Potential catalyst

Over the past year, ChatGPT-4 and DALL·E 2, both developed by OpenAI, have garnered significant attention and praise. These are examples of "Generative AI," a form of Artificial Intelligence (AI) that leverages Machine Learning (ML) and Deep Learning (DL) to generate synthetic data that closely resembles real data. Generative AI has been employed in a variety of applications, from harmless endeavors such as creating images and inventing new video game levels to more contentious ones such as detecting financial fraud and passing the bar exam. However, the efficacy, ethics, and impact of Generative AI remain topics of debate and scrutiny.

It is widely agreed that the potential of Generative AI is enormous, provided that certain issues are addressed. Therefore, it is crucial to tackle these problems. Without discussing the hype surrounding the emergence of generative AI, Federated Learning (FL), and healthcare, instead, focus on the intersection of these technologies. The medical field presents numerous fascinating applications for Generative AI. However, the full potential of these applications can only be realised with the implementation of privacy-preserving methods such as FL.

The advancement of AI in healthcare is hindered by inadequate access to therapeutically relevant data gathered from diverse patient populations across various locations and updated over time. To ensure that all patients have equal opportunities for advanced AI-based diagnoses and treatments, it is crucial to prioritise patient privacy, diverse data sets, and trustworthy collaboration throughout the entire lifecycle of healthcare AI. Federated Learning (FL) enables this by facilitating secure and decentralised data sharing among healthcare providers.

Genomics research is leveraging Generative AI to create synthetic DNA sequences to assess the effects of different genetic variations in gene editing experiments. In drug design, Generative AI is being used to generate new molecular structures that conform to the principles of chemistry and physics and can be optimised for specific therapeutic outcomes. These structures can be simulated and assessed to identify the most promising candidates for a particular treatment. Furthermore, Generative AI is being used to produce large amounts of synthetic data for simulation and hypothesis testing in clinical trials. This approach can help accelerate the typically costly and time-consuming process of clinical trial design by generating data spanning diverse patient groups, treatment outcomes, and adverse events.

Generative AI finds use in medical imaging by generating synthetic images with desired characteristics not found in the original dataset. These images supplement existing datasets and aid model development. Furthermore, Generative AI transforms images between modalities and generates images from text inputs. Personalised medicine creates customised treatment programs based on patient-specific features such as diet and exercise regimen. Generative AI serves as an educational tool for medical students to refine their diagnostic and treatment planning skills by practicing with computer-generated patient scenarios, akin to creating new levels in a video game.

Although Generative AI testing has exhibited potential, its outcomes have not been widely applicable. This is because the initial models were not designed with healthcare professionals in mind. FL may be a solution. A significant obstacle to the development and implementation of AI/ML/DL in healthcare is the lack of large, diverse datasets. Models developed in one location may not be transferable to another due to differences in patient demographics, hospital practices, data formats, or even image creation tools.

To achieve true success, ChatGPT-4 and other Generative AI models require extensive training on diverse patient data from a variety of healthcare institutions, including images, provider notes, and outcome data. Healthcare data controllers must balance data sharing with privacy concerns and regulations, limiting sharing to snapshots of patient journeys. Although comprehensive multimodal data is valuable for AI/ML/DL development, there is a higher risk of re-identification.

The challenge of data diversity in healthcare AI/ML/DL can be addressed through FL. Generative AI models can augment existing data sets with synthetic data, but the resulting models may still exhibit bias if the training data lacks sufficient diversity. FL offers a solution by delivering the model to the data instead of centralising the data. This allows developers to access the full range of multimodal data while giving data custodians control over data usage and the ability to opt-in or out. By leaving the data at rest on hospital servers, information security teams can also feel more comfortable with this approach. FL effectively cuts through this "Gordian Knot" and allows for the development of more robust AI/ML/DL models in healthcare.

Similar to working with centralised data, utilising an FL platform enables data analysis, AI model development, and deployment without the necessity of transferring or relocating the data. This accelerates access, reduces risk, and generates more outcomes and benefits that can be broadly utilised. By incorporating FL and edge computing, the approach facilitates collaboration among healthcare organisations, researchers, and the industry without compromising patient privacy, data mobility, or ownership transfer. Consequently, it enhances healthcare and life sciences outcomes while reducing costs and simplifying implementation processes.

The FL, distributed computing, and edge computing enabled hospital-hospital, hospital-industry, and industry-industry collaborations. Different FL models include DL for cancer diagnosis, novel privacy-preserving methods, picture quality assessment, and hospital outcomes evaluation. These are just a few of the numerous effective FL and edge computing applications.

In FL, data remains in its original location, ensuring the confidentiality of patient information. Each site receives a copy of the AI model, and local training is conducted on-site, eliminating the need for data duplication or massive data transportation. By leveraging aggregate learnings, an optimised model is created. This approach harnesses the power of diverse data to develop AI-based healthcare solutions that can rapidly spread across international borders. Consequently, the models operate more reliably and enhance the standards of medical treatment. FL's rich data accessibility empowers generative AI model developers to unleash the full potential of these remarkable tools. This is achieved by training them on comprehensive and diverse samples, while simultaneously ensuring hospital infosec teams can rest easy. While ChatGPT-4 showcased the potential of Generative AI, FL is assisting healthcare AI developers in realising it.

The writer is an HoD and Assistant Professor of Dept of Computer Sc & Electronics, Ramakrishna Mission Vidyamandira. Views expressed are personal