of GM‐density in bilateral amygdala, pallidum, and hippocampus areas, as well as decreases of WM‐FA in the right fornix, bilateral corticospinal tract, and cerebellar peduncles. Interestingly, ALS_limbic also displays slight increases of GM‐density in the occipital lobe, specifically in bilateral cuneus and pericalcarine areas.

Pairwise comparisons of demographic variables between all subtrajectories show differences. ALS_motor was generally younger than ALS_limbic, while ALS_diffuse had the lowest ALS‐IPS. No differences were found in symptom duration and DPR. Disease composite score shows an increase in ALS_motor compared to ALS_diffuse and ALS_limbic (Figure 6B). The ALS subtrajectories also exhibit differences in clinical symptom presentation, as shown in Figure 6C. ALS_motor has significantly lower ALSFRS‐R compared to ALS_diffuse and is trending towards the highest UMN burden, indicating worse motor performance compared to the other groups. No differences in cognitive performance are observed. ALSmotor has a significantly lower ratio of cognitively impaired subjects than ALS_diffuse, while ALS_limbic had the highest ratio of bulbar to spinal onset, but no differences in DPR can be observed, Figure 6D.

3.5. Subtrajectories Similarity Across Studies

The analysis identified three ALS subtrajectories in CALSNIC data and three subtrajectories in the Utrecht data that show spatially distinct alterations between subtrajectories but are largely overlapping between two independent cohorts. Based on the pairwise correlations of regional t‐stats (from comparing each neuroimaging modality to HC) across the two studies, we observe good correspondence between the two data sets. In both patient cohorts we were able to extract subtrajectories that (i) exhibit large changes in the motor system, (ii) show widespread changes with connections through the thalamus, and (iii) display changes in regions associated with the limbic system. This correspondence pattern is visualized in Figure 7, where the correlation coefficients used for matching the subtrajectories across studies are shown for the GM and WM features separately. Using a weighted DICE coefficient to assess the similarity yields comparable results and can be seen in Figure S4.

FIGURE 7.

Pairwise similarities of identified CALSNIC and Utrecht ALS subtrajectories. Left panel shows similarities of GM and right panel displays the similarities of WM changes.

4. Discussion

Our analysis is a step forward in the joint characterization of ALS subjects' according to their disease progression and putative disease subtrajectories, and across independent cohorts. We have shown that we are able to obtain a robust and personalized in vivo disease progression index that is predictive of current clinical impairment utilizing multimodal neuroimaging data. Additionally, the disease progression index was used to inform the neuroimaging‐based clustering to identify and group subjects with common patterns of progressive pathologies. Thus, leading to the extraction of disease subtrajectories which sets the analysis apart from conventional clustering to reveal disease sub‐groups (Bede et al. 2022; Tan et al. 2022). While another study attempted a joint characterization of ALS and ALS‐FTD subjects using neuroimaging data (Shen et al. 2023), our analysis goes beyond that by leveraging the complementary information of multiple imaging modalities to arrive at a patient‐centered characterization of disease progression and disease subtrajectories. Our identified disease subtrajectories show distinct patterns of GM and WM pathological alterations that translate into different clinical symptom presentations. The discussion that immediately follows applies to the CALSNIC dataset, followed by that for the independent Utrecht dataset.

4.1. Neuroimaging Data Encodes Pathophysiological Changes That Is Helpful in Describing ALS Disease Progression

We make use of cPCA, which when applied to multimodal neuroimaging data as part of the cTI methodology, allows for the identification of ALS‐specific patterns, free from the effects of normal aging or other physiological changes in the control group. The cPCA output thus encapsulates the ALS disease space in a low dimension that is subsequently used to establish subject‐specific indices of disease progression as well as to identify subgroup‐specific disease trajectories.

Pathological changes in both GM and WM were detected using the cTI methodology. The most contributing WM and GM features, in describing the ALS disease space, are the bilateral superior corona radiata areas and the splenium of the corpus callosum, which were all found to show significant reductions in FA in a large‐scale meta‐analysis comprising nearly 396 ALS patients and 360 HC (Zhang et al. 2018). Furthermore, the left entorhinal and right middle temporal GM areas have been shown to be important discriminators in ALS from HC, as well as distinguishing between UMN and LMN predominant ALS patients in a cortical thickness study (Ferrea et al. 2021). Structural and functional network studies reported altered connectivities of the orbitofrontal node to the basal ganglia, which were in turn associated with lower cognitive performance (Masuda et al. 2016; Bharti et al. 2022). Overall, the areas in our results that underlined general pathological changes in ALS were partially consistent with previous neuroimaging studies.

4.2. Objective and Multi‐Modal Neuroimaging‐Based Disease Progression Index Corresponds to Clinical Disability

Common clinical staging systems for ALS, such as the King's College system (Roche et al. 2012) and the D50 model (Steinbach, Batyrbekova, et al. 2020), rely on interpretations of medical cues and lack granularity, objectivity, incorporation of extra‐motor features, and/or require longitudinal assessments. Neuroimaging studies typically use clinical stages to stratify subjects, but there is a gap in neuroimaging‐based studies defining ALS disease stages. To address these limitations, we aim to characterize ALS progression using a data‐driven approach that infers pseudotemporal information from population‐level ALS data. This analysis leverages high‐dimensional multi‐modal neuroimaging data to identify latent patterns of ALS pathology, distinguishing them from natural physiological changes. These patterns allow for ranking ALS subjects based on individual disease progression relative to a normative background. Previous applications of cTI have displayed its ability to robustly generate a subject‐specific disease progression score based on molecular (Iturria‐Medina, Khan, et al. 2020; Iturria‐Medina, Carbonell, et al. 2021; Iturria‐Medina et al. 2022) and neuroimaging data (McCarthy et al. 2022) that is reflective of clinical disease severity.

Our neuroimaging‐derived index for ALS progression, ALS‐IPS, displayed a positive correlation with the disease composite score, indicating that it is reflective of overall disease severity. Additionally, we note that subjects in later stages of the disease, as categorized by the King's College stage, typically also exhibit a higher ALS‐IPS, suggesting some degree of overlap between the in vivo obtained ALS‐IPS and King's College stages. However, as the ALS‐IPS is based on multivariate and multimodal neuroimaging data, it is likely to be influenced by additional non‐motor/function pathological effects and could provide a more comprehensive score of overall disease progression compared to the mostly functional assessment with ALSFRS‐R. Considering that ALSFRS‐R or King's College stages focus on motor/function aspects to ascribe disease progression or stage, the ALS‐IPS summarizes the disease in a multiparametric way reflecting the multisystem nature of ALS. Taken together, this suggests that not only ALS‐specific pathology but also the degree of clinical severity can be recovered, at an individual level, by carefully examining pseudotemporal patterns at the population level using cross‐sectional neuroimaging data, and thus ALS‐IPS can be used as an in vivo proxy for disease severity. In vivo generated disease progression scores could provide pivotal tools in clinical research and development of treatments as they have the potential to track disease progression over time or stratify patients according to the pathological processes observed in neuroimaging.

4.3. ALS Subject Stratification Based on Neuroimaging Disease Trajectories

Objective patient stratification offers significant benefits in healthcare by enabling more personalized and effective treatment. It allows for tailored interventions based on individual patient characteristics such as disease progression or disease subtrajectory and can aid in early disease detection, better prognostics, and more efficient resource allocation for clinical trials. ALS subjects are often stratified based on univariate clinical observations such as site of onset, progression rate, or cognitive/behavioral symptoms; however, these clinical phenotypes may only encompass limited information of this complex disease. Related neuroimaging studies stratify the ALS subjects on clinical characteristics first and then aim to uncover associated radiological signatures (Senda et al. 2017; Cardenas‐Blanco et al. 2014; Mazón et al. 2018). Subject stratification based on high‐dimensional multi‐modal neuroimaging data, where different pathological aspects are embedded in the radiological profiles of each individual, could segregate distinct ALS sub‐groups without relying on clinical a priori knowledge. Studies investigating neuroimaging‐based sub‐grouping in ALS have found mixed results. A previous study that used a combination of cortical thickness, sub‐cortical volumetric, and WM integrity indices found two ALS clusters that did not differ in demographics or symptom presentation but in their genetic profiles (Bede et al. 2022). Another study revealed three distinct clusters utilizing a combination of cortical thickness and structural connectivity indices from DTI. The clusters differed in several clinical evaluations, including ALSFRS‐R and site of onset (Tan et al. 2022).

Applying cTI to multi‐modal neuroimaging data revealed three ALS‐related subtrajectories. Predominant GM‐density decreases in the right postcentral and left precentral gyri, together with decreases in WM‐FA in the bilateral corticospinal tract and body of the corpus callosum, characterize ALS_motor, which encapsulates previously described alterations of the GM and WM motor system (Tan et al. 2022; Hsueh et al. 2023). ALS_diffuse exhibited widespread changes with predominately GM‐density decreases in the bilateral thalamus, combined with WM‐FA decreases in the bilateral superior corona radiata. The thalamus acts as a relay to connect subcortical and cortical structures across the brain, and multiple studies have reported altered structural as well as functional connectivity in ALS (Castelnovo et al. 2023). Additionally, reductions in size and microstructural integrity of the thalamus have been reported with clinical relevancy to identify distinct modulation patterns along the ALS‐FTD spectrum (Ahmed et al. 2021; Chipika et al. 2020); however, we could not find any differences in cognitive performance comparing ALS_diffuse with the other subtrajectories. ALS_limbic is characterized by changes in areas constituting the limbic system. Decreases in GM‐density in the left hippocampus and left amygdala, as well as reductions in WM‐FA in the right fornix and body of the corpus callosum, can be observed. The limbic system has been implicated in ALS pathology before (Shen et al. 2023; Passamonti et al. 2013; Westeneng et al. 2015; Christidi et al. 2024). Atrophy in the amygdala and hippocampus has previously been noted as an important structure to differentiate ALS‐FTD, behavioral FTD, and pure ALS when compared to controls (Ahmed et al. 2021) pointing towards distinct atrophy profiles along the ALS and FTD spectrum. Moreover, increases in mean diffusivity were observed in the thalamus, hippocampus, and amygdala in ALS when compared to controls, where the change in microstructural properties was also associated with the degree of extra‐motor disability (Barbagallo et al. 2014), highlighting the possibility of distinct neuroanatomical changes along the spectrum of ALS and FTD. The limbic system was also implicated with overall disease severity and could lead to anxiety‐induced motor impairment (Passamonti et al. 2013). The body of the corpus callosum has been shown to have altered microstructural integrity in ALS when compared to controls (Steinbach, Gaur, et al. 2020; Menke et al. 2018), whereas the fornix has been implicated to be associated with memory dysfunction in ALS (Sarro et al. 2011). Of note here is that both the predominately motor and limbic‐associated ALS subtrajectories display some level of motor impairment when compared to subtrajectory ALS_diffuse, but we could not observe significantly altered levels of cognitive performance between the three subtrajectories. Of note, a comparison of UMN/LMN predominance across subtrajectories (Figure S3) revealed that ALS_diffuse encapsulates the least LMN predominant subjects. Previous literature investigating UMN/LMN predominance found that when looking at a priori dichotomized subject groups, the UMN predominant group exhibits cortical atrophy beyond the motor system to frontal, temporal, parietal, and occipital lobes (Ferrea et al. 2021), all areas that were also implicated in our ALS_diffuse subtrajectory.

Overall, despite methodological differences, our findings partly overlap with previous reports on ALS subject stratification. A large‐scale study by Tan et al., of which the Utrecht data used here is a subset and a focus on static sub‐grouping of ALS patients, reported three sub‐groups with distinct neuropathological features (Tan et al. 2022) that encompass a pure motor sub‐type, a fronto‐temporal sub‐type, and a cingulate‐parietal–temporal sub‐type, where we can support the finding of the pure motor sub‐type in both the CALSNIC and the Utrecht data. The fronto‐temporal sub‐type overlaps with subtrajectories ALS_diffuse in CALSNIC and Utrecht data. Another recent study utilizing ALS and ALS‐FTD subjects, with a focus on identifying sub‐groups with distinct progression patterns, reported two sub‐groups, a prefrontal‐somatomotor and limbic predominant sub‐type (Shen et al. 2023). While they identified a sub‐type comprising motor and frontal areas together, it might be due to the combined ALS and ALS‐FTD patients in their analysis. Nevertheless, a motor or motor‐inclusive sub‐type appears to be a robust finding across different ALS data sets, investigated neuroimaging features, and analysis methods. Of note should be that they also mention that a large portion of patients could not be reliably assigned to either of their two sub‐types, highlighting the large heterogeneity in ALS and along the ALS FTD disease spectrum.

It is important to recognize that there is no requirement or expectation for a one‐to‐one mapping between differing neuroimaging metrics of our identified subtrajectories and clinical disability. In fact, similar levels of clinical impairment can arise from distinct patterns of neurobiological disruption, meaning that patients with comparable symptom profiles may exhibit markedly different alterations in underlying biological data, as shown by others (Khan et al. 2022; Corriveau‐Lecavalier et al. 2023). This heterogeneity reflects the complexity of underlying disease mechanisms, where divergent biological pathways can converge on a shared clinical phenotype. As a result, therapeutic needs may differ substantially among patients who appear clinically similar, as results from many clinical trials have shown for the neurodegenerative spectrum. This disconnect underscores a major limitation of relying solely on clinical symptoms to guide treatment selection, a strategy that has repeatedly failed to generalize across heterogeneous patient populations in neurodegenerative diseases such as ALS, Alzheimer's disease, and Parkinson's disease. More nuanced, biologically informed approaches are essential to improve precision in diagnosis, stratification, and treatment.

4.4. ALS Sub‐Grouping and Disease Progression Indexing Can Largely Be Replicated in a Secondary ALS Cohort

Applying the cTI methodology to a second ALS cohort (Utrecht dataset), we again identified three sub‐groups that share similarities in their respective neuroimaging pathological changes with the CALSNIC cohort and are summarized in Figure 7. The distribution of subjects per subtrajectory is comparable, while the subtrajectories show differences in ALS‐IPS. Notably in the Utrecht data, the split of the top 10 most contributing features per modality in describing the ALS disease space is now shifted to favor GM features. The top contributing GM features are the right middle temporal and left entorhinal areas, both of which were also the most contributing GM areas in the CALSNIC data. The most contributing WM areas are the left cerebral peduncle and the left superior corona radiata, which are different from the ones observed in CALSNIC (albeit the right superior corona radiata instead of the left one and the left superior corona radiate is the third most contributing feature) is observed there. Overall, we observe 3/10 overlapping WM areas as the main contributing factors between CALSNIC and Utrecht data and 6/10 overlapping GM areas (three with exact overlap and three where the opposite hemisphere is observed between studies). The cerebral peduncle is an important structure facilitating the corticospinal as well as the corticobulbar tract and thus is an integral part of the motor system. Altered microstructural integrity in the cerebral peduncles were associated with a faster disease progression (Spinelli et al. 2020). A meta‐analysis reported consistent decreases in FA in ALS in the right cerebral peduncle, left corona radiata, together with the body and splenium of the corpus callosum (Zhang et al. 2018), highlighting the importance of these areas in characterizing the ALS disease space. Atrophy in the temporal gyrus has been previously observed (Mezzapesa et al. 2007) and was reported to be linked to cognitive decline in ALS (Jellinger 2023).

As in the CALSNIC cohort, the ALS‐IPS is a good proxy for overall disease severity in all subjects as it correlates with the disease composite score and King's College stages in all subjects. Additionally, it also correlates with the individual motor and cognitive assessments across all subtrajectories and also shows positive trends with the King's College stages.

The ALS‐IPS correlates with all evaluated clinical domains of motor and cognitive performance, as well as the disease composite score in all three identified ALS subtrajectories. Taken together, this highlights again that individual pathological patterns are encoded in structural and microstructural MRI data that can be recovered and leveraged to provide an objective index of disease progression on a subject specific level.

The subtrajectories display distinct patterns of pathological alterations that encompass the motor system, the limbic system, and widespread cortical and subcortical areas, similarly to the CALSNIC cohort. ALS_motor appeared to have the worst motor performance assessment out of all three subtrajectories, as it displayed lower ALSFRS‐R scores than both other subtrajectories and showed a trend towards the highest UMN scores among the three sub‐groups. The cognitive performance between the three identified sub‐groups was comparable; on the other hand, it is important to note that the differences in motor performance were not due to demographic or symptom duration discrepancies, indicating an ALS subtrajectory predominately affected by motor symptoms. Indeed, the largest degree of change in structural and microstructural integrity is found within the GM and WM motor system. No cognitive differences between any of the subtrajectories could be found when analyzing ECAS total and ECAS ALS‐specific (as well as ECAS sub‐domains of language, executive, verbal fluency, memory, and visuospatial, data not shown); however, it is interesting to note that ALS_motor includes the least, relatively, subjects with cognitive impairment. The lack of direct quantitative differences between the subtrajectories could stem from the general insensitivity of the ECAS when trying to identify subtle differences between patient cohorts, but a dichotomization may provide a larger picture. Overall, there is substantial overlap in radiological subtrajectories among the two investigated ALS cohorts, CALSNIC and Utrecht.

5. Limitations

This study is not without limitations. Given the nature of neuroimaging studies, the distribution of disease severity is skewed towards earlier/mildly affected subjects, which could in turn influence the staging and sub‐typing methodologies, even though they have been internally cross‐validated. The ability to investigate large sample sizes is crucial for revealing robust disease‐related patterns and is a major benefit of multi‐center studies; however, there are also drawbacks associated with such a study design, and that is the aggregation of data stemming from different scanner types and manufacturers. Although we specifically accounted for potential scanner differences by harmonizing within and across studies, there may still be residual variability in the data that is related to these differences and could be a reason why the CALSNIC subtypes appear to be less clear and distinct from each other than in the Utrecht data, due to the higher number of aggregated scanner models and manufacturers. Frequently reported behavioral impairments were not considered in this work due to the misalignment of assessed tests across the studies and cohorts. This was an intentional decision not to confound the generated disease composite score with potentially non‐overlapping tests so that the subsequent analyses were more directly comparable across cohorts. Future studies should include a carefully selected or harmonized assessment of behavioral changes. Additionally, we found that genetic data of subjects was largely unavailable to either be included in cTI or for genotypic comparisons across subtrajectories to draw robust conclusions. Genetic mutations have been shown to correspond to distinct symptom manifestations and pathological alterations in the past and will need to be included in a future study for a more integrative approach of identifying ALS subtrajectories.

6. Conclusion

To conclude, we demonstrate the feasibility of extracting personalized disease progression markers as well as distinct disease subtrajectories from in vivo multimodal neuroimaging data in a heterogeneous ALS population. This individualized disease progression index can be obtained from cross‐sectional data alone by utilizing advanced data‐driven and unsupervised methods on large‐scale data sets. This highlights that individual pathological patterns are encoded in structural and microstructural MRI data that can be recovered and leveraged to provide an objective index of disease progression on a subject‐specific level. Obtaining an individualized in vivo disease progression index can have tremendous potential in objectively assessing one's disease progression as well as monitoring longitudinal changes. It has the potential to improve patient stratification of clinical trials aiming to study early disease characteristics by utilizing the objective marker of disease progression rather than subjective and noisy clinically observed medical cues. Additionally, a clinical trial could potentially directly monitor the progression of an in vivo marker of interest and how the intervention may change it, instead of evaluating indirect, secondary outcome measures. Moreover, we provided evidence of reproducible ALS subtrajectories using complementary multimodal neuroimaging modalities characterizing GM and WM degeneration across disease stages. Our approach could be a promising tool for characterizing the spatiotemporal pathology of different tissue types in ALS, while providing crucial insights of quantitative and individualized indices of disease progression and phenotype, moving closer to precision medicine.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: hbm70341‐sup‐0001‐Supinfo.docx.

Baumeister, T. R. , Westeneng H.‐J., van den Berg L., et al. 2025. “Multimodal Neuroimaging‐Guided Stratification in Amyotrophic Lateral Sclerosis Reveals Three Disease Subtypes: A Multi‐Cohort Analysis.” Human Brain Mapping 46, no. 14: e70341. 10.1002/hbm.70341.

Data Availability Statement

The data used can be made available to researchers upon reasonable request. The cTI methodology is part of the NeuroPM box (Iturria‐Medina, Carbonell, et al. 2021), available at: https://www.neuropm‐lab.com/neuropm‐box‐download.html.