Training Together, Diagnosing Better: Federated Learning for Collagen VI–Related Dystrophies

Abstract

The application of Machine Learning (ML) to the diagnosis of rare diseases, such as collagen VI-related dystrophies (COL6-RD), is fundamentally limited by the scarcity and fragmentation of available data. Attempts to expand sampling across hospitals, institutions, or countries with differing regulations face severe privacy, regulatory, and logistical obstacles that are often difficult to overcome. The Federated Learning (FL) provides a promising solution by enabling collaborative model training across decentralized datasets while keeping patient data local and private. Here, we report a novel global FL initiative using the Sherpa.ai FL platform, which leverages FL across distributed datasets in two international organizations for the diagnosis of COL6-RD, using collagen VI immunofluorescence microscopy images from patient-derived fibroblast cultures. Our solution resulted in an ML model capable of classifying collagen VI patient images into the three primary pathogenic mechanism groups associated with COL6-RD: exon skipping, glycine substitution, and pseudoexon insertion. This new approach achieved an F1-score of 0.82, outperforming single-organization models (0.57–0.75). These results demonstrate that FL substantially improves diagnostic utility and generalizability compared to isolated institutional models (see Figure 1). Beyond enabling more accurate diagnosis, we anticipate that this approach will support the interpretation of variants of uncertain significance and guide the prioritization of sequencing strategies to identify novel pathogenic variants.

1. Introduction

Collagen VI is an extracellular matrix (ECM) protein that forms a beaded microfibrillar network surrounding skeletal muscle fibers, contributing to muscle function and myofiber integrity. It is composed of three alpha chains, α1(VI), α2(VI) and α3(VI), encoded by COL6A1, COL6A2 and COL6A3 genes, respectively. The assembly of collagen VI occurs as a highly orchestrated, multistep process: the three alpha chains intertwine into a triple-helical, heterotrimeric monomer, which subsequently dimerizes and assembles into tetramers. Tetramers are secreted into the extracellular space where they associate to form a beaded microfibrillar network. Pathogenic variants in any of the three COL6 genes disrupt collagen VI assembly and function, leading to a spectrum of COL6-related muscular dystrophies, ranging from the mild Bethlem muscular dystrophy (BM), via an intermediate phenotype, to the severe Ullrich congenital muscular dystrophy (UCMD).¹

A diagnosis of collagen VI-related dystrophies (COL6-RD) can be suspected in an individual with characteristic clinical features, muscle imaging features, muscle immunohistochemical features, and collagen VI abnormalities in dermal fibroblast cultures.² This initial subjective evaluation is mostly qualitative, and definitive diagnosis relies on the identification of a pathogenic variant in one of the three COL6 genes via molecular genetic testing. However, diagnosis remains challenging despite advances in next-generation sequencing technologies. The complexity of variant identification has increased with the discovery of more intronic variants that retain introns at low levels, complicating the interpretation of genetic data (C.B., unpublished data). Furthermore, there are still many variants of unknown significance, adding another layer of difficulty in making an accurate diagnosis. Although some methods for quantification have been developed³, they do not consider the microfibrillar structure that is crucial in assessing the function of collagen VI; therefore, alternative quantitative methodologies to evaluate collagen VI abnormalities are needed.

COL6-RD is caused either by recessive loss-of-function variants or, more commonly, by de novo dominant-negative pathogenic variants. Biallelic recessive variants typically result in loss-of-function or nonsense-mediated RNA decay, and are readily identified by collagen VI immunostaining as the collagen VI signal is reduced or absent. Heterozygous dominant-negative variants, in contrast, can be broadly categorized into three main groups based on their pathogenic mechanism: (i) pathogenic variants causing in-frame exon skipping (e.g. splice site variants, in-frame deletions) in any of the three COL6 genes, (ii) pathogenic variants causing single-amino acid substitutions which disrupt the Gly-X-Y motif of the highly conserved N-terminal triple helical domain of any of the three COL6 genes, and (iii) the recurrent deep-intronic pathogenic variant in intron 11 of COL6A1 gene (c.930+189C>T), which creates a cryptic donor splice site that prompts the insertion of a 72-nt in-frame pseudoexon^1,4–9. De novo dominant-negative pathogenic variants result in the production of structurally abnormal but still assembly-competent alpha-chains. These mutated chains thus retain the ability to assemble into dimers and tetramers but ultimately lead to the deposition of a dysfunctional collagen VI matrix, as revealed by collagen VI immunostaining. Although collagen VI immunostaining of patient-derived fibroblast cultures can detect collagen VI matrix abnormalities, distinguishing among these three groups based on visual inspection remains challenging due to subtle differences in staining patterns (Figure 2).

Figure 2:

Collagen VI immunofluorescence images from control and patient-derived dermal fibroblast cultures. Each image represents one of the four classes. Scale bar = 50 μm.

Machine Learning (ML) has become a powerful tool in medical diagnosis and treatment, enabling the identification of complex patterns within biomedical data. However, its successful application in rare diseases research, such as studies involving COL6-RD, is hindered by the limited availability of large and diverse datasets necessary for training accurate and generalizable models. Traditional approaches to data sharing, such as centralized aggregation (Figure 3 Left), consolidate data from multiple sources onto a single server. While this can increase dataset size and heterogeneity, it raises significant concerns regarding privacy, regulatory compliance, and logistical challenges, including compliance with the General Data Protection Regulation (GDPR)¹⁰, data transmission costs, and restrictions imposed by data localization laws.

Figure 3:

Traditional (Centralized) training (left) compared to Federated training implemented on the Sherpa.ai FL platform (right). In FL, only model updates (not raw images) are shared with a central aggregator, enhancing privacy and regulatory compliance.

Federated Learning (FL) offers a promising alternative by enabling the training of ML models across distributed institutions (nodes) without sharing raw data (Figure 3, Right). In FL, each node keeps its data locally and only shares model updates with a central aggregator, which integrates these updates to improve a global model.¹¹. This approach enhances data privacy, facilitates regulatory compliance, reduces bandwidth requirements, and allows for the inclusion of diverse, geographically distributed datasets. Despite these advantages, FL still faces several challenges, such as security vulnerabilities, increased communication overhead, and the presence of non-independent and identically distributed (non-IID) data across nodes¹². In real-world scenarios, non-IID data distributions are common due to variations in patient populations, imaging equipment, and data acquisition protocols. Such heterogeneity can hinder model convergence, reduce performance, and introduce instability or bias during training ^13,14. To address these issues, recent research has explored various strategies, including personalization techniques, adaptive aggregation schemes, and robust optimization methods¹⁵, aimed at improving model performance under non-IID conditions.

This study presents a novel multi-institutional FL effort for diagnosing COL6-RD using collagen VI immunofluorescence microscopy images from patient-derived fibroblast cultures. The model was trained on data from 90 COL6-RD patients from two geographically distinct sites across two countries. Specifically, this work describes a new approach to develop a model capable of further subclassifying collagen VI immunofluorescence microscopy images from dermal fibroblast cultures into the three primary pathogenic mechanism groups associated with COL6-RD: exon skipping, glycine substitution, and pseudoexon insertion. By leveraging FL, the model gained knowledge from a larger and more diverse dataset, thereby enhancing its diagnostic utility.

2. Results

2.1. Limited generalization of previous models to external cohorts

One of the standard laboratory techniques used to investigate suspected cases of COL6-RD is the analysis of collagen VI immunofluorescence microscopy images from dermal fibroblast cultures derived from the proband. This method allows qualitative assessment of collagen VI morphology and extracellular deposition by comparing patient samples with those from healthy controls.

Building on prior single-institution efforts, Bazaga et al.¹⁶ reported a two-class convolutional neural network (CNN) trained on z-stacked confocal images, while Frías et al.¹⁷ systematically compared classical and deep learning approaches—finding that pretrained deep features coupled with Support Vector Machine (SVM) outperformed alternatives (see also¹⁸). Although these results indicate strong within-site performance, both studies were evaluated exclusively on images from a single institution, leaving out-of-distribution robustness and cross-site generalizability untested.

To overcome these limitations, our study utilized two independently acquired image datasets from geographically distinct sites, i.e., the National Institutes of Health (NIH) dataset and the University College London (UCL) dataset. Each dataset was collected using different microscopy systems and imaging parameters, resulting in a heterogeneous and non-IID collection of images that better represents real-world variability.

To assess the generalizability of Bazaga et al.’s model, we first applied it to the NIH dataset, categorizing images into two classes: healthy control and patient. The model performance was assessed using the F1-score and Accuracy metrics (see Section 4.4 for metric definitions).

As shown in Table 1 and Figure 4, the model’s performance decreases substantially on this external dataset, achieving an F1-score of 0.69 and an Accuracy of 0.74, considerably lower than the originally reported values of 0.95 for both metrics. These results indicate that Bazaga’s model has only limited ability to generalize across cohorts with different imaging characteristics and institutional origins.

Table 1:

Bazaga et al. model’s performance comparison using different datasets.

Metric	Bazaga et al. dataset	NIH dataset
Mean F1-score (STD)	0.95 (—)	0.689 (0.07)
Mean Accuracy (STD)	0.95 (—)	0.735 (0.07)

Figure 4:

Mean metrics using Bazaga et al. model with different datasets.

2.2. A ML model for subclassifying collagen VI immunofluorescence microscopy images by COL6-RD pathogenic mechanism

Next, we developed an ML model capable of classifying patient collagen VI immunofluorescence images into one of the three pathogenic mechanism groups, thereby performing a four-class classification task (healthy control, glycine substitution, pseudoexon insertion, or exon skipping). The image dataset was defined as:

$$\mathbf{X}={\left[x_{s_1}^1,\dots ,x_{s_i}^m\right]}^T$$ (1)

where each element $x_{s_i}$ represents a collection of $m$ microscopy images from site $s_i$. The corresponding label vectors are denoted as: $\mathbf{Y}=\left[y_1,\dots ,y_i\right]$ where each $y_i\in \{1,2,3,4\}$, corresponding to the classes: $c=\{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{o}\mathrm{l},\mathrm{g}\mathrm{l}\mathrm{y}\mathrm{c}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{u}\mathrm{b}\mathrm{s}\mathrm{t}\mathrm{i}\mathrm{t}\mathrm{u}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n},\mathrm{p}\mathrm{s}\mathrm{e}\mathrm{u}\mathrm{d}\mathrm{o}\mathrm{e}\mathrm{x}\mathrm{o}\mathrm{n}\mathrm{i}\mathrm{n}\mathrm{s}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n},\mathrm{e}\mathrm{x}\mathrm{o}\mathrm{n}\mathrm{s}\mathrm{k}\mathrm{i}\mathrm{p}\mathrm{p}\mathrm{i}\mathrm{n}\mathrm{g}\}$.

We assessed the model’s performance locally in a non-federated configuration at the NIH site using F1-score and Accuracy as evaluation metrics. The model classified images into four categories: control or one of the three pathogenic mechanism groups. For comparison with the approach described by Bazaga et al., we reformulated this four-class problem into a binary classification task, where all pathogenic groups were combined into a single positive class and the control group was retained as the negative class. Under this formulation, the proposed model achieved an F1-score of 0.94 and an Accuracy of 0.92, demonstrating strong performance in distinguishing patients from controls (see Table 2 and Figure 5).

Table 2:

Classification performance comparison between the Bazaga et al. model and our approach for 10-fold cross-validation (NIH local model).

Metric	Bazaga et al. (2 classes)	Sherpa.ai (converted to 2 classes)
Mean F1-score (STD)	0.689 (0.07)	0.94 (0.06)
Mean Accuracy (STD)	0.735 (0.07)	0.92 (0.06)

Figure 5:

Mean metrics comparison between the Bazaga et al. model and our approach for 10-fold cross-validation (NIH local model).

2.3. FL enables enhanced pathogenic mechanism-based diagnosis of COL6-RD

Given the limited data availability, typical of rare diseases such as COL6-RD, we trained our model using FL, which enables multi-institutional training without sharing raw data using the NIH and UCL datasets as nodes (see Figure 6).

Figure 6:

Federated training implemented on the Sherpa.ai FL platform, including the nodes from NIH and UCL.

We employed Horizontal FL, a setup where both participating institutions share a common feature space but retain different patient datasets. Each institution (node) trained the local model on its private dataset. These locally learned model parameters were then aggregated using Federated Averaging (FedAvg)¹¹.

Each dataset remains locally stored at its respective institution. Due to differences in imaging protocols, equipment, and preprocessing pipelines, data across sites may vary in quality, resolution, and distribution. To ensure effective model training in this heterogeneous FL setting, we made the following assumptions: (i) Complete class coverage: Nodes complement each other, as some classes may be underrepresented at individual sites, ensuring that the global model benefits from a more balanced dataset. (ii) Standardized image format: All images are resized to a fixed resolution of 256 × 256 pixels and normalized to ensure compatibility with the convolutional neural network architecture. (iii) Domain heterogeneity: We account for inter-site variability in imaging modalities, equipment, patient demographics, and label distributions, introducing data drift typical in real-world, multi-center deployments. This exemplifies a heterogeneous FL scenario. (iv) Label reliability: Labels are annotated by domain experts and assumed to be reliable for supervised training. However, we acknowledge the possible presence of minor to moderate label noise.

For the federated classification task, we utilized the same EfficientNet-B0 feature extractor with global average pooling, as described in Section 2.2. To evaluate model performance, we assembled a hold-out set of 24 images, 20 from the NIH dataset and 4 from the UCL dataset, the latter limited by data availability. Ten images were selected based on expert criteria, while the remaining images were selected randomly to ensure a balanced class distribution. All splits were performed at the patient level (no patient’s images appear in both training and test).

We used the F1-score as the primary evaluation metric, given its robustness to class imbalance. Accuracy was also reported, consistent with previous FL studies. The model was trained and evaluated under two scenarios: (i) at the single-node level (using only individual datasets, we refer to it in further as the single-node model) and (ii) in the federated setting (see Table 3 and Figure 7).

Table 3:

Classification performance for UCL and NIH over the test subset — federated and local baseline models.

Metric	Single-node (NIH only)	Single-node (UCL only)	Federated model (Sherpa.ai)
Mean F1-score (STD)	0.747 (0.024)	0.582 (0.037)	0.820 (0.032)
Mean Accuracy (STD)	0.754 (0.022)	0.567 (0.038)	0.825 (0.031)

Results shown as means and standard deviations over 10 runs.

Figure 7:

Mean metrics across ten random trials for single-node (NIH-UCL) and federated model.

As summarized in Table 3, federated training yielded a mean F1-score of 0.820 ± 0.032 and Accuracy of 0.825 ± 0.031, outperforming both single-node models (NIH-only: F1-score 0.747 ± 0.024, Accuracy 0.754 ± 0.022; UCL-only: F1-score 0.582 ± 0.037, Accuracy 0.567 ± 0.038). In absolute terms, FL improved F1-score by +0.073 over NIH and +0.238 over UCL (≈ +9.8% and ≈ +40.9% relative, respectively), with corresponding Accuracy gains of +0.071 (≈ +9.4%) and +0.258 (≈ +45.5%).

Figure 8 shows confusion matrices for single-node performance: (Left) for the NIH dataset and (Right) for the UCL dataset. The first row in each matrix reflects the classification of control images, with the remaining rows corresponding to patient subgroups. In the NIH dataset (Figure 8 Left), 18 out of 24 test images were correctly classified. In contrast, performance on the UCL dataset (Figure 8 Right) was lower, with 14 correct classifications and 3 false positives in the control class.

Figure 8:

Confusion matrices of the model trained with the NIH dataset (Left) or the UCL dataset (Right), evaluated on the test dataset (i.e., before participating in FL). Cell counts reflect images.

Figure 9 presents the confusion matrix of the federated model, which outperforms both the NIH single-node model and the UCL single-node results (Figure 8). These results demonstrate the benefits of federated collaboration, highlighting improved generalization and robustness across institutions.

Figure 9:

Confusion matrix of the FL model trained with both NIH- and UCL datasets, evaluated on the test dataset (i.e., after participating in FL). Cell counts reflect images.

The confusion matrices in Figures 8 and 9 further illustrate the effects highlighted in Table 3: while the NIH model maintained stable performance, the UCL site, constrained by a smaller and more heterogeneous dataset, showed marked reductions in misclassifications, particularly fewer false positives in the control class after federation, highlighting improved cross-site generalization from collaborative training.

3. Discussion

In this study, we report a novel real-world application of FL for COL6-RD, using data from 90 patients across two geographically distinct sites in two countries. Our objective was to develop a robust, generalizable ML model capable of accurately diagnosing COL6-RD and classifying its underlying pathogenic mechanisms through the analysis of collagen VI matrix immunofluorescence images. In contrast to the public model developed by Bazaga et al.¹⁶, which was limited to distinguishing between control and patient images, our approach further stratifies patient images into three distinct pathogenic subgroups. By leveraging the Sherpa.ai privacy-preserving FL platform, we (i) accessed a larger and more diverse dataset, crucial in the context of rare diseases, and (ii) trained our model in a decentralized manner, thereby improving its generalizability to out-of-sample data.

Although previous approaches constituted a substantial improvement on collagen VI matrix assessment within institution baselines, they both left cross-site robustness untested. Bazaga et al.¹⁶ trained a two-class CNN on z-stacked confocal microscopy images with integrated data augmentation, reporting an Accuracy and F1-score of 0.95, and Frías et al.¹⁷ showed that pretrained deep features with a linear SVM achieved a biopsy-level AUC of 0.996 (see also¹⁸). However, these single-institution studies present several limitations that might restrict their broader applicability: (i) limited sample sizes inherent to COL6-RD, yielding small training sets despite data augmentation; (ii) restricted data diversity, as all images originated from a single institution, potentially introducing dataset bias and limiting the model’s generalizability to external cohorts; and (iii) imaging system dependency. The dataset was acquired using a specific confocal microscopy system with fixed parameters. While confocal microscopy is widely used, not all laboratories have access to the same equipment or standardized imaging protocols. Moreover, many laboratories primarily rely on wide-field microscopy for rapid, qualitative evaluation of collagen VI matrix, which may further affect the model’s performance when applied across diverse settings. These constraints motivated our federated design, which explicitly targets generalization across non-IID, privacy-restricted datasets from different sites.

Over the past decade, ML has shown considerable promise in rare disease research, particularly in the study of rare neuromuscular disorders^16,22–28. However, the development of effective ML models is often hindered by the limited availability of high-quality patient data, an inherent limitation in rare diseases. This limitation highlights the importance of multi-institutional collaboration and data sharing to aggregate sufficient training data. However, such efforts introduce critical concerns about patient privacy and data security. FL addresses this challenge by enabling collaborative model training without requiring the exchange of raw patient data. Our study demonstrates that FL can facilitate the development of robust and generalizable ML models while preserving data privacy, even in highly data-constrained settings.

The performance of FL systems is highly influenced by the characteristics of the participating institutions. For rare diseases, where data scarcity is pronounced, site selection is limited, and heterogeneity in data quality, quantity, and label distribution across institutions is often unavoidable. In our case, significant non-IID data were present across institutions. While such variability can introduce biases and reduce global model performance, it also reflects the real-world constraints of rare diseases research. Therefore, we incorporated all available data sources in the federation, accepting the resulting heterogeneity as a necessary compromise to maximize data utilization. This highlights the need for FL frameworks that are robust to domain variability and can accommodate the complexity of low-data systems.

Previous studies have investigated FL in data-constrained scenarios, particularly within medical imaging applications²⁹. To address this limitation and improve model performance, various strategies have been used, including data augmentation, transfer learning, and feature extraction using pretrained models.^30,31. In our study, we adopted a feature extraction approach using models pretrained on the ImageNet dataset, a widely used resource comprising over 14 million labeled images. Transfer learning from ImageNet enabled us to train efficient models with fewer parameters, an approach well-suited to the FL paradigm and small, specialized datasets.

Somewhat unexpectedly, the resulting model also accurately classified collagen VI immunofluorescence patient images according to their underlying pathogenic mechanism. Each institution participating in the FL collaboration benefited differently: For the NIH, the primary benefit was increased model robustness. Although the model weights changed substantially during FL training, the confusion matrices before and after FL (Figures 8 and 9, respectively) remained very similar. This consistency suggests that the model retained stable performance despite significant parameter updates. At UCL, we observed substantial improvements in predictive performance, particularly in the accurate identification of control images, which the local model had previously misclassified.

Despite these advances, our study has several limitations. First, the dataset was small, reflecting the challenges inherent in rare diseases research. Second, only two sites participated in the FL collaboration, which constrained the diversity and volume of training data. Expanding the federation to include more institutions would enhance data volume and diversity, key factors in improving model performance. Third, considerable heterogeneity was present in the imaging data, including differences in acquisition protocols, equipment, and label distributions. Standardizing image acquisition procedures across sites would help reduce this heterogeneity and enhance model performance. Fourth, the integration of additional data modalities such as magnetic resonance imaging (MRI) images, clinical data, or genomic information warrants further investigation to enhance diagnostic accuracy and model interpretability.

Although FL enables privacy-preserving, multi-institutional collaboration, it is not without risks. The distribution of model parameters and updates across participating sites may introduce vulnerabilities to attacks such as membership inference, gradient leakage, and model inversion.³² Ensuring privacy and security in FL remains an active area of research. Moreover, FL requires each participating site to have sufficient computational infrastructure and reliable communication channels. A privacy-preserving FL platform addresses these challenges, offering a user-friendly interface for coordinated and secure model training across multiple sites.

In conclusion, we have demonstrated the feasibility and utility of an FL platform approach to develop a robust and generalizable ML model capable of classifying collagen VI immunofluorescence images according to their underlying pathogenic mechanisms. This work supports the integration of ML into the diagnostic pipeline for COL6-RD, offering potential support for clinicians, geneticists, and researchers. Specifically, the model may (i) aid in interpreting variants of unknown significance, (ii) guide the prioritization of sequencing strategies to identify novel pathogenic variants, thereby reducing diagnostic time and cost, and (iii) support the development and evaluation of targeted therapies by quantifying their impact on the collagen VI matrix.

In future work, expanding the federated network to include additional institutions will be essential to enhance the model’s robustness and generalizability across broader patient populations and imaging settings (see Figure 10). We will prioritize partnerships with international groups and clinical centers that actively study collagen VI–related dystrophies, such as those in France, Italy, and Japan. As participating institutions collect more patient-derived samples, we aim to increase both the quantity and quality of imaging data per node, thereby improving intra-node diversity and overall model performance.

Figure 10:

Global map illustrating current and prospective institutions joining the federated learning initiative to enhance model generalizability for collagen VI–related dystrophies.

4. Materials and methods

4.1. Data acquisition

The data used in this study describe patient populations with a confirmed diagnosis of COL6-RD, specifically displaying the collagen VI matrix features, investigated by collagen VI immunofluorescence analysis in dermal fibroblast cultures³³. Detailed genetic and demographic characteristics of the study population are provided in Tables S1, S2, and S3.

Dermal fibroblast cultures were established from skin biopsies obtained based on standard operating procedures (12-N-0095, approved by the institutional review board of the National Institute of Neurological Disorders and Stroke (NINDS), and 06/Q0406/33, approved by the institutional review board of the University College London (UCL). Informed consent was obtained from each participant.

Primary dermal fibroblasts were cultured in Dulbecco’s modified Eagle’s medium (DMEM) supplemented with 10% fetal bovine serum (FBS) and 1% penicillin/streptomycin at 37°C under 5% CO₂. For immunostaining analysis, primary dermal fibroblasts were seeded into 8-chamber tissue culture slides or on collagen-precoated glass coverslips (Corning) in a six-well plate. The following day, the culture medium was replaced with DMEM supplemented with 10% FBS and 50 μg/mL of L-ascorbic acid (Wako Chemicals USA, Richmond, VA). Every 48–72h, the medium was again replaced and supplemented with L-ascorbic acid. After 3–7 days of L-ascorbic supplementation, cells were fixed with 4% paraformaldehyde for 10 min at room temperature and matrix immunostaining was performed following the protocol and using antibodies previously described³⁴. Alternatively, a collagen VI mouse monoclonal antibody (MAB3303, Chemicon) was also used.

Images were acquired from multiple microscope systems: a TCS SP5 II system, a Stellaris 8 system (Leica Microsystems, Buffalo Grove, IL), and a Nikon Eclipse Ti microscope (Nikon Instruments, Melville, NY) equipped with an sCMOS pco.edge 4.2 LT camera (Excelitas PCO GmbH, Kelheim, Germany), LSM710 Zeiss Confocal microscope and Leica DMR fluorescence microscope. For confocal images, the pinhole was set to 1 Airy unit, and stack images were acquired using an x40 objective with system-optimized settings. Stacks were then merged. Non-stacked wide-field microscopy images were acquired with x20 or x40 objectives.

4.2. Data processing

To develop the model, we assembled two image datasets (i.e. NIH and UCL datasets) comprising 331 images from 90 patients collected at two international sites, with varying contributing images across sites (Table 4).

Table 4:

Image distribution across sites and COL6-RD pathogenic mechanism groups.

Site ID	Pathogenic mechanism’s group	# images	% images	Total # images
NIH	Control	84	28.0%	300
	Exon skipping	71	23.7%
	Gly substitution	94	31.3%
	Pseudoexon insertion	51	17.0%
UCL	Control	7	22.6%	31
	Exon skipping	11	35.5%
	Gly substitution	5	16.1%
	Pseudoexon insertion	8	25.8%

For NIH-only experiments reported in Section 2.2, we employed 10-fold cross-validation. For federated experiments reported in Section 2.3, evaluation was performed on a fixed hold-out test set of 24 images (20 NIH, 4 UCL), with the remaining images used for training. Given the limited sample size, a data augmentation and preprocessing pipeline was applied to the training set to increase data diversity and improve model robustness. Data augmentation techniques included (i) 45-degree rotations to improve orientation invariance, (ii) horizontal flipping to simulate positional variability, and (iii) brightness adjustments to enhance contrast between collagen VI and background. For brightness augmentation, images were converted to HSV (Hue, Saturation, Value) color space, and the Value (V) channel was scaled by factors 1.25 and 1.5 before reconversion to RGB (Red, Green, Blue). Subsequently, the images were normalized by scaling the pixel values to the [0, 1] range and resized to 256 × 256 pixels to conform to the model’s input specifications. For the validation and test data, no preprocessing was performed besides resizing images to 256 × 256 and normalizing them to the [0, 1] range.

4.3. ML approach

We formulated the prediction task as a supervised ML problem, aiming to approximate an unknown function that maps inputs to outputs using a finite set of labeled examples. This was framed as an empirical risk minimization problem, with an added regularization term to prevent overfitting. Model training involves optimizing a parameterized function class, such as neural networks or ensemble methods, based on this objective.

To enable collaborative learning across multiple data holders without requiring data centralization, we extended this framework to a Horizontal FL setting (see the supplementary material). In this setup, each participating node trains a local model on its own data subset and contributes to the global model through iterative aggregation of locally optimized updates. This approach preserves data privacy while aiming to approximate the performance of a fully centralized model.

A detailed mathematical formulation of both the centralized and FL setups, including conditions for optimization and existence of solutions, is provided in the supplementary material (Sections B and C).

All experiments were conducted on a machine equipped with 1TB of disk space, an Intel Core i7–7700 4-core CPU at 3.60 GHz, 64 GB of RAM, and running Ubuntu 24.04 as the operating system. We used Python 3.11, TensorFlow 2.14.0, and scikit-learn 1.6.1 for feature extraction and model training.

In the idealized centralized scenario, where all data is pooled in one location (typically prohibited by data protection laws such as the GDPR¹⁰), model performance serves as a benchmark for comparison. Conversely, in the fully private single-node scenario, where neither data nor model parameters are shared, performance typically reflects a lower limit. The federated setting thus aims to achieve performance within this spectrum, striking a balance between utility and privacy.

4.4. Metrics

In healthcare and medical imaging, performance metrics such as Accuracy, macro-averaged F1-score, and class-specific F1-scores are widely used to evaluate multiclass classification models^35,36. While Accuracy provides an overall measure of the proportion of correctly predicted samples across all classes, the F1-score is particularly valuable in the presence of class imbalance, as it balances precision and recall for each class.

For a given class $c$, precision and recall are defined as:

$${\text{Precision}}_c=\frac{{\mathrm{T}\mathrm{P}}_c}{{\mathrm{T}\mathrm{P}}_c+{\mathrm{F}\mathrm{P}}_c},{\text{Recall}}_c=\frac{{\mathrm{T}\mathrm{P}}_c}{{\mathrm{T}\mathrm{P}}_c+{\mathrm{F}\mathrm{N}}_c},$$ (2)

where ${\mathrm{T}\mathrm{P}}_c,{\mathrm{F}\mathrm{P}}_c$, and ${\mathrm{F}\mathrm{N}}_c$ denote the number of true positives, false positives, and false negatives for class $c$, respectively. The F1-score for class $c$ is the harmonic mean of its precision and recall:

$$\mathrm{F}{1}_c=2\cdot \frac{{\text{Precision}}_c\cdot {\text{Recall}}_c}{{\text{Precision}}_c+{\text{Recall}}_c}.$$ (3)

To provide a single performance indicator over all $K$ classes, the macro-averaged F1-score is computed as the arithmetic mean of the class-wise F1-scores:

$$\mathrm{F}{1}_{\mathrm{m}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o}}=\frac{1}{K}\sum _{c=1}^K\mathrm{F}{1}_c.$$ (4)

The overall Accuracy of the model is calculated as:

$$\text{Accuracy}=\frac{1}{N}\sum _{i=1}^N\mathbb{1}\left({\hat{y}}_i=y_i\right)$$ (5)

where $N$ is the total number of samples.

This evaluation framework ensures that performance is not dominated by the majority classes, providing a balanced view of classification quality across all disease phenotypes.

Figure 1:

Overview of the Sherpa.ai FL setup showing individual institutional accuracies and the resulting global federated model performance.

5. Data availability

The datasets used in this study are not made available due to restrictions imposed by acquiring sites.

6. Code availability

The code used in this study is not publicly available and cannot be shared due to institutional policies.