Healthcare AI Models from Microsoft

Healthcare AI innovation with Zero Trust technology

Facebook Twitter LinkedIn

Posted on October 26, 2021

John Doyle Chief Technology Officer, Microsoft Health & Life Sciences

From research to diagnosis to treatment, AI has the potential to improve outcomes for some treatments by 30 to 40 percent and reduce costs by up to 50 percent. Although healthcare algorithms are predicted to represent a $42.5B market by 2026, less than 35 algorithms have been approved by the FDA, and only two of those are classified as truly novel.1 Obtaining the large data sets necessary for generalizability, transparency, and reducing bias has historically been difficult and time-consuming, due in large part to regulatory restrictions enacted to protect patient data privacy. That’s why the University of California, San Francisco (UCSF) collaborated with Microsoft, Fortanix, and Intel to create BeeKeeperAI. It enables secure collaboration between algorithm owners and data stewards (for example, healthy systems, etc.) in a Zero Trust environment (enabled by Azure Confidential Computing), protecting the algorithm intellectual property (IP) and the data in ways that eliminate the need to de-identify or anonymize Protected Health Information (PHI)—because the data is never visible or exposed.

Enabling better healthcare with AI

By uncovering powerful insights in vast amounts of information, AI and machine learning can help healthcare providers to improve care, increase efficiency, and reduce costs. For example:

AI analysis of chest x-rays predicted the progression of critical illness in COVID-19 patients with a high degree of accuracy.2

An image-based deep learning model developed at MIT can predict breast cancer up to five years in advance.3

An algorithm developed at the University of California, San Francisco can detect pneumothorax (collapsed lung) from CT scans, helping prioritize and treat patients with this life-threatening condition—the first algorithm embedded in a medical device to achieve FDA approval.4

At the same time, the adoption of clinical AI has been slow. More than 12,000 life-science papers described AI and machine learning in 2019 alone.5 Yet the U.S. Food and Drug Administration (FDA) has only approved a little over 30 AI- and machine learning-based medical technologies to date.6 Data access is a major barrier to clinical approval. The FDA requires proof that a model is generalizable, which means that it will perform consistently regardless of patients, environments, or equipment. This standard requires access to highly diverse, real-world data so that the algorithm can train against all the variables it will face in the real world. However, privacy protections and security concerns make such data difficult to access.

Breaking through barriers to model approval

As both an AI innovator and a healthcare data steward, UCSF wanted to break through these challenges. “We needed to find a way that allowed data owners and algorithm developers to share so we could develop bigger data sets, more representative data sets, as well as allowing [data owners] to get exposed to algorithm developers without risking the privacy of the data,” says Dr. Michael Blum, Executive Director of the Center for Digital Health Innovation (CDHI) at UCSF.7

With support from Microsoft, Intel, and Fortanix, UCSF created a platform called BeeKeeperAI. It allows data stewards and algorithm developers to securely collaborate in ways that provide access to real-world, highly diverse data sets from multiple institutions, where AI models are validated and tested without moving or sharing the data or revealing the algorithm. The result is a Zero Trust environment that can dramatically accelerate the development and approval of clinical AI.  ..... '