Federated Learning in Healthcare



1. Introduction to Federated Learning in Healthcare  

 

  

Federated Learning, AI in Healthcare, Machine Learning  

 

The recent electronic transformation in healthcare enables the collection of vast amounts of patient data, such as electronic health records, imaging, genomics, and data from wearable devices. However, the sensitive nature of health data halts the sharing and centralizing of data because of privacy concerns as well as legal frameworks like HIPAA in the United States and GDPR in Europe.  

 

FL (Federated Learning) is a new approach to solve such central challenges of sharing sensitive data. Machine Learning (ML) can be performed on the distributed data without the need for centralizing it, and FL enables training of models on decentralized data. Federated Learning in healthcare makes it possible for hospitals and clinics to collaborate in building advanced AI models without sharing sensitive patient data. Only model parameters or gradients are exchanged. This new approach makes it possible to train models in a way that privacy is maintained, and rich diverse patient data can be utilized.  

 

As highlighted in the McKinsey report of 2023, over 60% of US hospitals are pursuing FL AI applications, in particular, focusing on advanced applications in cancer diagnostics, COVID 19 forecasting, and the classification of rare diseases.

 

2. Technical Architecture and Workflow  

 

Healthcare Federated Learning is implemented in a client-server configuration. Each individual hospital acts as a client that starts with an onsite dataset and performs preliminary training on the model. Only the model derivatives are transmitted to the central cloud server, which performs aggregation and builds the global model with FedAvg or similar algorithms.  

 

Key Components:  

 

Local training: Each client trains a model instance on their proprietary data and optimizes it using SGD or Adam.  

 

Model Update Transmission: Partial updates comprising the gradients and the model’s weights are transmitted.  

 

Aggregation: A central aggregator builds the global model by averaging the received updates and other data received.  

 

Iteration: The global model is sent to the clients, who then further refine the training and return it to the aggregator.  

 

To protect model updates, Supplemental model update security mechanisms have been recently introduced, including Differential Privacy (DP), Secure Multi-Party Computation (SMPC), and Homomorphic Encryption (HE). These frameworks are implemented in Google’s TensorFlow Federated and NVIDIA Clara FL.  

 

3. Key Applications in Healthcare  

 

Federated Learning is penetrating new domains within medicine, facilitating collaboration toward the development of AI systems while ensuring data confidentiality and control over the information.

 

i) Radiology and Medical Imaging  

 

FeTS Challenge 2022 showcased the collaboration of over 30 institutions to teach cooperative AI for the task of tumor segmentation in MRI scans. The FeTS models yielded comparable results to centralized models, achieving over 85% in the Dice coefficient value.  

 

ii) Genomic Research  

 

In the domain of genomics, the branch has experienced the emergence of mutation prediction models, particularly associated with the breast cancer genes, BRCA1 and BRCA2, which have benefitted from FL. The iGenomics-FL architecture performed mutation prediction with an advantage of 15-20% over local models after model aggregation from several research centers.  

 

iii) Wearables and Remote Monitoring  

 

Currently, Apple and Fitbit are using FL to develop global models for the identification of arrhythmias, sleep apnea, and stress using data from wearables. This approach enhances the appropriateness of the models for individual users without compromising privacy.  

 

iv) Electronic Health Records (EHRs)  

 

Federated Learning allows the NLP models to automate patient data extraction from structured EHRs for chronic illnesses using data from diverse hospitals to train, as seen in the NLP model FederatedMIMIC, which improved AUC in disease classification tasks by 9% through collaboration.  

 

v) Rare Disease Diagnosis  

 

The use of FL in collaborative model development is crucial for small hospitals targeting advanced methods for early detection and intervention of rare diseases.

 

This form of collaboration is essential for efficient training of the model.

 

4. Issues and Solutions  

 

Due to the healthcare sector's operational and technical infrastructure, integrating Federated Learning (FL) poses various concerns.  

 

In healthcare, the data from within and between hospitals does not meet the IID (Identically and Independently Distributed) assumption. Factors such as demographic data and standardization of labels add volatility during the training stage. Efforts are being made toward such solutions as personalized federated FL and clustered FL.  

 

System heterogeneity within healthcare poses problems for FL. Different hospitals have distinct technologies and equipment. This creates challenges with FL traditional asynchronous training, edge computing, and resource-poor optimization strategies.  

 

Frequent communication to transmit model updates poses a risk of excessive bandwidth usage. There are ways to mitigate this to some degree, such as sparse updates, model compression, and reducing the frequency of model communication.  

 

FL poses vulnerabilities to model inversion, backdoor, and poisoning attacks. Protective strategies for these attacks have incorporated DP (distributed perturbation) noise, blockchain audit trails, and trust-based client selection.  

 

FL frameworks like NVIDIA Clara, IBM FL and PySyft have started to implement Kubernetes-based orchestration for scaling these frameworks.

 

These frameworks, however, struggle with the issues of smart aggregation and client selection with the fluctuation in the number of clients.  

 

5. Future Analytics and Forecasting  

 

By the year 2030, the healthcare industry is projected to grow with the adoption of Federated Learning to USD 290 million, increasing by 290 million dollors. Between 2024 and 2030, the Compound Annual Growth Rate (CAGR) is expected to be 46.3%, (Source MarketsandMarkets, 2024). This surge is anticipated because of growing AI adoption in healthcare, enhanced laws on privacy and data protection, and a greater need for private collaborative research.  

 

Promising Research Directions:  

 

Federated reinforcement learning for personalized treatment protocol recommendations.  

 

Cross-silo FL drug discovery collaborations between hospitals and pharmaceutical companies.  

 

Multimodal FL for triadic diagnosis with EHRs, images, and genomics in FL.  

 

Integrated BioGPT for federated abstractive summarization of medical reports within LLMs.  

 

In 2023, twenty European hospitals participated in a Federated Learning predictive analytics project to forecast ICU patient deterioration. The federated model achieved an AUC of 0.92, outperforming the best local model by 12%.

 

Conclusion  

 

In the context of employing AI technologies in the healthcare industry, FL, or Federated Learning possesses interesting opportunities since it preserves the privacy of delicate patient information. This form of machine learning can facilitate large-scale distributed collaboration in delicate healthcare as well as in scientific research. Federated learning can be used in the improvement of clinical practice and many other areas. Although there are many outstanding technical and policy questions, growing attention in the area of optimization algorithms, secure aggregation protocols, and cloud-edge orchestration systems is encouraging. FL will be one of the principal technologies needed to apply AI to the privacy-sensitive domain of healthcare.

 

 

 

 

 

Prepared by

 

Dr Balajee Maram,

Professor,

School of Computer Science and Artificial Intelligence, SR University, Warangal, Telangana, 506371.

 

 

Comments

Post a Comment

Popular posts from this blog

Setting a Question Paper Using Bloom's Taxonomy

    Setting a Question Paper Using Bloom's Taxonomy I) Understanding Bloom's Taxonomy The Blooms Taxonomy is another classification framework for educational aims in designing assessments that create opportunities for learning at higher or lower cognitive levels. Created by Benjamin Bloom and his colleagues in 1956, the model includes thinking at six levels, classifying it from basic to sophisticated levels: Remembering, Understanding, Applying, Analyzing, Evaluating, and Creating. The taxonomy is a guide for teachers in developing appropriate learning objectives and assessments. Each level represents a progression in cognitive complexity. For example, Remembering is the recall of facts and information, while Understanding emphasizes comprehension and interpretation. Applying requires using learned knowledge in real-world contexts, and Analyzing requires breaking down information into components to understand relationships. Evaluating involves making judgments based on criteri...
Read more

TIPS TO WRITE A SURVEY RESEARCH PAPER

                                    TIPS TO WRITE A SURVEY RESEARCH PAPER   I . Introduction Research serves as the backbone for advancement of knowledge as well as for the solution of real life problems. For an entry researcher, be it PhD student, is to know as much as one can on how much has already been written and studied on a specific topic area. A good research survey paper forms part of the necessary steps towards successful attainment of this aim. Synthesizing extant literature and summarizing the previous work helps identify gaps and suggests possible directions for further research. A survey paper in a research paper gives a general understanding of a particular field by gathering relevant studies and trends in an organized manner. Unlike original research, which actually addresses a specific problem, the objective of a survey paper is to walk the reader through the landscape of the ...
Read more
      Outcome-Based Education (OBE) in Engineering Colleges in India   1. Introduction to Outcome-Based Education (OBE) Outcome-Based Education (OBE) is a student-centered framework of education in which achieving predefined learning outcomes is more important. Unlike other education systems, which concentrate on what is to be delivered or memorized, OBE focuses on what students ought to know, be able to do, and value by the time they complete their education. It represents a shift in teaching and assessment paradigms, as success here is measured not by hours spent in class or coverage of a syllabus but rather by the attainment of very clearly defined objectives. OBE traces its origins back to global educational reforms in the 1980s and has since become prominent in countries such as the United States, Australia, and Singapore. It has been widely adopted as best practice in higher education across the world, particularly in engineering disciplines, due to its alignme...
Read more