Multi-scale machine learning model predicts muscle and functional disease progression

Abstract

Facioscapulohumeral muscular dystrophy (FSHD) is a genetic neuromuscular disorder characterized by progressive muscle degeneration with substantial variability in severity and progression patterns. FSHD is a highly heterogeneous disease; however, current clinical metrics used for tracking disease progression lack sensitivity for personalized assessment, which greatly limits the design and execution of clinical trials. This study introduces a multi-scale machine learning framework leveraging whole-body magnetic resonance imaging (MRI) and clinical data to predict regional, muscle, joint, and functional progression in FSHD. The goal this work is to create a ‘digital twin’ of individual FSHD patients that can be leveraged in clinical trials. Using a combined dataset of over 100 patients from seven studies, MRI-derived metrics—including fat fraction, lean muscle volume, and fat spatial heterogeneity at baseline—were integrated with clinical and functional measures. A three-stage random forest model was developed to predict annualized changes in muscle composition and a functional outcome (timed up-and-go (TUG)). All model stages revealed strong predictive performance in separate holdout datasets. After training, the models predicted fat fraction change with a root mean square error (RMSE) of 2.16% and lean volume change with a RMSE of 8.1 ml in a holdout testing dataset. Feature analysis revealed that metrics of fat heterogeneity within muscle predicts muscle-level progression. The stage 3 model, which combined functional muscle groups, predicted change in TUG with a RMSE of 0.6 s in the holdout testing dataset. This study demonstrates the machine learning models incorporating individual muscle and performance data can effectively predict MRI disease progression and functional performance of complex tasks, addressing the heterogeneity and nonlinearity inherent in FSHD. Further studies incorporating larger longitudinal cohorts, as well as comprehensive clinical and functional measures, will allow for expanding and refining this model. As many neuromuscular diseases are characterized by variability and heterogeneity similar to FSHD, such approaches have broad applicability.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-09516-8.

Introduction

Facioscapulohumeral muscular dystrophy (FSHD) is a genetic neuromuscular disorder characterized by progressive muscle weakness and wasting, affecting approximately 1 in 7,500 individuals¹. The genetic basis for FSHD—mutations leading to aberrant expression of the protein DUX4, which is toxic to muscles—is well established. Disease progression typically occurs over years in adult-onset FSHD and can affect skeletal muscles across the entire body^2–10. However, there are no approved treatments for FSHD, and no effective approaches exist to mitigate the progressive weakness and associated loss of key functions such as ambulation, reaching, and facial expression.

Current clinical metrics for tracking FSHD progression primarily rely on functional assessments, including manual muscle testing, timed movement tests (e.g., timed up-and-go (TUG)), computer-vision-analyzed movements (e.g., reachable workspace (RWS)¹¹), and patient-reported outcomes. While these measures can provide valuable insights into function over longer periods, they have significant limitations in shorter-term settings, such as clinical trials lasting for shorter periods of time, such as a year or less. Functional measures often exhibit inherent variability from test to test and day to day, reducing their reliability for detecting meaningful changes over short durations. Furthermore, the heterogeneity of FSHD means that the most relevant functional task for tracking disease progression likely differs across patients. For instance, only a subset of participants in a clinical trial may experience measurable declines in shoulder girdle strength, impacting their RWS performance, particularly in the absence of treatment. The limitations of such measures were recently highlighted in the Fulcrum Therapeutics Phase III trial (NCT05397470), where both placebo and treatment groups showed improvements in RWS, ultimately preventing the detection of meaningful treatment effects. The reasons for this placebo effect remain unclear but could involve motor learning effects or placebo responses. Similar limitations extend to other functional metrics, such as TUG and the 6-minute walk test, which may not capture progression uniformly across all patients.

To address the limitations of global clinical metrics, magnetic resonance imaging (MRI) has emerged as the gold standard for quantifying muscle-level involvement in FSHD. MRI enables a detailed assessment of muscle composition, particularly the amount of fat infiltration, which serves as a reliable metric of disease expression. Despite the relative ease of acquiring whole-body imaging data, current analytic methods for segmenting and quantifying features at the individual muscle level remain limited. Consequently, much of the existing literature in adults^{3–5,8,10,12–24} and children^25,26 has relied on qualitative ratings or quantification confined to single slices or central muscle sub-regions. Recent advances have introduced semi-automated²¹ and fully automated methods^22,24 to quantify select individual muscles or combined muscle groups, leading to updated conclusions about muscle involvement and fat progression patterns. Key revisions include: (1) while FSHD was historically thought to primarily affect specific muscles (e.g., scapular fixation muscles, hamstrings, tibialis anterior, and gastrocnemius), it is now evident that all muscles can be affected across the lifespan; (2) the pattern of muscle involvement and progression is highly variable across patients and muscle groups; and (3) within individual muscles, disease progression is heterogeneous, with varying degrees of fat infiltration and diverse patterns of structural degradation.

As trials have progressed—such as the recently completed trial (NCT05397470) and ongoing studies (NCT05747924 and NCT05548556)—the common approach has been to monitor a subset of individual muscles or combined muscle groups. This involves grouping muscles with a prespecified range of fat infiltration, deemed “at risk,” into a single category and using this aggregated number as a treatment response metric²². While this streamlined strategy aligns with the single-biomarker approach used in DMD²⁷ it has several notable limitations: (1) it omits muscles above and below the specified threshold that may also be changing; (2) it disregards baseline muscle status and identity, which are known to influence change rates; (3) it overlooks fat distribution patterns, with confluence of fat being a characteristic feature of the disease; and (4) it averages data from muscles with distinct trajectories, effectively diluting the signal from muscles undergoing meaningful change by combining them with muscles showing little or no change. This signal dilution may reduce sensitivity to detect subtle muscle-level changes that could result clinically meaningful functional changes, especially in the context of heterogeneous disease progression. Collectively, these limitations have constrained our ability to fully leverage the capabilities of MRI to capture the complexity of FSHD progression and its functional implications.

While a whole-body MRI analysis provides a more in-depth, personalized description of FSHD disease state, the vast amount of data generated also introduce significant analytical challenges. Furthermore, muscle MRI measurements do not provide the full picture of an individual’s disease state: other data types, such as clinical assessments and disease descriptors (both descriptive modulators like D4Z4 allele length and functional task performance), can also be incorporated to provide an integrated patient assessment tool. These challenges point to the need for updated methods beyond traditional parametric statistics to interpret and analyze the data effectively. Machine learning techniques, particularly ensemble methods, have shown considerable promise in addressing the complexities of heterogeneous disease in order to predict patient-specific disease progression. Indeed, machine learning has been applied in the context of clinical trials for other diseases through synthetic control arms^28,29 and AI-driven digital twins of patients³⁰. These examples highlight the increasing role of AI-based models in addressing disease heterogeneity, which remains a major challenge in FSHD trials.

The primary objective of this study is to develop a machine learning model that incorporates detailed MRI measures along with clinical data and use the model to predict and advance our understanding of FSHD disease progression. Beyond individual outcome prediction, our goal is to create an FSHD patient digital twin—a model that simulates the natural course of disease progression in untreated patients.

Methods

Data collection

We leveraged multiple FSHD retrospective datasets, combined into a “data lake.” The datasets included MRI scans collected from > 100 patients with FSHD from seven different studies: Wellstone³¹ cohort (n = 34, P50 AR065139, J. Chamberlain PI, S. Tapscott co-PI); Kennedy Krieger Institute (KKI) cohort (n = 30) (1K23NS091379, D. Leung PI), a subset of the Fulcrum Phase II DUX4 placebo cohort (n = 20, NCT05397470), FSHD Global Research Foundation Registry cohort (n = 28), and MOVE Plus patients (n = 24) (Statland PI, funded by Avidity Biosciences, FSHD Canada). All subjects were adults and provided informed consent. All experimental protocols used to collect the data were in accordance with relevant guidelines/regulations and were approved by institutional review boards (IRB) or ethics committees associated with the University of Kansas Medical Center, Kennedy Kreiger Institute, Fulcrum Therapeutics, University of Washington, University of Rochester, or FSHD Global Research Foundation. Subjects provided informed consent for data collection and aggregation. The IRB at Seattle Children’s Hospital approved the protocol for analysis of aggregated data as presented in this paper. The scans varied in coverage and acquisition method (Table 1).

Table 1.

Study cohort and sample characteristics. Demographic information, coverage, and time point distribution for each study cohort included. Means ± standard deviations are reported.

	Region of Coverage	# of Subjects	Timepoint Distribution	Height (cm)	Weight (kg)
Site	Distribution of scans	Sex Distribution	# Scans per Timepoint	Mean +/- Stdev (Range)	Mean +/- Stdev (Range)
WELLSTONE	LE (96)	34 M - 15, F - 19	Baseline: 34 52 Weeks: 30 104 Weeks: 27 156 Weeks: 5	175.11 +/- 11.34 (152.0 - 201.0)	81.08 +/- 13.63 (53.4 - 111.0)
KKI	LE/CO (69/73)	30 M - 12, F - 18	Baseline: 30 12 Weeks: 27 36 Weeks: 26 60 Weeks: 30 84 Weeks: 29	173.13 +/- 11.29 (154.0 - 196.0)	81.94 +/- 19.18 (47.8 - 118.9)
MOVEPlus	LE/FB (19/13)	24 M - 20, F - 4	Baseline: 24 52 Weeks: 8	177.63 +/- 10.42 (157.0 - 193.0)	82.81 +/- 16.22 (55.7 - 122.3)
Fulcrum	LE/FB (21/39)	20 M - 15, F - 5	Baseline: 20 48 Weeks: 20 96 Weeks: 20	178.05 +/- 11.21 (152.4 - 199.0)	86.0 +/- 15.44 (54.9 - 119.0)
FSHD Global	FB (29)	29 M - 17, F - 12	Baseline: 29	175.25 +/- 10.78 (155.0 - 200.0)	78.18 +/- 15.73 (55.0 - 110.0)

MRI segmentation

We utilized an AI-based approach similar to our previously published algorithm and described methods³² to segment the boundaries of up to 118 muscles (depending on coverage) across the whole body from the Dixon water images. As per our prior work, trained segmentation engineers manually reviewed and edited for accuracy (process we call “vetting”) utilizing the manual editing tools (i.e. paint, draw, erase) in 3D Slicer (v4.11). This process required between one and three hours, depending on the level of fatty replacement (with higher fatty replacement muscles requiring more user interaction time for contour verification).

Muscle distribution metrics

Since disease progression is often heterogeneous within each muscle in FSHD, the Stage 1 model focused on slice-level, regionally sensitive level metrics. First, we computed muscle and fat quantities within the cross-section of each muscle in each axial slice and represented those metrics as a function of superior-inferior position along the muscle (as described previously²⁴). Slice-level measurements are particularly sensitive to localized changes in fat distribution, which often occur within discrete anatomical zones. To summarize regional variation along the length of each muscle, we divided the muscle into eight longitudinal intervals along the superior-inferior direction spanning 10% increments — specifically: 10–20%, 20–30%, 30–40%, 40–50%, 50–60%, 60–70%, 70–80%, and 80–90%. Within each of these regions, we calculated averaged values for slice-level metrics such as cross-sectional area (CSA), lean muscle CSA, fat fraction (FF), and spatial heterogeneity metrics— fat variation, kurtosis, skewness, and Moran’s index. This approach enabled us to assess how fat infiltration and structural changes varied along the muscle’s length in a standardized and interpretable way. Fat variation (in units of %) for each region was calculated as the standard deviation of fat fraction across all pixels in the boundary volume. Kurtosis and skewness (expressed as percentages) for each region were calculated based on the distribution of pixel fat fraction values within the boundary cross-section. Moran’s index (normalized) for each region assesses whether nearby or neighboring areas are more similar (or dissimilar) than what would be expected under random spatial distribution; our implementation was based on Santago et al.³³.

Muscle-level metrics derived from MRI data

To complement the slice-level analysis, we computed muscle-level aggregate metrics by collapsing values across all slices within each muscle. These metrics provide an overall summary of muscle condition. Fat fraction (FF, %) was calculated using the ratio of fat to the combined fat and water signal intensities. Boundary volume (BV, ml) was determined as the total segmented muscle volume. Lean muscle volume (LMV, ml) was calculated by subtracting fat volume from the total boundary volume using the formula: LMV = BV – (FF × BV).

Bone metrics

The volume, average cross-sectional area, and fat fraction of each femur bone was determined. The CSA and volume were used to standardize muscle volume metrics and fat fraction was incorporated because prior work has demonstrated a correlation between trabecular fat fraction and bone mineral density measured by quantitative computed tomography³⁴.

Analysis of baseline MRI muscle data for model feature development

The baseline muscle fat fraction and lean muscle volume data were analyzed to examine trends across the FSHD population and to establish normative ranges for metrics within the FSHD population. Analysis included: fat fraction for each muscle, fat variation for each muscle, and lean volume (adjusted for body size) for each muscle. In addition to standard descriptive statistics, heatmap visualizations of muscle involvement, with elements organized by the average fat fraction across muscles (within each patient) and TUG times were used to characterize the relationships between pattern of muscle involvement and overall functional and global fat replacement. The relationships between total volume and the product of femur cross-sectional area and lean muscle volume across patients with FSHD were determined, to provide a size adjusted value.

To integrate contextual information about muscle involvement into our predictive model, we developed two derived metrics based on the relative ranking of muscles by disease involvement. First, we calculated the average fat fraction and standard deviation for each muscle across all patients at baseline. Using these values, we computed a fat fraction z-score for each muscle in each patient to standardize its involvement relative to the cohort distribution. Next, we ranked all muscles based on their cohort-wide average fat fraction, from lowest (least typically involved) to highest (most typically involved). This ranking allowed us to define two contextual metrics for each muscle in each patient. The higher-ranking muscle score for a given muscle was calculated as the average fat fraction z-score of all muscles in that patient ranked higher (i.e., more typically involved muscles with lower ranking numbers). The lower-ranking muscle score was similarly calculated as the average fat fraction z-score of all muscles in that patient ranked lower (i.e., less typically involved muscles with higher ranking numbers). Biologically, a higher higher-ranking muscle score indicates that muscles normally involved earlier in the disease process are already highly infiltrated in this patient, suggesting a higher likelihood that the current muscle will also progress. In contrast, a higher lower-ranking muscle score suggests that muscles usually spared until later stages are already highly involved, indicating the disease may have reached a plateau phase, and the current muscle is therefore less likely to progress.

Clinical and functional data

The D4Z4 repeat length and timed up-and-go (TUG) was included in the baseline description of each patient. For a subset of patients, TUG measurements were also available at follow up time points. All subjects that had longitudinal MRI measures and thus were simulated with the disease progression model were confirmed FSHD Type 1. Unfortunately, disease severity metrics were not the same across all cohorts and thus were not used in the model. Individual patient age and sex were not included in the model because: (1) preliminary analysis indicated that they were not features of high significance, and (2) excluding these features allowed for maximal use of the data because age and sex were not known for all patients.

Three-Stage, Multi-scale model development

The random forest algorithm was chosen for its ability to handle complex, non-linear relationships and its resistance to overfitting. Prior research has established the suitability of this approach for predicting disease progression^35,36. To incorporate both regional and total muscle metrics, we developed a multi-stage random forest training and validation approach (Fig. 1). The first two models are designed to predict individual muscle progression over time, with the third model designed to incorporate the output from the first two model stages to predict change in TUG performance. The first stage, one unique random forest model, included training on a subset of the data to predict the changes in fat fraction and lean cross-sectional area at the regional level. This allows the model to capture how the local tissue structure influences progression of fat infiltration and atrophy in a specific region. This model was trained on 10% of the full data set, which included 6330 muscle regions. The data used for the stage 1 model training were not included in any other training or testing datasets. This approach ensured that each stage was trained and tested on different subsets of data, minimizing the risk of overfitting from incorporating outputs of earlier model stages. The second stage, a two parallel-running random forest model, included training another subset of the data (that was not included in the prior training datasets) to predict the whole-muscle level changes in fat fraction and lean muscle volume. Each of these samples were first run through the stage 1 model to predict regionalized changes; then the cross-sectional area weighted average of the regional fat fraction changes was calculated and inputted into the stage 2 model. This model was trained on 70% of the full data set, which included 4513 muscles, and the model was tested on 833 muscles. Cases with incomplete inputs for key variables were excluded from model training and testing, and no imputation was performed, ensuring that only complete data were used to maintain prediction reliability. For data points where fat fraction decreased or lean muscle volume increased over time, these changes were set to zero. As a result, the model focuses on predicting progression (i.e., degeneration) and does not capture potential improvements. Therefore, a prediction of no change indicates either stabilization or improvement The outputs of the stage two models were combined to calculate functional group changes, by calculating the volume-weighted average of the individual fat fractions at the joint level, and summing the lean muscle volumes, normalized by the product of height and mass. Finally, the right and left group-level predicted changes were averaged and incorporated into the stage 3 random forest model to predict change in TUG measurements. The stage 3 model was trained with 26 data points and tested with 4 data points (only a subset of patients included had both coverage of major lower limb muscle groups as well as multiple measurement of TUG).

Fig. 1.

Schematic of the multi-scale disease progression modeling framework. Flow diagram illustrating the three-stage modeling framework used to predict muscle-level and functional progression in FSHD. The pipeline begins with MRI-based segmentation and analysis of muscle and bone from whole-body mDixon MRI scans at baseline. Stage 1 predicts regional fat fraction changes using sub-region data (eight regions per muscle). Stage 2 aggregates regional predictions to muscle-level, combining them with patient data to predict changes in muscle fat fraction (FF) and lean muscle volume (LMV). Stage 3 integrates muscle-level predictions with baseline timed up-and-go (TUG) measurements to predict changes in functional performance at follow-up. This framework captures the progression of disease from local muscle regions to functional-level outcomes.

For each stage, a random forest model consisting of around 500 trees was trained on the training dataset. Models were implemented in MATLAB (Mathworks, Natick, MA, USA), using the combination of templateTree and fitrensemble functions. Grid search was conducted to optimize key hyperparameters such as the number of trees, maximum depth, and minimum samples per split. The out-of-bag method was used to prevent overfitting. Shapley Additive exPlanations (SHAP) analysis^37,38 was used to identify the key variables contributing to predicted outcome. Final list of features used in each model are listed and described in Table 2.

Table 2.

Features used in the multi-stage models. Parameters and descriptions for features used in each stage of the multi-stage model: Stage 1 (regional-level model), Stage 2 (muscle-level model), and Stage 3 (functional-level model). For details on feature derivation and integration, refer to the Methods section.

Parameter	Description
Stage 1. Regional-level model
Region Fat Fraction	Fat fraction within a region
Region CSA	Region CSA
Relative CSA	Ratio of region CSA to average muscle CSA
Difference From FFneighbors	Difference between region FF and neighboring regions FF
FF Range Region	Max regional FF - Minimum regional FF
Region FF Variation	Standard deviation of FF in region
Region FF Skewness	Skewness of FF in region
Region FF Kurtosis	Kurtosis of FF in region
Region FF Moran	Moran’s Index of FF in region
Stage 2. Muscle-level model
Region Prediction	Output of stage 1 prediction
Fat Fraction Z score	Z score of fat fraction relative to FSHD population at baseline
Fat Fraction	Fat fraction at baseline
Higher ranking muscles score	Average of Fat fraction z scores in muscles with higher ranking
Lower ranking muscles score	Average of Fat fraction z scores in muscles with lower ranking
LMV to bone Z score	Z score of lean muscle volume to bone CSA, compared to FSHD population
Femur CSA	Cross sectional area of femur
Fat fraction range across regions	Max fat fraction across sub regions – min fat fraction across sub regions
Femur Fat Fraction	Fat fraction within femur
D4Z4 repeat length	D4Z4 repeat length (integer value)
Bone to Height Z-score	Z score of bone to height
TUG time	Time to up and go
Muscle	Muscle name (classifier)
Stage 3. Functional-level model
TUG at Baseline	Timed up-and-go at baseline
Knee extensors normalized change in LMV from baseline	Output from combined prediction of Stage 2 LMV model, normalized to height*mass
Hip abductors normalized change in LMV from baseline	Output from combined prediction of Stage 2 LMV model, normalized to height*mass
Hip extensors normalized change in LMV from baseline	Output from combined prediction of Stage 2 LMV model, normalized to height*mass
Hip flexors normalized change in LMV from baseline	Output from combined prediction of Stage 2 LMV model, normalized to height*mass
Trunk extensors normalized change in LMV from baseline	Output from combined prediction of Stage 2 LMV model, normalized to height*mass

Model evaluation

Each model was evaluated by assessing the ability to predict annualized change in fat fraction for each region (for Stage 1), muscle-level fat fraction or lean muscle volume (for Stage 2) for both the training and testing data samples. The Stage 3 model was evaluated by assessing the ability to predict change in timed up-and-go. The performance accuracy metrics included: mean error (ME), root-mean-square error (RMSE), as well as Pearson’s correlation coefficient between predicted and actual fat fraction or lean muscle volume change (r) and its p-value. Feature importance scores were extracted from the random forest model using SHAP analysis to identify the most influential predictors of fat fraction or lean muscle volume changes.

Resuts

Fat fraction varied substantially across muscles and patients (Fig. 2). On average, the semimembranosus (hamstring) muscle was the most affected in the cohort while the popliteus muscle was least affected. Heatmap visualization of fat fraction across all muscles (organized average fat fraction across individuals) and patients at baseline (organized by TUG) reveal an overall pattern of muscle involvement across disease severities (Fig. 3). As TUG time increases, more muscles have fat fraction levels in the high range, consistent with the progressive nature of the disease. The order of muscles is not regional: the muscles in the ‘earlier’ group are in both upper and lower regions of the body; similarly, muscles in the ‘later’ group are muscles from all regions as well. Substantial heterogeneity across patients and muscles is most apparent in the patients in the middle range of involvement.

Fig. 2.

3D visualizations of muscle fat infiltration in patients with FSHD. 3D whole-body muscle visualizations showing the distribution of fat infiltration (color-coded) across patients with varying levels of disease severity. Muscles are colored from low (yellow) to high (orange/red) fat infiltration, illustrating the heterogeneous nature of FSHD muscle involvement. The images highlight regional differences in muscle degeneration, with some muscles showing early or severe fat replacement while others remain relatively spared. This visualization underscores both the widespread and individualized progression patterns of FSHD.

Fig. 3.

Baseline muscle-level fat fraction heatmap ordered by timed up-and-go (TUG). Heatmap showing baseline muscle-level fat fraction (%) for each subject, ordered by TUG time (seconds) at baseline (lower plot). Muscles are arranged from most commonly involved (top) to least involved (bottom) across the cohort. Color indicates fat fraction (white/yellow = low, orange/red = high), with white cells for missing data. The pattern reveals a relationship between higher TUG times and increased fat infiltration in multiple muscles, illustrating disease progression and heterogeneity in FSHD.

The region-specific random forest model (Stage 1) showed a significant correlation with measured regional fat fraction per muscle (r = 0.43, p < 0.001), with a mean error of −0.32% and an RMSE of 3.5% in the testing dataset (Fig. 4). SHAP analysis revealed the parameters: Regional Fat Fraction Kurtosis, Regional Fat Fraction Variation, Difference from FF Neighbors, and Relative CSA had the greatest influences on prediction of fat fraction change at the regional level.

Fig. 4.

Stage 1 Model: Regional-Level Fat Fraction Change Prediction. (A, D) Predicted vs. measured regional fat fraction change (%) for training (A) and testing (D) datasets. Strong correlation in training (r = 0.89, p < 0.001) and moderate in testing (r = 0.43, p < 0.001) demonstrate predictive performance. (B, E) Bland-Altman plots comparing measured and modeled changes. Training shows minimal error (mean = 0.00%, RMSE = 1.62%), while testing shows modest bias (mean = 0.32%, RMSE = 3.50%). (C, F) Histograms of prediction errors (model – measured) with centered distributions in training (n=6330) and testing (n=7218). (G) SHAP beeswarm plot showing the contribution of key features to fat fraction change predictions. Each dot represents a regional sample, with SHAP values indicating feature impact and colors reflecting feature values (red = high, blue = low). Top predictors include fat fraction kurtosis, skewness, variation, neighbor differences, and CSA metrics.

At the muscle level (Stage 2), the random forest model predicted fat fraction change with a significant correlation to measured changes (r = 0.5, p < 0.001), a mean error of −0.3%, and an RMSE of 2.2% (Fig. 5). Regional Prediction (output from Stage 1 model), femur cross-sectional area, higher-ranking and lower-ranking muscle scores, and muscle had the greatest influence on prediction of muscle-level fat fraction change. For predicting lean muscle volume change, the Stage 2 model showed significant correlation with measured changes (r = 0.57, p < 0.001), a mean error of −0.8 ml, and an RMSE of 8.3 ml (Fig. 6). Femur cross-sectional area, D4Z4 repeat length, LMV to femur CSA z-score, higher-ranking muscle scores, and muscle had the greatest influences on prediction of muscle-level lean muscle volume change. Combining the predicted individual fat fractions and lean muscle volumes to calculate functional group-level changes in fat fraction and lean muscle volume accurately predicted variability in functional-group level changes in the testing dataset (Fig. 7). For group-level fat fraction change prediction, all RMSE values were less than 2.5%, with the exception of the trunk flexors (5.18%). Group-level lean muscle volume change predictions showed RMSE values ranging from 3.3 ml (hip external rotators) to 26.7 ml (knee extensors).

Fig. 5.

Stage 2 Model: Muscle-Level Fat Fraction Change Prediction. (A, D) Predicted vs. measured muscle-level fat fraction change (%) for training (A) and testing (D) datasets. Strong training correlation (r = 0.95, p < 0.001) and moderate testing correlation (r = 0.50, p < 0.001) demonstrate good model performance. (B, E) Bland-Altman plots comparing measured and modeled changes. Training error is minimal (mean = 0.04%, RMSE = 0.93%) while testing shows a slight negative bias (mean = −0.31%, RMSE = 2.16%). (C, F) Histograms of prediction errors (model – measured) for training (n=4513) and testing (n=833) samples, centered around zero. (G) SHAP beeswarm plot showing contributions of key features to muscle-level fat fraction change prediction. Each dot represents a muscle sample, with SHAP values indicating feature impact and color representing feature values (red = high, blue = low). Top predictors include regional prediction (Stage 1 output), D4Z4 repeat length, baseline fat fractions (muscle and femur), femur CSA, ranking scores, fat fraction ranges, and muscle identity.

Fig. 6.

Stage 2 Model: Lean Muscle Volume Changes Prediction. (A, D) Predicted vs. measured lean muscle volume change (ml) for training (A) and testing (D) datasets. The model showed strong correlation in training (r = 0.95, p < 0.001) and moderate correlation in testing (r = 0.57, p < 0.001). (B, E) Bland-Altman plots showing measured minus modeled differences against their average for training and testing datasets. Training error is minimal (mean = 0.02 ml, RMSE = 3.05 ml), while testing shows a small negative bias (mean = −0.80 ml, RMSE = 8.13 ml). (C, F) Histograms of prediction errors (model – measured) for training (n=4513) and testing (n=833) samples, with distributions centered around zero. (G) SHAP beeswarm plot illustrating the impact of features on lean muscle volume change predictions. Each dot represents a muscle sample; the x-axis shows SHAP values (feature contributions), and the color indicates feature value (red = high, blue = low). Key predictors include Stage 1 regional predictions, LMV to bone z-score, D4Z4 repeat length, baseline femur fat fraction, femur CSA, ranking scores, and fat fraction range across regions.

Fig. 7.

Evaluation of the output of stage 2 when determining group-level fat fraction and lean muscle volume. (A) Predicted vs. measured fat fraction (FF) changes (%) for various muscle functional groups across the body. Each scatter plot shows predictions for a specific functional group, including shoulder abductors, hip abductors, trunk extensors, hip flexors, knee extensors, and ankle dorsiflexors, among others. Filled circles represent training data, and open circles represent testing data. Model performance metrics—mean error and RMSE (root mean square error)—are reported for the testing data of each group. The diagonal line indicates perfect prediction. (B) Predicted vs. measured lean muscle volume (LMV) changes (ml) for the same functional groups. Each plot shows model performance for predicting LMV changes in the corresponding muscle group, with training and testing samples shown as filled and open circles, respectively. Model accuracy, including mean error and RMSE, is reported for the testing dataset in each group. The plots demonstrate the model’s ability to predict group-level changes in both FF and LMV across multiple muscle groups, reflecting functional variations in disease progression.

The functional-level model (Stage 3) predicted change in TUG from the first to the second time point with an RMSE of 0.75 s in the testing dataset (Fig. 8). SHAP analyses revealed that the TUG at baseline had the largest influence on the predicted change in TUG, followed by the normalized change in lean muscle volume of the Trunk Exensors, Hip Extensors, and Knee Extensors muscle functional groups.

Fig. 8.

Stage 3 Model – Prediction of change in timed up-and-go (TUG) – performance. (A) Predicted vs. measured change in TUG (seconds) for the model. Filled circles represent training data (n=26) and open circles represent testing data (n=4). The diagonal line indicates perfect prediction. The test dataset shows a mean error of 0.60 seconds and an RMSE of 0.75 seconds. (B) SHAP beeswarm plot illustrating the contribution of key features to the model’s TUG prediction. Each point represents a sample, with horizontal position indicating the SHAP value (impact on prediction) and color representing the feature value (red = high, blue = low). The most influential predictors include TUG at baseline, and normalized changes in lean muscle volume (LV) of knee extensors, hip abductors, hip extensors, hip flexors, and trunk extensors. This analysis highlights the factors driving functional performance prediction in the model.

Discussion

A major challenge for drug development in neuromuscular disease is how to measure the effects of treatment in chronic progressive disorders that are marked by high variability in severity at baseline and in rates of progression. Clinical trials in FSHD currently adopt a “one-size-fits-all” approach to evaluating drug efficacy. Strategies that rely on composite metrics from MRI or performance tests risk regressing small changes to the mean, thereby reducing sensitivity in detecting meaningful differences and prolonging the required trial duration. Conversely, using a single task as a reporter is only effective if participants are pre-selected based on muscle involvement relevant to that task. Otherwise, a substantial proportion of participants—potentially unevenly distributed across placebo and treatment groups—may fail to exhibit measurable change in task performance. In this study, we present a novel ‘digital twin’ approach to predicting personalized disease progression at multiple length scales: muscle region, muscle, group, and functional levels. The developed method integrates a three-tiered machine-learning model, each component designed to capture progression at a distinct scale. This framework enables the prediction of which muscles, and to what extent, will undergo changes over time during the natural course of the disease, offering a more precise and individualized approach to monitoring FSHD progression in clinical trials. Furthermore, the feature importance analyses provide insight into which factors best predict disease progression. In these models, metrics of fat fraction variability (within regions and whole muscle) were most associated with fat progression changes over one year, with bone metrics playing an important role in lean muscle volume decline over one year.

Given the heterogeneity and nonlinearity of the disease, each patient has a specific set of muscles that are most likely to change at a given time. As might be expected in a heterogeneous disease, some muscle groups changed more than others in each patient, which has implications for the functional metrics that best identify disease progression in a particular patient. For example, the muscles that best predict TUG – the trunk extensors, hip extensors, and knee extensors – only changed in a subset of patients; therefore, TUG only changed in a subset of patients as well. Not surprisingly, the TUG at the baseline timepoint demonstrated a strong relationship with predicted change in TUG, as well. Taken together, these results suggest that the disease progression model provides a new approach for clinical trials that allows for patient-specific selection of functional metrics that are most likely to detect deviations from each individual’s disease progression and/or selection of qualified patients if specific functional metrics are of interest.

Analysis of the models provided insight into the parameters that influenced model prediction of disease progression. At the regional (Stage 1 model) level, the initial fat fraction, the fat variation, and the kurtosis had the greatest influence on predicted fat fraction changes, illustrating that progression patterns are associated with the heterogeneity of fat within the muscle. At the whole muscle (Stage 2 model) level, the predicted fat fraction change from the Stage 1 model, the muscle, the D4Z4 repeat length, the associated muscle group, femur size, femur fat fraction, and metric of overall involvement influenced change in fat fraction prediction. The explanations for some of these predictive features are intuitive and consistent with current knowledge, while others are novel and require further investigation. The fact that ‘muscle’ was a key feature contributing to prediction accuracy indicates that each muscle’s identity influences its likelihood of progression, supporting the notion of muscle-specific progression patterns that has already been demonstrated³⁹. Similarly, the fact that ‘D4Z4 length’ is associated with faster progression is also expected as per prior work^3,16,20. However, the mechanisms by which bone size and composition affect disease progression represent additional complexities yet to be unraveled. It is possible that these bone metrics (e.g., femur CSA and fat fraction) reflect unmodeled confounders such as sex, body size, or age, which were not included in our model. Future studies explicitly incorporating these variables will be important to clarify these associations and their potential biological relevance. Overall, these results demonstrate that the complex nature of disease progression can be handled within iterative model building, allowing insights into primary predictors and the individualized nature of FSHD.

The model presented in this study builds upon significant prior work in the field demonstrating general features of disease progression as determined by muscle & fat measurements on MRI^{3,6,7,10,14,15,18,21–23,40–48}. For example, a 2017 study by Andersen et al. followed 45 patients over one year, demonstrating that quantitative MRI could detect disease progression, with increased fat infiltration corresponding with declines in muscle strength and function². Similarly, Fatehi et al. conducted a long-term follow-up of thigh muscles in FSHD patients, finding that quantitative MRI effectively monitored progressive fatty infiltration over time⁵. Another common finding in the literature is that a small percentage (~ 5%) of fat-affected muscles also exhibit signal elevation on short tau-inversion-recovery (STIR+) sequences^3,7,9,48. This STIR + signal can be indicative of intramuscular edema or inflammation^3,49–51. While this biomarker has been suggested by our group and others to foreshadow faster progression, as examples^3,4,6,7,42these studies did not discriminate between muscle fat fraction at baseline, fat pattern, or muscle identity. Notably, the stability of the STIR + signal—even when challenged with immunosuppression and steroid treatment³⁹—combined with long imaging times required by STIR sequences⁵², highlights a limitation of STIR as a dynamic biomarker for disease progression. STIR imaging was not consistently available across the cohorts included in this study, further limiting its potential utility in this analysis. Future work incorporating both STIR imaging and fat fraction/fat variation metrics, as well as complementary imaging markers such as water T2 (wT2), could help clarify how these biomarkers independently or synergistically predict disease progression.

There are limitations to this study that should be acknowledged. First, while the model was developed based on what is likely the most complete dataset of muscle-level MRI in FSHD patients to date, more data would improve the robustness and predictive abilities of the model. Second, several parameters not included in this study would likely improve model performance. Additional measures like methylation, further clinical assessments, and blood biomarkers represent promising possible additions to the progression model. Third, we had limited access to functional measurements at the second time point; therefore, the Stage 3 model analysis was somewhat limited and only focused on one functional measurement (TUG) and a smaller subset of subjects and muscle groups based on available longitudinal task data. Despite these factors, the model lends itself to rapid future extension, with incorporation of more functional measurements spanning the lower and upper bodies. Lastly, the model was trained based only on adult FSHD patient data; therefore, its robustness for predicting disease progression in pediatric patients with FSHD has not been evaluated. A new training dataset, that incorporates the effects of growth assessed by a similar metric of femur volume, would provide the opportunity for expanding the applicability of the model to pediatric populations. Additionally, the model was designed to focus on progression (i.e., fat fraction increase and lean muscle volume decrease), with data points showing improvement (fat fraction decrease or lean muscle volume increase) set to zero. Consequently, predictions of no change may reflect either stabilization or actual improvement, representing a limitation in the model’s ability to capture recovery processes. To ensure transparency in reporting, we have aligned this study with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD)⁵³ Level 2 (model development with internal validation; see Supplemental Table S1 for a mapping of the manuscript sections to TRIPOD items) and followed key guidelines for transparent and reproducible model development.

This work represents a significant advancement in the use of machine learning and imaging to address the challenges of disease heterogeneity in FSHD. By integrating advanced MRI-derived muscle metrics with clinical and biological data, the proposed multi-scale framework reveals that individual muscle progression over a year interval is predictable, given a comprehensive integration of personalized data at baseline. The digital twin model serves as a powerful benchmark for assessing untreated progression, enhancing the precision and efficiency of tracking natural history changes over time. This advancement has broad potential applications in clinical trial design, including the use of personalized digital twins as surrogate placebos and leveraging the progression model’s identified features to better isolate ‘at-risk’ muscles in traditional trial designs. While incorporating additional biomarkers and expanding datasets will further refine predictive capabilities, this study lays the foundation for applying personalized machine learning and imaging approaches to address variability inherent in neuromuscular diseases and to enable more precise, efficient, and patient-tailored clinical trials.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Author Contributions Statement: SB, LR, KC, MP, OD, DL, SF contributed to design, analysis, and paper writing. ST, RT, JS, DS, LW, MW, LL, MJ, DL, and SF helped with data collection and paper revision. OD, MC, KC, MP, and JM assisted in data analysis and workflow. DS, SF, and DL assisted in study design, data interpretation, and revision of the paper. All authors have read and approved the final submitted manuscript.

Data availability

Access to model outputs or key components may be granted upon reasonable request for academic research purposes under appropriate agreements. Requests should be made to: Silvia Blemker at silvia.blemker@springbokanalytics.com.

Declarations

Competing interests

Lara Riem, Olivia DuCharme, Megan Pinette, Kathryn Eve Costanzo, and Silvia Blemker are employees of Springbok Analytics. LR, EC, and OD have stock options to the company. SB is co-founder and owns stock in the company. FSHD Global is an investor in SB Analytics. JS is consultant or on advisory board for Avidity, Dyne, Fulcrum, Roche, Kate, Alnylam, Epicrispr, MiRecule, and Sanofi. SF is a consultant for Avidity, Dyne, Fulcrum, Kate, and Epicrispr. The remaining authors have no competing interests to declare.