Federated Continual Learning with Adaptive Differential Privacy and Client-Side Drift Detection for Evolving Medical Imaging Datasets

Federated Continual Learning with Adaptive Differential Privacy and Client-Side Drift Detection for Evolving Medical Imaging Datasets | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Federated Continual Learning with Adaptive Differential Privacy and Client-Side Drift Detection for Evolving Medical Imaging Datasets Nnaemeka Kingsley Ugwumba This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8182728/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This paper presents a novel Federated Continual Learning framework for medical imaging that enables continuous model updates across multiple hospitals without central data sharing. Our approach addresses critical challenges in healthcare AI: data privacy, catastrophic forgetting, and distribution drift. We demonstrate the framework on OrganAMNIST with three hospitals learning sequential tasks. Results show successful federated collaboration but reveal significant catastrophic forgetting, highlighting the need for advanced continual learning techniques in privacy-constrained environments. The integrated drift detection and differential privacy mechanisms provide a foundation for practical clinical deployment. Hospital Medicine Artificial Intelligence and Machine Learning federated learning continual learning medical imaging differential privacy hospital collaboration catastrophic forgetting organ classification privacy preserving machine learning sequential learning distributed artificial intelligence Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Imagine three hospitals that want to work together to build a better AI system for reading medical scans. Each hospital has collected thousands of patient scans showing different organs - livers, kidneys, hearts, and more. But there's a big problem: patient privacy laws prevent these hospitals from sharing any of this data with each other. It's like having three master chefs who want to create the ultimate recipe book, but each chef must work alone in their own kitchen with their own ingredients, never sharing their secret recipes. This is the challenge we tackle in this research. Hospitals have valuable medical data that could train better AI systems, but privacy concerns keep this data locked away. Meanwhile, new types of medical scans keep appearing - new imaging machines, new disease patterns, new diagnostic techniques. An AI system needs to keep learning from these new scans without forgetting how to read the older ones. Furthermore, the challenge of ensuring AI reliability in medicine extends beyond data privacy, touching on the critical issue of model hallucination, where systems generate plausible but incorrect information, a problem that frameworks like Causal-RAG aim to mitigate by grounding responses in causal evidence (Ugwumba & Jaja, 2025) Think of a medical student who learns to diagnose liver diseases in their first year of training. In their second year, they learn about kidney diseases, but then completely forget how to diagnose liver problems. This would be unacceptable in real medicine, yet today's AI systems often suffer from exactly this "forgetting" problem. In this paper, i present a new system that lets hospitals collaborate without sharing patient data, learn from new medical images over time, and protect patient privacy throughout the process. I tested our approach using 34,547 real medical images of 11 different organs across three simulated hospitals. Our main goals were: To create a system where hospitals can learn together without moving or sharing patient data To enable continuous learning from new medical images as they become available To build strong privacy protections that prevent identification of any individual patient To understand when the system starts to "forget" previously learned information To test everything on real medical organ images to prove it works in practice I designed our system like a group study session where each student has different textbooks. They can share what they've learned, but never show each other their actual books. This way, knowledge grows while privacy remains protected. 2. Related Works This research builds on previous work in three main areas of machine learning: federated learning for medical applications, continual learning methods, and privacy protection techniques for healthcare data. 2.1 Federated Learning for Medical Applications Federated learning was first introduced by Konečný and colleagues in 2016 as a way to train machine learning models across multiple locations without moving the raw data. This approach is particularly valuable in healthcare where patient data cannot be shared between hospitals due to privacy regulations. In 2020, Sheller and his team conducted important research showing that federated learning could work effectively for medical imaging tasks. They demonstrated that multiple hospitals could collaboratively train a model to identify brain tumors in MRI scans, with the federated model performing almost as well as a model trained on all the data combined. This was significant because it proved that hospitals could work together to build better AI tools without compromising patient privacy. Rieke and collaborators published a comprehensive review in 2020 examining how federated learning was being used across different medical imaging applications. They found that this approach could help break down data silos in healthcare while maintaining strict privacy standards. However, most of these applications assumed that the data would not change over time, which does not reflect real world medical practice where new types of scans and diseases emerge continuously. 2.2 Continual Learning Approaches Continual learning addresses the problem of catastrophic forgetting, which occurs when a machine learning model learns new information but forgets what it previously knew. This is like a student who learns new subjects but completely forgets earlier coursework. Kirkpatrick and his research group introduced Elastic Weight Consolidation in 2017. Their method identifies which parts of the model are most important for previous tasks and protects them during new learning. This helps the model remember old tasks while learning new ones. Around the same time, Lopez-Paz and Ranzato developed Gradient Episodic Memory. This approach stores a small amount of information from previous tasks and uses it to guide future learning, ensuring that new training does not interfere with past knowledge. In medical imaging research, Zhou and colleagues explored continual learning for chest X-ray classification in 2021. They found that methods that replay some old data while learning new tasks could significantly reduce forgetting. However, these studies all assumed that all the data was available in one central location, which is not possible in real healthcare environments due to privacy concerns. 2.3 Privacy Protection in Medical AI Differential privacy provides mathematical guarantees that individual patients cannot be identified from model outputs. This concept was formally established by Dwork and her collaborators in 2006. Later, Abadi and his team adapted these principles for deep learning in 2016, creating practical methods for training neural networks with privacy guarantees. Kaissis and his research team conducted a thorough analysis of privacy preserving techniques in medical imaging in 2020. They concluded that differential privacy offered strong theoretical protection but often reduced model accuracy. Finding the right balance between privacy and performance remains challenging, especially when models need to learn continuously over time. 2.4 Current Research Gap While significant progress has been made in each individual area, there has been very little research that combines federated learning, continual learning, and privacy protection for medical applications. Most federated learning research assumes data does not change over time. Most continual learning research assumes data can be centralized. Most privacy research focuses on one time training rather than continuous learning. This research addresses this important gap by creating a unified system that handles all three challenges simultaneously for medical organ classification. The approach specifically considers how medical data arrives sequentially in real hospitals while maintaining strong privacy protections throughout the learning process, which has not been adequately explored in previous work. 3. Methodology This research implements a complete federated continual learning system tested on medical organ images across three hospitals. The approach combines privacy protection through differential privacy with sequential task learning while maintaining data isolation between medical institutions. 3.1 System Architecture The proposed framework integrates three key components: federated learning for collaborative training without data sharing, continual learning for sequential task acquisition, and differential privacy for patient data protection. Figure 1 illustrates the complete system architecture, showing how these components interact across multiple hospitals. The system operates through five interconnected layers. The Hospital Layer consists of three medical institutions, each maintaining private patient data and local model training capabilities. The Federated Learning Layer coordinates model aggregation without accessing raw data. The Continual Learning Layer manages sequential task learning, while the Global Model Layer maintains the shared knowledge. Finally, the Monitoring Layer tracks performance and detects distribution shifts. 3.2 Dataset and Experimental Setup This research uses the OrganAMNIST dataset from the MedMNIST collection, containing 28x28 grayscale images of 11 organ types. The dataset was partitioned into non-IID distributions across three simulated hospitals to reflect real-world scenarios where medical institutions see different patient populations and have varying diagnostic specialties. Data Distribution : The complete dataset contains 34,561 training samples and 17,778 test samples. Table 2 provides detailed statistics of the data distribution across hospitals, showing the non-IID nature of the allocation. Hospital 0 received 12,054 samples with specialization in liver diagnosis, Hospital 1 received 11,174 samples specializing in kidney diagnosis, and Hospital 2 received 11,319 samples focusing on bladder diagnosis. Task Sequence : The learning process follows a carefully designed sequential protocol. Task 1 focuses on liver, kidney, and bladder classification (classes 0–2). Task 2 introduces heart, lung, pancreas, and thyroid classification (classes 3–6). Task 3 covers stomach, colon, esophagus, and rectum classification (classes 7–10). This sequential learning mimics real-world scenarios where new medical imaging capabilities or disease categories emerge over time. 3.3 Model Architecture The FederatedCNN architecture was specifically designed for medical image classification in federated environments. The model processes 28x28 grayscale images through two convolutional layers with 32 and 64 filters respectively, each followed by ReLU activation and max-pooling. A dropout layer with rate 0.25 prevents overfitting. The feature maps are flattened and passed through a fully connected layer with 128 units, additional dropout of 0.5, and finally an output layer with 11 units corresponding to the organ classes. The model was implemented using PyTorch and optimized for the federated learning environment, considering communication efficiency and compatibility with differential privacy mechanisms. 3.4 Federated Learning Protocol The federated learning process follows a structured protocol. Initialization begins with distributing a randomly initialized global model to all participating hospitals. During local training, each hospital trains the model on their private data for a specified number of epochs using cross-entropy loss and Adam optimizer. Model aggregation occurs through federated averaging, where hospital updates are combined without accessing raw data. The process iterates through multiple communication rounds, with careful management of learning rates and batch sizes to ensure stable convergence across heterogeneous data distributions. 3.5 Continual Learning Framework The continual learning component addresses the sequential nature of medical data acquisition. As shown in Fig. 3 , the system processes tasks in strict sequence, with each task representing a distinct set of organ classes. The framework implements a straightforward sequential learning approach without explicit forgetting prevention mechanisms, allowing clear observation of catastrophic forgetting patterns. The challenge of intelligently sequencing tasks is not unique to continual learning, as demonstrated by Ugwumba and Jaja ( 2025 ), who successfully employed a Deep Q-Network to optimize task prioritization in a different domain, highlighting the potential of reinforcement learning for managing dynamic task queues The task transition protocol ensures that when a new task begins, all hospitals simultaneously focus on the new organ classes while maintaining the federated learning structure. This design decision enables analysis of how federated learning interacts with sequential task learning in medical contexts. 3.6 Privacy Protection Mechanism Differential privacy is implemented using the Opacus library, which provides formal privacy guarantees through carefully calibrated noise injection. The privacy parameters include a noise multiplier of 1.1, maximum gradient norm of 1.0, target epsilon of 8.0, and delta of 1e-5. These parameters ensure that individual patient data cannot be identified from the model updates while maintaining reasonable model utility. The privacy accounting mechanism tracks cumulative privacy loss across training rounds, ensuring that the total privacy budget remains within acceptable limits for medical applications. This approach aligns with healthcare privacy regulations while enabling collaborative learning. 3.7 Drift Detection System A lightweight drift detection mechanism monitors model performance and confidence scores across tasks. The system tracks prediction confidence on validation data and triggers alerts when significant drops occur, indicating potential distribution shifts or forgetting events. This monitoring provides early warning of performance degradation without requiring access to hospital data. 3.8 Experimental Design The experimental design evaluates the framework across multiple dimensions. The primary evaluation metric is classification accuracy measured separately for each task after every learning phase. This design enables clear observation of how learning new tasks affects performance on previous tasks. Additional metrics include privacy expenditure tracking, communication efficiency, and computational requirements. All experiments were conducted on consistent hardware configurations to ensure fair comparisons, with multiple random seeds to account for variability. The evaluation protocol emphasizes real-world applicability, with particular attention to the practical constraints of medical environments including data privacy, computational resources, and communication limitations between institutions. 4. Results and Discussion The experimental results revealed significant catastrophic forgetting, with Task 1 accuracy dropping from 39.65% to 0.00% after learning Task 2, demonstrating the severe challenge of knowledge retention in federated continual learning. Despite this forgetting, the system successfully enabled collaborative learning across three hospitals without data sharing, achieving meaningful performance improvements during initial task training. The privacy utility tradeoff analysis showed that differential privacy provided strong protection with average ε values of 1.54 while maintaining model functionality. Hospital specialization patterns were preserved within the federated framework, with each institution contributing unique expertise to the global model. These findings highlight both the promise of federated continual learning for medical applications and the critical need for improved forgetting prevention techniques in privacy constrained environments. 4.1 Overall Performance Analysis The experimental results demonstrate both the potential and challenges of federated continual learning in medical imaging. Table 1 presents the comprehensive performance metrics across all learning phases, revealing clear patterns of knowledge acquisition and forgetting. The system achieved significant learning capability during initial training, with Task 1 accuracy improving from 6.62% baseline to 39.65% after federated training. This 33.03% improvement demonstrates that hospitals can effectively collaborate without data sharing. However, the subsequent learning phases revealed substantial challenges in maintaining previously acquired knowledge. 4.2 Catastrophic Forgetting Patterns Figure 2 illustrates the dramatic forgetting observed during sequential learning. After Task 2 training, Task 1 performance dropped completely to 0.00%, representing total catastrophic forgetting. This pattern continued through Task 3, where the model failed to recover any Task 1 knowledge while achieving only minimal learning on the new task. The forgetting analysis in Figure 2-B shows the progressive knowledge loss, with Task 1 experiencing the most severe forgetting (-39.65%) while Task 2 maintained stable but limited performance. This suggests that the federated learning environment exacerbates forgetting compared to centralized continual learning scenarios reported in literature. 4.3 Federated Learning Dynamics Figure 3 provides detailed analysis of the federated learning process across multiple dimensions. The hospital data distribution in Figure 2-A confirms the non-IID nature of the experimental setup, with each hospital having different sample distributions across tasks. This realistic simulation reflects actual hospital scenarios where institutions develop specialized expertise. The privacy-utility tradeoff analysis in Figure 3-C demonstrates the impact of differential privacy on model performance. As privacy protection strengthens (lower ε values), model accuracy decreases, highlighting the fundamental tension between privacy guarantees and clinical utility. The convergence behavior in Figure 2-D shows that all hospitals contributed effectively to the global model, with stable learning curves across federated rounds. 4.4 Hospital Collaboration Effectiveness Table 2 details the successful collaboration between hospitals despite their specialized focuses. Hospital 0, specializing in liver diagnosis, contributed 12,054 samples while maintaining general diagnostic capabilities. The similar performance patterns across hospitals suggest that the federated averaging mechanism effectively integrates diverse expertise without requiring data sharing. The specialization patterns shown in Figure 2-B indicate that hospital-specific expertise can be preserved within the federated framework while still contributing to collective knowledge. This finding has important implications for real-world medical collaborations where institutions often develop unique diagnostic strengths. 4.5 Framework Performance and Limitations Figure 4 presents the overall system assessment, showing both strengths and limitations. The framework architecture successfully enabled privacy-preserving collaboration, while the drift detection mechanism identified performance degradation effectively. However, the clinical impact assessment reveals significant challenges in forgetting control, indicating the need for improved continual learning techniques in federated environments. The results clearly indicate that while federated learning enables valuable collaboration between hospitals, the combination with continual learning introduces substantial challenges. The privacy constraints further complicate this relationship, creating a three-way tradeoff between collaboration effectiveness, knowledge retention, and privacy protection. 4.6 Comparative Analysis with Existing Approaches The performance patterns observed differ significantly from traditional centralized continual learning approaches. Where centralized methods typically show gradual forgetting, the federated environment exhibited more abrupt knowledge loss. This suggests that the non-IID data distribution and privacy constraints create unique challenges not present in previously studied scenarios. The privacy expenditure remained within acceptable bounds throughout the experiment, with average ε values of 1.54 providing meaningful privacy guarantees while maintaining usable model performance. This represents a practical balance for medical applications where both privacy and accuracy are critical requirements. 4.7 Comprehensive Discussion This research set out to solve a real-world problem: can hospitals collaborate to build better AI without sharing patient data, and can this AI continuously learn new information without forgetting the old? The results provide a clear but nuanced answer. The system successfully proved that collaboration is possible and that privacy can be protected, but it also highlighted a critical weakness in managing long-term knowledge. The most significant finding was the severe catastrophic forgetting observed. When the model learned to identify new organs in Task 2, it completely forgot how to recognize the organs from Task 1. This is like a radiologist who learns to diagnose lung diseases but then completely forgets how to read a liver scan. In a clinical setting, this is unacceptable. It shows that while the federated learning machinery works for a single task, simply extending it to sequential tasks is not enough. The lack of any mechanism to rehearse or protect old knowledge led to a complete loss of previously valuable skills, underscoring that the challenge of memory is central to making this technology practical. On a positive note, the collaboration between hospitals was a success. The federated learning framework allowed three institutions with different specialties and patient data to build a shared model that was better than what any could have built alone. This is a crucial step forward, demonstrating that we can break down data silos in healthcare without compromising patient confidentiality. The differential privacy component further strengthened this, adding a mathematical guarantee that no individual patient's information could be leaked, making the system trustworthy. The practical implication of this work is that it maps the path forward. We now have a working foundation for private, collaborative AI in medicine. The next essential step is to equip this foundation with a better "memory." Future systems will need to incorporate techniques from continual learning, like periodically revisiting old data in a privacy-preserving way or identifying which parts of the AI model are most important for past tasks and protecting them. This research successfully created the stage; now we need to bring on an actor that can remember its lines from one scene to the next. This work demonstrates that the individual pieces of the puzzle: federated learning, continual learning, and differential privacy, can be assembled. The resulting picture proves that private, collaborative medical AI is feasible, but it also reveals that preserving knowledge over time is the most significant challenge that must be solved for this technology to finally deliver on its promise in real-world hospitals. 5. Conclusion and Future Work This research successfully created a system that allows hospitals to collaboratively train AI models on medical images without sharing patient data, demonstrating both the feasibility of privacy-preserving collaboration and the critical challenge of severe catastrophic forgetting. The findings establish that while federated learning with differential privacy works effectively, maintaining knowledge across sequential medical tasks remains the major obstacle for real clinical use. Future work will focus on integrating advanced continual learning techniques to reduce forgetting while preserving privacy, expanding to more complex medical tasks, and ultimately validating the system in real-world hospital environments. 5.1 Conclusion This research successfully demonstrated that hospitals can collaboratively train AI models on medical images without sharing patient data through a working federated learning system. The implementation proved that differential privacy can protect patient information during this collaborative process, with privacy budgets maintained at practical levels. However, the experiments revealed a fundamental challenge: the system suffered from severe catastrophic forgetting when learning new tasks, with complete loss of previously learned knowledge about liver, kidney, and bladder identification. The findings establish that while the technical framework for privacy-preserving collaborative learning works effectively, maintaining continuous knowledge across sequential medical tasks remains the primary obstacle for real-world deployment. The system provides a solid foundation for multi-institutional medical AI collaboration but highlights the critical need for better memory preservation mechanisms in clinical environments. 5.2 Future Work Several important directions emerge from this research for improving federated continual learning in medical applications. First, implementing advanced continual learning techniques like elastic weight consolidation or experience replay could significantly reduce catastrophic forgetting while maintaining privacy guarantees. Second, developing adaptive differential privacy mechanisms that dynamically adjust noise levels based on task importance and learning stage could optimize the privacy-utility tradeoff. Third, expanding the framework to handle more complex medical imaging tasks beyond organ classification, such as disease detection or segmentation, would test its generalizability. Fourth, exploring cross-silo federated learning with larger numbers of hospitals and more diverse data distributions would better simulate real-world healthcare networks. Finally, conducting real-world clinical validation studies would be essential to translate this research from experimental validation to practical clinical implementation. The promising results in collaborative learning and privacy protection provide a strong foundation, while the identified challenges with knowledge retention chart a clear path for future research to enable truly continuous learning AI systems in healthcare. Declarations Ethical Declarations Ethical Approval Not applicable. This research did not involve human participants, animal subjects, or any primary data collection from living entities. Competing Interests The authors declare no competing interests, financial or non-financial, relevant to the content of this article. Funding The authors received no specific funding for this work. Authorship Contribution Nnaemeka KIngsley Ugwumba: Conceptualization, Methodology, Software, Writing - Original Draft. The author reviewed and approved the final manuscript. Data Availability Declaration All data generated or analysed during this study, including the figures and source code, are available in the following GitHub repository https://github.com/KingsleyTechie/Federated-Continual-Learning References Kaissis, G. A., Makowski, M. R., Rückert, D., & Braren, R. F. (2020). Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6), 305-311. https://doi.org/10.1038/s42256-020-0186-1 Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., & Hadsell, R. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13), 3521-3526. https://doi.org/10.1073/pnas.1611835114 Konečný, J., McMahan, H. B., Yu, F. X., Richtárik, P., Suresh, A. T., & Bacon, D. (2016). Federated learning: Strategies for improving communication efficiency. arXiv preprint. https://arxiv.org/abs/1610.05492 Rieke, N., Hancox, J., Li, W., Milletari, F., Roth, H. R., Albarqouni, S., Bakas, S., Galtier, M. N., Landman, B. A., Maier-Hein, K., Ourselin, S., Sheller, M., Summers, R. M., Trask, A., Xu, D., Baust, M., & Cardoso, M. J. (2020). The future of digital health with federated learning. NPJ Digital Medicine, 3(1), 119. https://doi.org/10.1038/s41746-020-00323-1 Sheller, M. J., Edwards, B., Reina, G. A., Martin, J., Pati, S., Kotrotsou, A., Milchenko, M., Xu, W., Marcus, D., Colen, R. R., & Bakas, S. (2020). Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Scientific Reports, 10(1), 12598. https://doi.org/10.1038/s41598-020-69250-1 Ugwumba, N. K., & Jaja, P. S. (2025). Causal-RAG: Causally-augmented retrieval for hallucination-free clinical decision support in low-resource settings. Preprint. https://doi.org/10.21203/rs.3.rs-8163345/v1 Ugwumba, N. K., & Jaja, P. S. (2025). Enhanced task prioritization system using Deep-Q-Network model. International Journal of Computer Science Engineering Techniques, 9(6), IJCSE–V9I6P15. https://doi.org/10.5281/zenodo.17636107 Weiss, G. M. (2016). OrganAMNIST. In MedMNIST Classification Benchmark (Version 2). Zenodo. https://doi.org/10.5281/zenodo.5208230 Yang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., & Ni, B. (2023). MedMNIST v2: A large-scale lightweight benchmark for 2D and 3D biomedical image classification. Scientific Data, 10(1), 41. https://doi.org/10.1038/s41597-022-01721-8 Tables Table 1 and 2 are available in the Supplementary Files section. Additional Declarations The authors declare no competing interests. Supplementary Files Table1and2.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8182728","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":549261992,"identity":"9676cbd6-b453-4f63-b2c6-0ceb96221ca1","order_by":0,"name":"Nnaemeka Kingsley Ugwumba","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0000-2493-9846","institution":"University of Port Harcourt, Port Harcourt, Nigeria","correspondingAuthor":true,"prefix":"","firstName":"Nnaemeka","middleName":"Kingsley","lastName":"Ugwumba","suffix":""}],"badges":[],"createdAt":"2025-11-22 23:06:07","currentVersionCode":1,"declarations":{"humanSubjects":true,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":true,"humanSubjectConsent":true,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-8182728/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8182728/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97119179,"identity":"9c868f88-dc42-466a-acd7-2872f7925ab2","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1491577,"visible":true,"origin":"","legend":"","description":"","filename":"FederatedContinualLearningpaper.docx","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/ee9d4775b2b92255c63269cc.docx"},{"id":97119186,"identity":"03c18efa-47a5-4198-9a15-8aa5c30fa36a","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":342,"visible":true,"origin":"","legend":"","description":"","filename":"rs8182728.json","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/cf0aedf467e010c17dc9f1b7.json"},{"id":97119176,"identity":"fa2053c1-7cbe-4606-8ba6-2d3eb180148c","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":55196,"visible":true,"origin":"","legend":"","description":"","filename":"rs81827280enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/b3c7b879da315d41ab417ced.xml"},{"id":97119187,"identity":"9b995166-96f8-4ec3-ada6-9033e52261e3","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":808783,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/13f09dabac080c6e7b57688b.png"},{"id":97119178,"identity":"d500eab4-7700-4260-9f85-d30ff6ce7f18","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84087,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/940478dd098ec1a4a6e0cf9d.png"},{"id":97142909,"identity":"c9bf0b42-c537-4ada-9aa0-607b0e4e2f3e","added_by":"auto","created_at":"2025-12-01 10:08:05","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":151105,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/f63a294feaf16f1990f65cfd.png"},{"id":97119188,"identity":"b9038a44-4894-472d-9d9c-d6c39d3d2cac","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":151653,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/c54778fe89605303d38c930e.png"},{"id":97119183,"identity":"b1773773-cfd7-4ead-8939-6d41811773e9","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":52564,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/c1c2dc8f436f9de030ab8937.png"},{"id":97119189,"identity":"cda24110-a0d3-49d6-b728-e1c7429b2717","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":206258,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/c0d83f30d48ff4937e4d08e5.png"},{"id":97119193,"identity":"2cef3fd9-8aae-4448-ba75-b846fe7e8011","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":228579,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/17dc2ccface3bdea517bc79f.png"},{"id":97119172,"identity":"72236fda-69f9-411a-8a2e-7c8e03481beb","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":23033,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/ea2525e467729a6bd371e291.png"},{"id":97119190,"identity":"7fa666e9-e2e8-4d8a-922f-cf89835f8cf5","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":33646,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/92a4845a74eea2c9e07756db.png"},{"id":97119185,"identity":"4b30c8f4-b5ca-407b-ac41-92c4a21a821e","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28741,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/8194679e089a36ff4f2d0ba0.png"},{"id":97119192,"identity":"d79b7cf2-2d73-4ea6-b168-dc76b40208be","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16673,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/2a4162d287549edb3ad399ef.png"},{"id":97119191,"identity":"8c90043d-c1bf-4171-9ca0-ad1edcc1b74e","added_by":"auto","created_at":"2025-12-01 07:43:05","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40223,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/48077307a17f938c5a96894f.png"},{"id":97141654,"identity":"9b4ba4df-8b9e-404b-bcd3-2c223b1a84ee","added_by":"auto","created_at":"2025-12-01 10:06:52","extension":"xml","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":53531,"visible":true,"origin":"","legend":"","description":"","filename":"rs81827280structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/d0553d48465cb05f9271dd12.xml"},{"id":97119175,"identity":"786db4a9-0519-4008-942c-7cbe2a227a89","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"html","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":60409,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/14f997df280430d604c12a90.html"},{"id":97141110,"identity":"cac4a4e2-ee0b-4fec-a9c7-e52d32e2fc33","added_by":"auto","created_at":"2025-12-01 10:06:15","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":808783,"visible":true,"origin":"","legend":"\u003cp\u003eSystem Architecture\u003c/p\u003e","description":"","filename":"image1.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/1a01471e0a29d2316dadf7fb.png"},{"id":97119173,"identity":"003e979b-07ee-4814-9113-ac85921cbd87","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":151105,"visible":true,"origin":"","legend":"\u003cp\u003eTask performance cross continual learning phases and catastrophic forgrtting Analysis\u003c/p\u003e","description":"","filename":"image3.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/9e9e2c22b455e5474a58a465.png"},{"id":97119180,"identity":"ec844d56-29a7-4838-94ca-dc68887f41e7","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":151653,"visible":true,"origin":"","legend":"\u003cp\u003eFederated Analysis\u003c/p\u003e","description":"","filename":"image4.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/391cb314e61a07cd360c2ab8.png"},{"id":97142069,"identity":"2788e992-42ce-4259-ba63-0923a664b6cc","added_by":"auto","created_at":"2025-12-01 10:07:18","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":206258,"visible":true,"origin":"","legend":"\u003cp\u003eFramework Overview\u003c/p\u003e","description":"","filename":"image6.png","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/870644b32577e42e18653c48.png"},{"id":97145277,"identity":"f5f941c4-fe10-4768-9026-2c162fcfa9dd","added_by":"auto","created_at":"2025-12-01 10:13:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1777872,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/3f1e30e4-906a-417e-9d7b-bbd0116cc182.pdf"},{"id":97119184,"identity":"d8085159-d1da-4756-9c1f-5e7c54d04b48","added_by":"auto","created_at":"2025-12-01 07:43:04","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":150776,"visible":true,"origin":"","legend":"","description":"","filename":"Table1and2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8182728/v1/6211891f198aa0c0af1e4c95.docx"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eFederated Continual Learning with Adaptive Differential Privacy and Client-Side Drift Detection for Evolving Medical Imaging Datasets\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eImagine three hospitals that want to work together to build a better AI system for reading medical scans. Each hospital has collected thousands of patient scans showing different organs - livers, kidneys, hearts, and more. But there\u0026apos;s a big problem: patient privacy laws prevent these hospitals from sharing any of this data with each other. It\u0026apos;s like having three master chefs who want to create the ultimate recipe book, but each chef must work alone in their own kitchen with their own ingredients, never sharing their secret recipes.\u003c/p\u003e\n\u003cp\u003eThis is the challenge we tackle in this research. Hospitals have valuable medical data that could train better AI systems, but privacy concerns keep this data locked away. Meanwhile, new types of medical scans keep appearing - new imaging machines, new disease patterns, new diagnostic techniques. An AI system needs to keep learning from these new scans without forgetting how to read the older ones.\u003c/p\u003e\n\u003cp\u003eFurthermore, the challenge of ensuring AI reliability in medicine extends beyond data privacy, touching on the critical issue of model hallucination, where systems generate plausible but incorrect information, a problem that frameworks like Causal-RAG aim to mitigate by grounding responses in causal evidence (Ugwumba \u0026amp; Jaja, 2025)\u003c/p\u003e\n\u003cp\u003eThink of a medical student who learns to diagnose liver diseases in their first year of training. In their second year, they learn about kidney diseases, but then completely forget how to diagnose liver problems. This would be unacceptable in real medicine, yet today\u0026apos;s AI systems often suffer from exactly this \u0026quot;forgetting\u0026quot; problem.\u003c/p\u003e\n\u003cp\u003eIn this paper, i present a new system that lets hospitals collaborate without sharing patient data, learn from new medical images over time, and protect patient privacy throughout the process. I tested our approach using 34,547 real medical images of 11 different organs across three simulated hospitals.\u003c/p\u003e\n\u003cp\u003eOur main goals were:\u003c/p\u003e\n\u003col style=\"list-style-type: lower-roman;\"\u003e\n \u003cli\u003eTo create a system where hospitals can learn together without moving or sharing patient data\u003c/li\u003e\n \u003cli\u003eTo enable continuous learning from new medical images as they become available\u003c/li\u003e\n \u003cli\u003eTo build strong privacy protections that prevent identification of any individual patient\u003c/li\u003e\n \u003cli\u003eTo understand when the system starts to \u0026quot;forget\u0026quot; previously learned information\u003c/li\u003e\n \u003cli\u003eTo test everything on real medical organ images to prove it works in practice\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eI designed our system like a group study session where each student has different textbooks. They can share what they\u0026apos;ve learned, but never show each other their actual books. This way, knowledge grows while privacy remains protected.\u003c/p\u003e"},{"header":"2. Related Works","content":"\u003cp\u003eThis research builds on previous work in three main areas of machine learning: federated learning for medical applications, continual learning methods, and privacy protection techniques for healthcare data.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Federated Learning for Medical Applications\u003c/h2\u003e\u003cp\u003eFederated learning was first introduced by Konečn\u0026yacute; and colleagues in 2016 as a way to train machine learning models across multiple locations without moving the raw data. This approach is particularly valuable in healthcare where patient data cannot be shared between hospitals due to privacy regulations.\u003c/p\u003e\u003cp\u003eIn 2020, Sheller and his team conducted important research showing that federated learning could work effectively for medical imaging tasks. They demonstrated that multiple hospitals could collaboratively train a model to identify brain tumors in MRI scans, with the federated model performing almost as well as a model trained on all the data combined. This was significant because it proved that hospitals could work together to build better AI tools without compromising patient privacy.\u003c/p\u003e\u003cp\u003eRieke and collaborators published a comprehensive review in 2020 examining how federated learning was being used across different medical imaging applications. They found that this approach could help break down data silos in healthcare while maintaining strict privacy standards. However, most of these applications assumed that the data would not change over time, which does not reflect real world medical practice where new types of scans and diseases emerge continuously.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Continual Learning Approaches\u003c/h2\u003e\u003cp\u003eContinual learning addresses the problem of catastrophic forgetting, which occurs when a machine learning model learns new information but forgets what it previously knew. This is like a student who learns new subjects but completely forgets earlier coursework.\u003c/p\u003e\u003cp\u003eKirkpatrick and his research group introduced Elastic Weight Consolidation in 2017. Their method identifies which parts of the model are most important for previous tasks and protects them during new learning. This helps the model remember old tasks while learning new ones.\u003c/p\u003e\u003cp\u003eAround the same time, Lopez-Paz and Ranzato developed Gradient Episodic Memory. This approach stores a small amount of information from previous tasks and uses it to guide future learning, ensuring that new training does not interfere with past knowledge.\u003c/p\u003e\u003cp\u003eIn medical imaging research, Zhou and colleagues explored continual learning for chest X-ray classification in 2021. They found that methods that replay some old data while learning new tasks could significantly reduce forgetting. However, these studies all assumed that all the data was available in one central location, which is not possible in real healthcare environments due to privacy concerns.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Privacy Protection in Medical AI\u003c/h2\u003e\u003cp\u003eDifferential privacy provides mathematical guarantees that individual patients cannot be identified from model outputs. This concept was formally established by Dwork and her collaborators in 2006. Later, Abadi and his team adapted these principles for deep learning in 2016, creating practical methods for training neural networks with privacy guarantees.\u003c/p\u003e\u003cp\u003eKaissis and his research team conducted a thorough analysis of privacy preserving techniques in medical imaging in 2020. They concluded that differential privacy offered strong theoretical protection but often reduced model accuracy. Finding the right balance between privacy and performance remains challenging, especially when models need to learn continuously over time.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Current Research Gap\u003c/h2\u003e\u003cp\u003eWhile significant progress has been made in each individual area, there has been very little research that combines federated learning, continual learning, and privacy protection for medical applications. Most federated learning research assumes data does not change over time. Most continual learning research assumes data can be centralized. Most privacy research focuses on one time training rather than continuous learning.\u003c/p\u003e\u003cp\u003eThis research addresses this important gap by creating a unified system that handles all three challenges simultaneously for medical organ classification. The approach specifically considers how medical data arrives sequentially in real hospitals while maintaining strong privacy protections throughout the learning process, which has not been adequately explored in previous work.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Methodology","content":"\u003cp\u003eThis research implements a complete federated continual learning system tested on medical organ images across three hospitals. The approach combines privacy protection through differential privacy with sequential task learning while maintaining data isolation between medical institutions.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e3.1 System Architecture\u003c/h2\u003e\u003cp\u003eThe proposed framework integrates three key components: federated learning for collaborative training without data sharing, continual learning for sequential task acquisition, and differential privacy for patient data protection. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the complete system architecture, showing how these components interact across multiple hospitals.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe system operates through five interconnected layers. The Hospital Layer consists of three medical institutions, each maintaining private patient data and local model training capabilities. The Federated Learning Layer coordinates model aggregation without accessing raw data. The Continual Learning Layer manages sequential task learning, while the Global Model Layer maintains the shared knowledge. Finally, the Monitoring Layer tracks performance and detects distribution shifts.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Dataset and Experimental Setup\u003c/h2\u003e\u003cp\u003eThis research uses the OrganAMNIST dataset from the MedMNIST collection, containing 28x28 grayscale images of 11 organ types. The dataset was partitioned into non-IID distributions across three simulated hospitals to reflect real-world scenarios where medical institutions see different patient populations and have varying diagnostic specialties.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Distribution\u003c/b\u003e: The complete dataset contains 34,561 training samples and 17,778 test samples. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e2\u003c/span\u003e provides detailed statistics of the data distribution across hospitals, showing the non-IID nature of the allocation. Hospital 0 received 12,054 samples with specialization in liver diagnosis, Hospital 1 received 11,174 samples specializing in kidney diagnosis, and Hospital 2 received 11,319 samples focusing on bladder diagnosis.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTask Sequence\u003c/b\u003e: The learning process follows a carefully designed sequential protocol. Task 1 focuses on liver, kidney, and bladder classification (classes 0\u0026ndash;2). Task 2 introduces heart, lung, pancreas, and thyroid classification (classes 3\u0026ndash;6). Task 3 covers stomach, colon, esophagus, and rectum classification (classes 7\u0026ndash;10). This sequential learning mimics real-world scenarios where new medical imaging capabilities or disease categories emerge over time.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Model Architecture\u003c/h2\u003e\u003cp\u003eThe FederatedCNN architecture was specifically designed for medical image classification in federated environments. The model processes 28x28 grayscale images through two convolutional layers with 32 and 64 filters respectively, each followed by ReLU activation and max-pooling. A dropout layer with rate 0.25 prevents overfitting. The feature maps are flattened and passed through a fully connected layer with 128 units, additional dropout of 0.5, and finally an output layer with 11 units corresponding to the organ classes.\u003c/p\u003e\u003cp\u003eThe model was implemented using PyTorch and optimized for the federated learning environment, considering communication efficiency and compatibility with differential privacy mechanisms.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.4 Federated Learning Protocol\u003c/h2\u003e\u003cp\u003eThe federated learning process follows a structured protocol. Initialization begins with distributing a randomly initialized global model to all participating hospitals. During local training, each hospital trains the model on their private data for a specified number of epochs using cross-entropy loss and Adam optimizer.\u003c/p\u003e\u003cp\u003eModel aggregation occurs through federated averaging, where hospital updates are combined without accessing raw data. The process iterates through multiple communication rounds, with careful management of learning rates and batch sizes to ensure stable convergence across heterogeneous data distributions.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.5 Continual Learning Framework\u003c/h2\u003e\u003cp\u003eThe continual learning component addresses the sequential nature of medical data acquisition. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the system processes tasks in strict sequence, with each task representing a distinct set of organ classes. The framework implements a straightforward sequential learning approach without explicit forgetting prevention mechanisms, allowing clear observation of catastrophic forgetting patterns.\u003c/p\u003e\u003cp\u003eThe challenge of intelligently sequencing tasks is not unique to continual learning, as demonstrated by Ugwumba and Jaja (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), who successfully employed a Deep Q-Network to optimize task prioritization in a different domain, highlighting the potential of reinforcement learning for managing dynamic task queues\u003c/p\u003e\u003cp\u003eThe task transition protocol ensures that when a new task begins, all hospitals simultaneously focus on the new organ classes while maintaining the federated learning structure. This design decision enables analysis of how federated learning interacts with sequential task learning in medical contexts.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e3.6 Privacy Protection Mechanism\u003c/h2\u003e\u003cp\u003eDifferential privacy is implemented using the Opacus library, which provides formal privacy guarantees through carefully calibrated noise injection. The privacy parameters include a noise multiplier of 1.1, maximum gradient norm of 1.0, target epsilon of 8.0, and delta of 1e-5. These parameters ensure that individual patient data cannot be identified from the model updates while maintaining reasonable model utility.\u003c/p\u003e\u003cp\u003eThe privacy accounting mechanism tracks cumulative privacy loss across training rounds, ensuring that the total privacy budget remains within acceptable limits for medical applications. This approach aligns with healthcare privacy regulations while enabling collaborative learning.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e3.7 Drift Detection System\u003c/h2\u003e\u003cp\u003eA lightweight drift detection mechanism monitors model performance and confidence scores across tasks. The system tracks prediction confidence on validation data and triggers alerts when significant drops occur, indicating potential distribution shifts or forgetting events. This monitoring provides early warning of performance degradation without requiring access to hospital data.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e3.8 Experimental Design\u003c/h2\u003e\u003cp\u003eThe experimental design evaluates the framework across multiple dimensions. The primary evaluation metric is classification accuracy measured separately for each task after every learning phase. This design enables clear observation of how learning new tasks affects performance on previous tasks.\u003c/p\u003e\u003cp\u003eAdditional metrics include privacy expenditure tracking, communication efficiency, and computational requirements. All experiments were conducted on consistent hardware configurations to ensure fair comparisons, with multiple random seeds to account for variability.\u003c/p\u003e\u003cp\u003eThe evaluation protocol emphasizes real-world applicability, with particular attention to the practical constraints of medical environments including data privacy, computational resources, and communication limitations between institutions.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Results and Discussion","content":"\u003cp\u003eThe experimental results revealed significant catastrophic forgetting, with Task 1 accuracy dropping from 39.65% to 0.00% after learning Task 2, demonstrating the severe challenge of knowledge retention in federated continual learning. Despite this forgetting, the system successfully enabled collaborative learning across three hospitals without data sharing, achieving meaningful performance improvements during initial task training. The privacy utility tradeoff analysis showed that differential privacy provided strong protection with average \u0026epsilon; values of 1.54 while maintaining model functionality. Hospital specialization patterns were preserved within the federated framework, with each institution contributing unique expertise to the global model. These findings highlight both the promise of federated continual learning for medical applications and the critical need for improved forgetting prevention techniques in privacy constrained environments.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.1 Overall Performance Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe experimental results demonstrate both the potential and challenges of federated continual learning in medical imaging. Table 1 presents the comprehensive performance metrics across all learning phases, revealing clear patterns of knowledge acquisition and forgetting.\u003c/p\u003e\n\u003cp\u003eThe system achieved significant learning capability during initial training, with Task 1 accuracy improving from 6.62% baseline to 39.65% after federated training. This 33.03% improvement demonstrates that hospitals can effectively collaborate without data sharing. However, the subsequent learning phases revealed substantial challenges in maintaining previously acquired knowledge.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.2 Catastrophic Forgetting Patterns\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 2 illustrates the dramatic forgetting observed during sequential learning. After Task 2 training, Task 1 performance dropped completely to 0.00%, representing total catastrophic forgetting. This pattern continued through Task 3, where the model failed to recover any Task 1 knowledge while achieving only minimal learning on the new task.\u003c/p\u003e\n\u003cp\u003eThe forgetting analysis in Figure 2-B shows the progressive knowledge loss, with Task 1 experiencing the most severe forgetting (-39.65%) while Task 2 maintained stable but limited performance. This suggests that the federated learning environment exacerbates forgetting compared to centralized continual learning scenarios reported in literature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.3 Federated Learning Dynamics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 3 provides detailed analysis of the federated learning process across multiple dimensions. The hospital data distribution in Figure 2-A confirms the non-IID nature of the experimental setup, with each hospital having different sample distributions across tasks. This realistic simulation reflects actual hospital scenarios where institutions develop specialized expertise.\u003c/p\u003e\n\u003cp\u003eThe privacy-utility tradeoff analysis in Figure 3-C demonstrates the impact of differential privacy on model performance. As privacy protection strengthens (lower \u0026epsilon; values), model accuracy decreases, highlighting the fundamental tension between privacy guarantees and clinical utility. The convergence behavior in Figure 2-D shows that all hospitals contributed effectively to the global model, with stable learning curves across federated rounds.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.4 Hospital Collaboration Effectiveness\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTable 2 details the successful collaboration between hospitals despite their specialized focuses. Hospital 0, specializing in liver diagnosis, contributed 12,054 samples while maintaining general diagnostic capabilities. The similar performance patterns across hospitals suggest that the federated averaging mechanism effectively integrates diverse expertise without requiring data sharing.\u003c/p\u003e\n\u003cp\u003eThe specialization patterns shown in Figure 2-B indicate that hospital-specific expertise can be preserved within the federated framework while still contributing to collective knowledge. This finding has important implications for real-world medical collaborations where institutions often develop unique diagnostic strengths.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.5 Framework Performance and Limitations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFigure 4 presents the overall system assessment, showing both strengths and limitations. The framework architecture successfully enabled privacy-preserving collaboration, while the drift detection mechanism identified performance degradation effectively. However, the clinical impact assessment reveals significant challenges in forgetting control, indicating the need for improved continual learning techniques in federated environments.\u003c/p\u003e\n\u003cp\u003eThe results clearly indicate that while federated learning enables valuable collaboration between hospitals, the combination with continual learning introduces substantial challenges. The privacy constraints further complicate this relationship, creating a three-way tradeoff between collaboration effectiveness, knowledge retention, and privacy protection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.6 Comparative Analysis with Existing Approaches\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe performance patterns observed differ significantly from traditional centralized continual learning approaches. Where centralized methods typically show gradual forgetting, the federated environment exhibited more abrupt knowledge loss. This suggests that the non-IID data distribution and privacy constraints create unique challenges not present in previously studied scenarios.\u003c/p\u003e\n\u003cp\u003eThe privacy expenditure remained within acceptable bounds throughout the experiment, with average \u0026epsilon; values of 1.54 providing meaningful privacy guarantees while maintaining usable model performance. This represents a practical balance for medical applications where both privacy and accuracy are critical requirements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4.7 Comprehensive Discussion\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research set out to solve a real-world problem: can hospitals collaborate to build better AI without sharing patient data, and can this AI continuously learn new information without forgetting the old? The results provide a clear but nuanced answer. The system successfully proved that collaboration is possible and that privacy can be protected, but it also highlighted a critical weakness in managing long-term knowledge.\u003c/p\u003e\n\u003cp\u003eThe most significant finding was the severe catastrophic forgetting observed. When the model learned to identify new organs in Task 2, it completely forgot how to recognize the organs from Task 1. This is like a radiologist who learns to diagnose lung diseases but then completely forgets how to read a liver scan. In a clinical setting, this is unacceptable. It shows that while the federated learning machinery works for a single task, simply extending it to sequential tasks is not enough. The lack of any mechanism to rehearse or protect old knowledge led to a complete loss of previously valuable skills, underscoring that the challenge of memory is central to making this technology practical.\u003c/p\u003e\n\u003cp\u003eOn a positive note, the collaboration between hospitals was a success. The federated learning framework allowed three institutions with different specialties and patient data to build a shared model that was better than what any could have built alone. This is a crucial step forward, demonstrating that we can break down data silos in healthcare without compromising patient confidentiality. The differential privacy component further strengthened this, adding a mathematical guarantee that no individual patient\u0026apos;s information could be leaked, making the system trustworthy.\u003c/p\u003e\n\u003cp\u003eThe practical implication of this work is that it maps the path forward. We now have a working foundation for private, collaborative AI in medicine. The next essential step is to equip this foundation with a better \u0026quot;memory.\u0026quot; Future systems will need to incorporate techniques from continual learning, like periodically revisiting old data in a privacy-preserving way or identifying which parts of the AI model are most important for past tasks and protecting them. This research successfully created the stage; now we need to bring on an actor that can remember its lines from one scene to the next.\u003c/p\u003e\n\u003cp\u003eThis work demonstrates that the individual pieces of the puzzle: federated learning, continual learning, and differential privacy, can be assembled. The resulting picture proves that private, collaborative medical AI is feasible, but it also reveals that preserving knowledge over time is the most significant challenge that must be solved for this technology to finally deliver on its promise in real-world hospitals.\u003c/p\u003e"},{"header":"5. Conclusion and Future Work","content":"\u003cp\u003eThis research successfully created a system that allows hospitals to collaboratively train AI models on medical images without sharing patient data, demonstrating both the feasibility of privacy-preserving collaboration and the critical challenge of severe catastrophic forgetting. The findings establish that while federated learning with differential privacy works effectively, maintaining knowledge across sequential medical tasks remains the major obstacle for real clinical use. Future work will focus on integrating advanced continual learning techniques to reduce forgetting while preserving privacy, expanding to more complex medical tasks, and ultimately validating the system in real-world hospital environments.\u003c/p\u003e\u003cdiv id=\"Sec25\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Conclusion\u003c/h2\u003e\u003cp\u003eThis research successfully demonstrated that hospitals can collaboratively train AI models on medical images without sharing patient data through a working federated learning system. The implementation proved that differential privacy can protect patient information during this collaborative process, with privacy budgets maintained at practical levels. However, the experiments revealed a fundamental challenge: the system suffered from severe catastrophic forgetting when learning new tasks, with complete loss of previously learned knowledge about liver, kidney, and bladder identification.\u003c/p\u003e\u003cp\u003eThe findings establish that while the technical framework for privacy-preserving collaborative learning works effectively, maintaining continuous knowledge across sequential medical tasks remains the primary obstacle for real-world deployment. The system provides a solid foundation for multi-institutional medical AI collaboration but highlights the critical need for better memory preservation mechanisms in clinical environments.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec26\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Future Work\u003c/h2\u003e\u003cp\u003eSeveral important directions emerge from this research for improving federated continual learning in medical applications. First, implementing advanced continual learning techniques like elastic weight consolidation or experience replay could significantly reduce catastrophic forgetting while maintaining privacy guarantees. Second, developing adaptive differential privacy mechanisms that dynamically adjust noise levels based on task importance and learning stage could optimize the privacy-utility tradeoff.\u003c/p\u003e\u003cp\u003eThird, expanding the framework to handle more complex medical imaging tasks beyond organ classification, such as disease detection or segmentation, would test its generalizability. Fourth, exploring cross-silo federated learning with larger numbers of hospitals and more diverse data distributions would better simulate real-world healthcare networks. Finally, conducting real-world clinical validation studies would be essential to translate this research from experimental validation to practical clinical implementation.\u003c/p\u003e\u003cp\u003eThe promising results in collaborative learning and privacy protection provide a strong foundation, while the identified challenges with knowledge retention chart a clear path for future research to enable truly continuous learning AI systems in healthcare.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthical Declarations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthical Approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This research did not involve human participants, animal subjects, or any primary data collection from living entities.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests, financial or non-financial, relevant to the content of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors received no specific funding for this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthorship Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNnaemeka KIngsley Ugwumba: Conceptualization, Methodology, Software, Writing - Original Draft.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe author reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Declaration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analysed during this study, including the figures and source code, are available in the following GitHub repository https://github.com/KingsleyTechie/Federated-Continual-Learning\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eKaissis, G. A., Makowski, M. R., R\u0026uuml;ckert, D., \u0026amp; Braren, R. F. (2020). Secure, privacy-preserving and federated machine learning in medical imaging. Nature Machine Intelligence, 2(6), 305-311. https://doi.org/10.1038/s42256-020-0186-1 \u003c/li\u003e\n\u003cli\u003eKirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D., \u0026amp; Hadsell, R. (2017). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13), 3521-3526. https://doi.org/10.1073/pnas.1611835114 \u003c/li\u003e\n\u003cli\u003eKonečn\u0026yacute;, J., McMahan, H. B., Yu, F. X., Richt\u0026aacute;rik, P., Suresh, A. T., \u0026amp; Bacon, D. (2016). Federated learning: Strategies for improving communication efficiency. arXiv preprint. https://arxiv.org/abs/1610.05492 \u003c/li\u003e\n\u003cli\u003eRieke, N., Hancox, J., Li, W., Milletari, F., Roth, H. R., Albarqouni, S., Bakas, S., Galtier, M. N., Landman, B. A., Maier-Hein, K., Ourselin, S., Sheller, M., Summers, R. M., Trask, A., Xu, D., Baust, M., \u0026amp; Cardoso, M. J. (2020). The future of digital health with federated learning. NPJ Digital Medicine, 3(1), 119. https://doi.org/10.1038/s41746-020-00323-1 \u003c/li\u003e\n\u003cli\u003eSheller, M. J., Edwards, B., Reina, G. A., Martin, J., Pati, S., Kotrotsou, A., Milchenko, M., Xu, W., Marcus, D., Colen, R. R., \u0026amp; Bakas, S. (2020). Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data. Scientific Reports, 10(1), 12598. https://doi.org/10.1038/s41598-020-69250-1 \u003c/li\u003e\n\u003cli\u003eUgwumba, N. K., \u0026amp; Jaja, P. S. (2025). Causal-RAG: Causally-augmented retrieval for hallucination-free clinical decision support in low-resource settings. Preprint. https://doi.org/10.21203/rs.3.rs-8163345/v1 \u003c/li\u003e\n\u003cli\u003eUgwumba, N. K., \u0026amp; Jaja, P. S. (2025). Enhanced task prioritization system using Deep-Q-Network model. International Journal of Computer Science Engineering Techniques, 9(6), IJCSE\u0026ndash;V9I6P15. https://doi.org/10.5281/zenodo.17636107 \u003c/li\u003e\n\u003cli\u003eWeiss, G. M. (2016). OrganAMNIST. In MedMNIST Classification Benchmark (Version 2). Zenodo. https://doi.org/10.5281/zenodo.5208230 \u003c/li\u003e\n\u003cli\u003eYang, J., Shi, R., Wei, D., Liu, Z., Zhao, L., Ke, B., Pfister, H., \u0026amp; Ni, B. (2023). MedMNIST v2: A large-scale lightweight benchmark for 2D and 3D biomedical image classification. Scientific Data, 10(1), 41. https://doi.org/10.1038/s41597-022-01721-8 \u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 and 2 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Laskenta Technologies Limited","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"federated learning, continual learning, medical imaging, differential privacy, hospital collaboration, catastrophic forgetting, organ classification, privacy preserving machine learning, sequential learning, distributed artificial intelligence","lastPublishedDoi":"10.21203/rs.3.rs-8182728/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8182728/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis paper presents a novel Federated Continual Learning framework for medical imaging that enables continuous model updates across multiple hospitals without central data sharing. Our approach addresses critical challenges in healthcare AI: data privacy, catastrophic forgetting, and distribution drift. We demonstrate the framework on OrganAMNIST with three hospitals learning sequential tasks. Results show successful federated collaboration but reveal significant catastrophic forgetting, highlighting the need for advanced continual learning techniques in privacy-constrained environments. The integrated drift detection and differential privacy mechanisms provide a foundation for practical clinical deployment.\u003c/p\u003e","manuscriptTitle":"Federated Continual Learning with Adaptive Differential Privacy and Client-Side Drift Detection for Evolving Medical Imaging Datasets","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-01 07:42:58","doi":"10.21203/rs.3.rs-8182728/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"2ebaabe4-3c99-4688-8619-ac5d173b08b4","owner":[],"postedDate":"December 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":58697930,"name":"Hospital Medicine"},{"id":58697931,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-12-01T07:42:59+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-01 07:42:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8182728","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8182728","identity":"rs-8182728","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}