The global reach and rapid transmission of viral pandemics call for diagnostic solutions that are not only accurate and fast but also respectful of patient privacy. While deep learning models have shown high predictive capabilities, they often lack critical features like explainability, data confidentiality, and support for collaborative analysis across different medical institutions (1, 4, 11). Furthermore, most models are opaque offering little to no insight into their decision-making which limits their clinical utility and acceptance (12, 26–28). There is a growing need for a robust, federated, and interpretable AI framework capable of handling varied and non-uniform healthcare data from multiple sources (3, 8).

This research aims to:

Build a CNN-based federated learning model for classifying COVID-19, Normal, and Viral Pneumonia cases using chest x-ray images (5, 7).

The proposed U-FDL-PPE framework offers several key contributions:

Privacy-First Design: Uses federated learning to keep data secure and localized while still enabling shared model development (2, 10).

Figure 1 illustrates how traditional diagnostic systems fall short in managing viral disease outbreaks and why there's an urgent demand for smarter, faster solutions. U-FDL-PPE steps in as a transformative framework, offering decentralized AI with built-in privacy and interpretability to meet real-world healthcare needs.

Federated learning is increasingly recognized as a game-changer in enabling collaborative model development across healthcare institutions—without compromising patient data privacy. Sharma and Guleria (1) provide a thorough review of FL in healthcare, emphasizing the need to move from theoretical frameworks to real-world clinical deployment. Zwiers et al. (3) focus on applying FL to infectious disease research, highlighting challenges like inconsistent data and compliance with privacy standards across organizations. Similarly, Teo et al. (9) analyze over 600 FL-related studies in clinical applications, noting rapid adoption alongside unresolved issues like standardization and interoperability. In terms of practical use, Rawas and Samala (2) introduced an Edge-Assisted Federated Learning (EAFL) model that incorporates differential privacy and supports real-time diagnosis—tested on COVID-19 Radiography data. Their work showcases FL's effectiveness in time-sensitive, distributed environments. Moshawrab et al. (10) further investigate FL in disease prediction, with a focus on maturing these algorithms for use in smart health devices.

Deep learning has proven highly effective for diagnosing respiratory illnesses such as COVID-19. Mondal et al. (4) explored the combined use of DL and FL for COVID-19 screening, calling for a unified framework that leverages the strengths of both. Kareem et al. (5) built a federated CNN model for pneumonia detection from x-ray images using established architectures like ResNet and VGG19. Ajagbe and Adigun (6) offered a broader view, covering multiple pandemic-related datasets and noting limitations such as class imbalance and noisy labels. In another notable research, Gopal et al. (7) developed a robust CNN-based model with transfer learning, which incorporated explainability tools to address the black-box nature of DL models. Their work reinforced the importance of transparency, particularly in clinical settings. Moreover, using pre-trained models not only improves prediction accuracy but also reduces training time an important factor in federated and resource-constrained environments.

One of the most pressing challenges in deploying AI in healthcare is ensuring data privacy. Centralized approaches often involve moving sensitive patient data to a single server, increasing the risk of breaches. Rawas and Samala (2) tackle this by embedding differential privacy into their EAFL model. Similarly, Badidi (8) presents an Edge AI approach tailored for real-time health monitoring using IoT and EHRs while maintaining privacy. Moshawrab et al. (10) highlight the evolving nature of FL algorithms and stress the need for improvement in handling diverse and non-uniform data distributions. Sharma et al. (1) also point out that successful implementation of FL in real-world clinical environments demands more adaptive and context-aware protocol designs. Collectively, these studies underline the importance of building secure, privacy-respecting AI systems that can handle the realities of clinical data (24, 25).

In medical AI, explainability isn't just a luxury it's essential. Tools like Grad-CAM (Gradient-weighted Class Activation Mapping) help visualize which parts of an image the model is focusing on when making predictions, thus providing critical insight for clinicians. Mondal et al. (4) and Singh and Kolekar (14) both emphasize how Grad-CAM improves the interpretability of DL models used for COVID-19 detection from chest x-rays. Gopal et al. (7) effectively incorporated Grad-CAM into their architecture, allowing for clinician-friendly visual explanations. Abdul Gafoor et al. (22) found that these heatmaps significantly enhance radiologists’ understanding of how and why a model reaches its conclusions. As AI systems evolve to support ambient and human-centric healthcare applications, integrating explainability tools like Grad-CAM becomes increasingly vital. Collectively, these studies provide strong support for the development of U-FDL-PPE a federated, secure, and explainable deep learning framework. This model directly addresses the gaps and challenges identified across four core areas: collaborative model training, accurate viral disease detection, privacy preservation, and interpretability in diagnostic decision-making.

At the heart of the U-FDL-PPE framework is a decentralized deep learning architecture that simulates training across three virtual healthcare institutions. Each institution or client is assigned data from one of the three categories: COVID-19, Normal, and Viral Pneumonia. These clients train their respective convolutional neural networks (CNNs) locally, ensuring that raw data remains confined to its origin. This approach mimics how real hospitals store and process patient information independently, allowing for privacy-preserving AI development. The model leverages MobileNetV2 as its backbone due to its computational efficiency and suitability for deployment on devices with limited resources. After each training round, model weights from all clients are aggregated using Federated Averaging (FedAvg) to build a shared global model without ever exchanging sensitive data. To ensure model predictions are interpretable, Grad-CAM is integrated, offering visual insights into the regions influencing each prediction, which is especially valuable in medical diagnostics.

The Figure 2 shows how the U-FDL-PPE system works by connecting hospitals through federated learning, where each site trains its own model without sharing patient data. A shared global model is built securely, and explainable AI like Grad-CAM helps doctors understand the predictions with clear, visual cues. The federated deep learning system operates over K=3 clients, each training a local MobileNetV2 CNN model on private data subsets Dk, where k∈{1,2,3} represents COVID-19, Normal, and Viral Pneumonia categories respectively.

The local client model parameters at training round t are denoted as θk(t), is represented in the Equation (1) while the global aggregated model parameters at t are θg(t).

Local training proceeds via stochastic gradient descent (SGD) or its adaptive variants. The update rule per client reads:θk(t+1)=θk(t)−η∇L(fθk(t),Dk) (1) where η is the learning rate, and L is the categorical cross-entropy loss.

Aggregation of client models to update global parameters uses Federated Averaging is shown in the Equation (2):θg(t+1)=∑k=1K|Dk|∑j=1K|Dj|θk(t+1) (2) This weighted average accounts for heterogeneous data sizes ensuring fair model contribution.

For privacy preservation, additive Gaussian noise is introduced in local updates as shown in Equation (3):θ~k(t+1)=θk(t+1)+N(0,σ2I) (3) where σ controls noise magnitude to satisfy differential privacy guarantees.

The overall training objective minimized through federated updates is expressed as:minθ∑k=1K|Dk|∑j=1K|Dj|E(x,y)∼Dk[L(fθ(x),y)] (4) where fθ is the global model parameterized by θ is represented in Equation (4).

The training and evaluation were performed using the publicly available COVID-19 Radiography Database from Kaggle (38, 39, 41). This dataset comprises a wide variety of labeled chest x-ray images across three categories: COVID-19 infections, Normal lung images, and Viral Pneumonia. The diversity and volume of data make it ideal for training a robust classification model. To maintain consistency across training inputs and ensure compatibility with MobileNetV2, all images were resized to 224 × 224 pixels. Each image was mapped accurately to its category using the accompanying metadata files (COVID.metadata.xlsx, Normal.metadata.xlsx, and Viral Pneumonia.metadata.xlsx). These metadata sources were crucial in validating class labels and organizing data into simulated clients, effectively representing distinct hospital environments. https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database (41). In the U-FDL-PPE framework, F1-score and AUC were calculated using a macro-averaged one-vs.-rest approach to fairly assess performance across the three categories COVID-19, Normal, and Viral Pneumonia. This method ensures that each class contributes equally, allowing a balanced evaluation even when class sizes differ, which is often the case in medical imaging datasets.

The COVID-19 Radiography Database consists of images X∈R224×224×3 labelled over 3 classes C={0,1,2} representing Normal, COVID-19 and Viral Pneumonia.

Label assignment function is shown in Equation (5):y=δ(c),c∈C,δ:C→{0,1}3(one-hot) (5) Each client k is assigned a subset Dk={(xi,yi)}i=1nk is calculated in the Equation (6) such that:⋃k=1KDk=D,Di∩Dj=∅fori≠j (6) Data augmentation T is applied as in Equation (7) to enhance generalization:x^=T(x),x^∈R224×224×3 (7) Normalization standardizes input as:x′=x^−μσ (8) with mean μ and standard deviation σ computed in Equation (8) over training images.

This algorithm describes how multiple hospitals collaboratively train a global machine learning model without sharing sensitive patient data. Each hospital locally trains the provided global model on its own dataset and then shares the updated model parameters—not the data—with a central server. The server aggregates these local updates into a new global model by weighting each hospital's contribution according to its data size. This process iteratively improves the global model while preserving data privacy.

Step-by-Step Example:

Suppose three hospitals have 1,000, 500, and 1,500 x-ray images respectively. Initially, a global model is initialized and sent to all hospitals. Each hospital trains the model on its local images for several epochs and obtains updated model parameters. Hospital 1's update is weighted by 1,000/(1,000 + 500 + 1,500) = 0.333, hospital 2's by 0.167, and hospital 3's by 0.5. The server combines these weighted parameters to form the updated global model, which is again sent back for further training rounds.

To protect patient privacy, hospitals add carefully calibrated Gaussian noise to their locally updated model parameters before sending them to the central server. The noise masks individual data contributions, guaranteeing differential privacy by making it statistically difficult to infer sensitive information. The noise magnitude is decided based on a privacy budget that balances privacy and model accuracy.

Step-by-Step Example:

A hospital's local model has 20 parameters after training. It computes the maximum expected change if a patient's data were altered. Then, it generates a vector of 20 independent Gaussian noise values with zero mean and variance aligned to this change and privacy budget. Adding this noise vector to its parameters yields privacy-protected updates. These updated parameters are transmitted securely to the server where they contribute to the federated aggregation.

The experiments were conducted on a machine configured with 16 GB RAM and an NVIDIA GPU, providing ample processing power for federated learning tasks. The Flower framework was chosen to simulate federated learning, as it supports scalable and flexible implementation of FL workflows. Other libraries used include TensorFlow for constructing and training neural networks and OpenCV for image processing and manipulation. The model underwent training over five federated learning rounds. During each round, the three clients trained their CNN models independently on their respective datasets. After local training, their model updates were sent to a central server where they were merged using the FedAvg algorithm. The performance of the global model was evaluated using metrics such as accuracy, F1-score, and AUC after every round. This simulation closely reflects a decentralized healthcare setting and demonstrates the system's ability to preserve data privacy while ensuring reliable and interpretable diagnostics.

Figure 3 presents a clear overview of the proposed U-FDL-PPE-6X framework, which is built to support timely and accurate prediction of viral diseases in a secure, decentralized environment. The system is structured around six key components working together seamlessly: (1) the Federated Learning Engine (FLE) coordinates local model training across different hospital clients without sharing raw patient data; (2) each client uses a MobileNetV2-based CNN, chosen for its lightweight design and suitability for edge devices; (3) Grad-CAM is integrated to enhance interpretability by highlighting the critical image regions that influence the model's decisions; (4) the Data Preprocessing Unit resizes all input images to 224 × 224 pixels and leverages metadata for accurate classification; (5) the Simulation and Communication Layer, implemented using the Flower framework, handles client-server interactions and federated weight updates while keeping data localized; and (6) the Evaluation and Visualization Engine tracks key performance indicators like accuracy, F1-score, AUC, and provides visual outputs such as confusion matrices and Grad-CAM heatmaps. Together, these components deliver a privacy-compliant, scalable, and explainable AI solution suitable for real-world deployment using the COVID-19 Radiography dataset. As the U-FDL-PPE framework was tested in a simulated federated setup, its real-world deployment may encounter additional challenges such as inconsistent network connectivity, variations in imaging protocols, and differences in data management policies between institutions. Future work aims to implement the framework across actual hospital networks to assess its stability and adaptability under real clinical conditions, including diverse communication environments and regulatory practices.

This algorithm generates a visual explanation indicating image regions most influential in the model's classification decision. It calculates the gradient of the predicted class score with respect to activations in the last convolutional layer. These gradients determine weights for each feature map, which are combined into a heatmap and overlaid onto the input image. This aids clinicians in verifying and trusting model predictions by visualizing key clinical features.

Step-by-Step Example:

Given a chest x-ray predicted to have COVID-19, the model extracts feature maps from its deepest convolutional layer. Gradients of the COVID-19 class score w.r.t. these maps are computed and averaged to form weights. The weighted sum of feature maps creates a heatmap after applying ReLU to retain only positive values. The heatmap is resized and overlaid on the x-ray, distinctly highlighting areas such as lung infections that influenced the model's diagnosis.

Figure 4 is said to illustrate the entire flow of the federated learning system applied to healthcare, starting from local healthcare institutions where patient data is preprocessed and models are trained locally on sensitive medical images. It is described how, after local training, differentially private noise is added to model updates to protect patient confidentiality before these encrypted updates are securely transmitted to a central server. The central server aggregates these updates into a global model which is then redistributed to all client sites. Additionally, the inclusion of explainability via Grad-CAM heatmaps is emphasized, which visually highlights the important regions in medical images influencing model predictions, assisting clinicians and radiologists in interpretation. This iterative cycle continues until the model achieves convergence, ensuring accurate, private, and interpretable AI-supported diagnostics. The principal end users identified include doctors and healthcare professionals, medical researchers, and ultimately patients, ensuring data privacy and clinical transparency throughout.

To measure how effectively the proposed model performed, several core evaluation metrics were recorded over five federated learning rounds. The model reached an overall accuracy of 88%, indicating strong performance in correctly classifying chest x-ray images into the three categories: COVID-19, Normal, and Viral Pneumonia. The F1 score of 89.66% reflects a healthy balance between precision and recall, which is essential in medical diagnosis to minimize both false positives and false negatives. The model's ability to distinguish between the three classes was also assessed using the multi-class AUC, which was recorded at 0.5192. While this score suggests moderate classification power across different threshold levels, it remains acceptable considering the inherent complexity and imbalance in medical datasets. The training also maintained a low average loss of 0.30, showing stable learning with minimal error accumulation. Training was conducted over five federated epochs, where each simulated client trained on its local dataset. The average execution time per round was approximately 0.32 s, demonstrating the model's computational efficiency especially important for real-time applications. These results collectively show that U-FDL-PPE is capable of delivering fast, reliable, and accurate diagnostics in decentralized healthcare environments.

The confusion matrix provides a granular look at how well the model performed in each class. A strong diagonal pattern in the matrix suggests that predictions for COVID-19, Normal, and Viral Pneumonia were generally correct, with minimal confusion between classes. This is especially critical in clinical settings where accurate triaging can influence treatment paths and isolation measures. The model demonstrated strong sensitivity in detecting COVID-19 cases, which were rarely misclassified as Normal or Pneumonia. Such high recall is essential during pandemics to prevent false negatives and ensure timely patient care. Additionally, the model successfully distinguished between Viral Pneumonia and COVID-19, which often show overlapping features in imaging. Some misclassifications occurred between Normal and Viral Pneumonia a known challenge in chest x-ray diagnostics due to subtle visual similarities. However, the confusion remained limited, and overall, the matrix supports the model's ability to generalize well across real-world-like, decentralized datasets.

To provide transparency into how the model makes decisions, Grad-CAM visualizations were used to generate heatmaps over sample predictions. These highlighted areas of the lungs that contributed most to the classification. For instance, the model focused on lung opacities and consolidations hallmarks of infection confirming that it learned relevant clinical features. In COVID-19 cases, Grad-CAM highlighted the bilateral and peripheral regions typically associated with the disease. For Viral Pneumonia, the heatmaps revealed localized intensity near infected regions, while Normal images showed dispersed or low-activation areas, consistent with healthy scans. These visual cues not only prove that the model is looking at the right anatomical regions but also offer clinicians a layer of interpretability. By combining predictions with these explainability overlays, the system promotes greater confidence among healthcare professionals, which is key to real-world adoption.

Figure 5 shows a clear upward trend in model accuracy as training progresses through five federated learning rounds. This improvement highlights how U-FDL-PPE continuously refines its predictions by effectively combining insights from multiple client nodes, leading to better generalization across distributed healthcare data.

Figure 6 shows a steady decrease in loss values throughout the five training rounds, highlighting the model's ability to converge efficiently. This consistent drop confirms that the U-FDL-PPE framework reduces prediction errors while effectively learning from decentralized client data.

Figure 7 highlights that the chest x-ray images are evenly distributed among the categories—COVID-19, Normal, and Viral Pneumonia. This balanced representation helps the U-FDL-PPE model learn each class fairly, reducing bias and enhancing the reliability of its predictions.

Figure 8 highlights an equal spread of chest x-ray images across the three diagnostic groups COVID-19, Normal, and Viral Pneumonia. This balanced distribution helps the U-FDL-PPE model learn fairly from each class, improving its ability to generalize and reducing the chances of biased or skewed predictions.

U-FDL-PPE was designed to solve three fundamental issues in AI for healthcare: safeguarding patient data, improving model interpretability, and enabling training across fragmented datasets. The experimental outcomes show that the framework successfully tackles these by combining federated learning with explainable AI and a modular architecture. Through local model training, hospitals can build collaborative models without ever sharing sensitive data. Additionally, the inclusion of Grad-CAM provides transparency into how predictions are made, helping clinicians trust and verify the model's decisions. This significantly improves the acceptance of AI in clinical environments where understanding model logic is just as important as accuracy.

U-FDL-PPE adheres to global data privacy regulations like GDPR and HIPAA by ensuring that patient data never leaves the originating institution. Instead of sharing raw images, only encrypted model updates are exchanged, which minimizes security risks and supports ethical AI deployment. By leveraging the Flower framework for secure communication and federated averaging, the model maintains strong performance while complying with privacy mandates. This makes U-FDL-PPE a viable option for integration into real healthcare infrastructures where regulatory compliance is non-negotiable.

A defining strength of the framework is its use of Grad-CAM, which generates intuitive heatmaps that show which parts of an x-ray influenced the AI's prediction. This allows clinicians to visually verify whether the model is focusing on medically relevant features. In testing, the Grad-CAM results corresponded well with known clinical indicators highlighting areas of infection in COVID-19 cases while showing minimal activation in normal scans. This transparency builds confidence in the system and supports decision-making in high-stakes environments like hospitals.

Medical data can vary significantly across hospitals due to different imaging equipment, protocols, and demographics. The U-FDL-PPE framework is designed to manage such non-IID (non-identical and independently distributed) data across clients. During evaluation, the model was able to generalize well despite differences among datasets for COVID-19, Normal, and Pneumonia categories. Even with this variability, the model maintained high performance, achieving an 88% accuracy and a strong F1-score. These results suggest that U-FDL-PPE is robust enough to be deployed across various real-world settings without requiring strict data uniformity.

By using MobileNetV2, a lightweight CNN model, the framework is optimized for environments with limited computing power such as rural clinics or small hospitals without GPUs. The low inference time (∼0.32 s per round) proves that U-FDL-PPE can perform efficiently even with modest hardware. This lightweight yet powerful design allows the framework to be rolled out in diverse healthcare settings, not just advanced hospitals. Its adaptability makes it especially useful for bringing AI-driven diagnostics to underserved areas.

Although U-FDL-PPE performed well in simulations, it was not tested in a live federated hospital network. Real-world deployment introduces additional variables such as unstable internet connections, different annotation practices, and complex administrative policies that were beyond the scope of this research. Additionally, the model's AUC score (0.5192) suggests moderate differentiation among classes and indicates potential for improvement. Future versions could explore attention-based networks, ensemble learning strategies, or vertical federated learning to enhance classification performance and robustness across different data types and institutions.

The U-FDL-PPE framework marks a clear evolution from earlier federated and centralized deep learning models used in healthcare diagnostics. Previous works, such as those by Sharma et al. (1), Zwiers et al. (3), and Teo et al. (9), have largely been limited to literature-based analyses or conceptual discussions without real-world deployment or validation. While some studies, including Kareem et al. (5) and Rawas & Samala (2), have ventured into practical implementations—either with centralized CNNs or edge-assisted federated models their focus is often narrow, lacking broad applicability or interpretability features. Others, like Badidi (8), do not integrate explainable AI at all and fall short in meeting modern privacy and compliance standards for healthcare. In contrast, U-FDL-PPE offers a comprehensive and forward-looking approach by combining privacy preservation through GDPR-aligned federated learning, interpretability using Grad-CAM heatmaps, and robust performance metrics, achieving 88% accuracy and an F1 score of 89.66%. Unlike earlier models that can be complex or rigid, it leverages MobileNetV2 for efficient processing, making it suitable even for resource-constrained environments. The framework's modular design, depicted in Figure 1, ensures transparency across all stages from preprocessing to prediction and delivers visual outputs like confusion matrices and class activation maps for clinical confidence. These enhancements make U-FDL-PPE a scalable, ethical, and practical solution for modern federated healthcare deployments.

Table 1 highlights the measurable advantages of the proposed U-FDL-PPE framework over traditional systems. With an accuracy of 88% compared to 82.5% and an F1 score of 89.66% vs. 80%, U-FDL-PPE demonstrates stronger and more consistent classification performance. The AUC value also shows a slight improvement (0.5192 over 0.48), indicating better ability to distinguish between multiple classes. In addition, the proposed model completes five federated training rounds efficiently, with a lower average loss (0.3 vs. 0.45) and a faster round time (0.32 s vs. 0.8 s), reflecting both stability in learning and computational speed. Unlike earlier systems that typically rely on single-node simulations and lack features like interpretability or data privacy, U-FDL-PPE runs across three client nodes, supports explainability through Grad-CAM visualizations, and complies with privacy regulations such as GDPR. Its integration of the lightweight MobileNetV2 architecture also makes it ideal for resource-constrained or decentralized clinical environments. Overall, these improvements clearly demonstrate that U-FDL-PPE is a scalable, privacy-aware, and clinically applicable solution ready for real-world deployment.

The performance review of the U-FDL-PPE framework highlights clear and measurable improvements over earlier centralized and federated deep learning methods in medical image classification. The framework achieved an impressive accuracy of 88% and an F1-score of 89.66%, exceeding the average performance of existing models, which typically score around 82.5% accuracy and 80% F1-score. These results show that U-FDL-PPE can more accurately distinguish between COVID-19, Normal, and Viral Pneumonia cases using chest x-rays. Although the AUC score of 0.5192 indicates moderate class separation, it still reflects an improvement over older models. The training was completed across five federated learning rounds, resulting in a low average loss of 0.30, showing reliable and consistent learning. Additionally, the model reduced training time per round from 0.8 s (in older systems) to just 0.32 s, demonstrating enhanced speed and efficiency suitable for real-time diagnostic scenarios. More than just numbers, U-FDL-PPE also integrates key features needed for real-world application. It was tested in a three-client federated setup, reflecting how hospitals would operate collaboratively without sharing sensitive data. The inclusion of Grad-CAM provided intuitive heatmaps that helped doctors visually interpret the AI's predictions, adding a layer of trust and transparency. The system complies with GDPR privacy standards, as it never transmits raw patient data—only encrypted model updates. The use of the lightweight MobileNetV2 model enables fast and resource-friendly processing, making the system deployable in rural or low-resource healthcare settings. With detailed visualizations such as confusion matrices and training performance graphs, the evaluation shows that U-FDL-PPE is a powerful, explainable, and privacy-conscious AI solution well-suited for real-world, multi-institutional healthcare environments. In the U-FDL-PPE framework, the F1-score and AUC metrics were calculated using a macro-averaged, one-vs.-rest method to effectively evaluate the model's performance across the three classes COVID-19, Normal, and Viral Pneumonia. This approach ensures that each category is weighted equally during evaluation, providing a fair and balanced assessment even when the dataset contains varying sample sizes or class imbalances common in medical imaging tasks.