Medical Imaging Datasets for India (MIDAS) is a new platform for quality-graded health data for AI-enabled healthcare in India. MIDAS is a joint initiative by ICMR, IISc, and ARTPARK to create high-quality and standardized medical datasets representative of the Indian population. The researchers from the institutions developed the model to host gold-standard biomedical datasets from medical establishments in India. This is being implemented through a technology-enabled hubs-and-spokes system, where each hub institution works with a set of spoke institutions to gather medical and health data for a specific disease. 

According to the research write-up in the Nature Journal, the platform will automatically de-identify, anonymize, and harmonize heterogeneous data from multiple sources while ensuring individual privacy and confidentiality. The initial offering will be medical imaging datasets and associated demographics for AI and machine learning (ML) development. However, it will also host videos, electrocardiograms (ECGs), electroencephalograms (EEGs), pathology reports, and any other modality used in clinical practice, biomedical research, or those that may become available.

Human-in-the-loop systems

Most medical imaging datasets are retrospectively created from data collected during clinical care instead of proactively causing a disconnect between the purpose of data collection and AI applications developed using them. The human-in-the-loop machine-learning systems on MIDAS will provide semi-automated annotation and adjudication of the data. Initially, MIDAS will use retrospective data collection to formulate the guidelines for creating each specific dataset. Then, it will iteratively improve the prospective data-collection process by incorporating expert opinions and feedback from clinicians and data scientists. The versioned datasets will be in standardized, open, and interoperable formats, with enough samples from each demographic and setting to represent India’s unique diversity.

India-specific datasets

AI algorithms are developed using datasets that lack diversity and generalizability to other populations or ethnicities. As a result, such models trained on data from ethnic groups outside India are likely to underperform when deployed in India.

The researchers believe that the data sets in MIDAS would better represent the data from equipment deployed in the field. Many start-ups and research groups are reluctant to develop AI models because of a lack of data. Despite the dearth of medical imaging datasets, Indian start-ups such as Artelus, NIRAMAI, Qure.ai and Sigtuple are making breakthroughs in digital health-care. Such innovations could be accelerated by the availability of medical datasets through the MIDAS platform, which fosters competition in developing new AI algorithms and devices for improved health outcomes. The data sets collected could also serve to validate the performance of existing tools or new ones built for clinical deployment.

The research paper states that the MIDAS platform will enable access to curated medical datasets in a usable, standardized, and interoperable format for research and development stakeholders across academic institutions, start-ups, and corporations. Modular development with modern technology stacks and frameworks will enable the evolution of additional functionality and scalability to avoid obsolescence and siloing while automating workflows for seamless and user-friendly usage and adoption. MIDAS will provide incentive mechanisms for participating institutions to create high-quality gold-standard datasets. It will track and ensure compatibility with international best practices and enable commercialization by providing a platform for start-ups. The standards, guidelines, and protocols developed from this initiative will foster trustworthy and responsible use of medical data for AI/ML applications for better public health outcomes.

Nature