Important announcement: Missinglink has shut down. Click here to learn more.

Deep Learning in Healthcare Cover

Deep Learning in Healthcare

FDA Artificial Intelligence: Regulating The Future of Healthcare

Deep Learning (DL) has the potential to propel the healthcare industry into the future, with great experimental results and a variety of critical applications such as improved cancer diagnosis and medical screening techniques.

To prepare the industry for the future of healthcare without exposing patients to unnecessary risks, many regulatory entities need to make drastic changes to their frameworks to ensure they are better equipped to deal with Software-as-a-Medical-Device solutions.

This article will focus on the efforts made by the U.S Food & Drug Administration (FDA) to make the necessary changes to its framework and how SaMD solutions approved by the FDA have already started to shape the future of healthcare.

What Is Software as a Medical Device?

The American Food & Drug Administration (FDA) defines Software as a Medical Device (SaMD) as “software intended to be used for one or more medical purposes without being part of a hardware medical device.” In other words, it is software that is designed to diagnose, treat, cure, mitigate or prevent diseases or medical conditions.

SaMD as a medical tool includes mobile apps, in-vitro (IVD) medical devices and any software that can run on computing platforms that aren’t specifically made for medical purposes. While the definition includes even SaMD that works in combination or interfaces with other products, it excludes software designed specifically to run on hardware medical devices, such as an MRI machine.

Artificial Intelligence and Machine Learning in Medicine

Artificial Intelligence (AI), Machine Learning (ML) and Deep Learning (DL) technologies are more adaptable than other SaMD technologies. A SaMD based on these technologies is capable of optimizing device performance in real-time scenarios and keep providing better patient care at all times.

Medical applications for AI include:

  • Medical image analysis—identifies diseases and anomalies in medical imaging.
  • Cancer screening—identifies various types of cancer with a high success rate using deep learning Convolutional Neural Networks (CNNs), a type of Artificial Neural Network (ANN).
  • Diagnosing neurological conditions—helps doctors diagnose Amyotrophic Lateral Sclerosis (ALS) and other neurological diseases, and predict many neurological conditions such as Alzheimer’s prior to manifestation.

Learn more about AI and deep learning in healthcare and possible applications.

FDA Artificial Intelligence Regulation

The current approach the FDA uses to regulate traditional medical devices was not designed for flexible technologies such as AI and ML, mainly in terms of software modifications. ML and AI are highly dynamic technologies and the FDA forecasts that a driven by these technologies will require constant premarket review for software modifications.

FDA good machine learning practices

To prepare for the necessary changes to its regulatory framework for AI and ML-driven SaMDs,, the FDA has released its first policy document, a discussion paper, and Request for Feedback on April 2019 entitled: “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD)” or simply, the “AI Framework”.

In this paper, the FDA introduces a Total Product Life Cycle (TPLC) regulatory approach to AI/ML-driven SaMD. The FDA uses 4 key points to explain how they will asses the quality and organizational quality of companies to have reasonable assurances of high-quality SaMD development, testing and performance monitoring of products to have the assurance of safety and effectiveness through the TPLC:

  • Establish a definition of quality systems and good ML practices.
  • Demonstrate assurance of safety measures by reviewing SaMD products’ premarket submission. Additionally, establish clear expectations for manufacturers of AI/ML-based SaMDs on how to continually manage patient risks throughout the entire product lifecycle.
  • Count on manufacturers to monitor their AI/ML devices in all stages of development and during all algorithm changes. Incorporate risk management approach or any other approach outlined in this guidance document.
  • Increase transparency with reports of postmarket, real-world performance and maintain continued assurance of effectiveness and safety.

To successfully launch a TPLC regulation approach, the FDA will need to make changes, which will require different procedures and resources from its current regulations and provide solutions for many currently unanswered questions, including the following:

How will these changes affect pharmaceutical companies who market and offer AI/ML-driven software solutions?

This question is reinforced by the lack of explanation on the framework’s effect on AI/ML-driven software associated with a drug or biologic solutions and the absence of the Drug Evaluation And Research (CDER) and the Center for Biologics Evaluation and Research (CBER) from the discussion draft.

This is further complicated by a framework for prescription drug-use-related software published by these entities in December 2018, which doesn’t align with the current approach suggested in the paper by the FDA. This complication may lead to situations in which similar software is subject to different regulatory requirements depending on which entity is tasked with regulating the product.

How will the new AI framework address regulatory purposes and use-cases unmentioned in the paper?

The new AI framework doesn’t address many regulatory purposes and use-cases of AI/ML-based software. Examples include research and development, tracking post-market safety and the use of data gathered in the real world.

Does the FDA have the resources required to apply the new AI framework?

Introducing a TPLC approach to regulation requires many resources since SaMDs are frequently modified to apply software fixes and integrate new features. When the FDA was asked if they are currently equipped to deal with the workload of auditing AI/ML-based SaMD, the agency declined to comment.

To fully prepare themselves to audit AI-based medical models and algorithms and establish the necessary regulatory framework for modifications to AI/ML-based SaMDs, the FDA will need to rely on deep learning management solutions like MissingLink.

FDA Approved AI in Healthcare Applications

Many research teams are experimenting with AI/ML-based healthcare applications. So far, only 26 have received FDA approval as of May 2019. However, 14 of those 26 have been approved in 2019 and the list is growing rapidly. Here are some key examples:

Apple Watch ECG

The ECG technology used by the Apple Watch Series was approved by the FDA in September 2018. The technology, which uses 4 electors to detect irregularities in heart rhythms that might indicate heart failures, was the first consumer-available of its kind. The feature allows the Watch owners to take an ECG via their wrist that can also be used by physicians. For example, check out this story about how Apple Watch alert has potentially saved the life of a man.

ProFound AI by iCAD

This deep learning-enabled cancer detection app analyzes data gathered from several imaging techniques, such as 3D mammograms and breast tomosynthesis, to help detect malignancy. It was approved by the FDA in August 2019. Doctors use this app to improve the diagnosis rate, minimize false positives and reduce the time it takes them to screen for cancer.


Aidoc is an AI-driven app for analyzing radiology scans that received FDA approval in August 2018. It has already helped doctors diagnose over a million patients in more than a hundred hospitals worldwide.

After initial struggles with resource management and inflated GPU cloud services costs, The team at Aidoc started using MissingLink’s automated deep learning platform with great results and significantly scale up their business. MissingLink’s platform allowed Aidoc to automate and control their deep learning lifecycle, their core cloud infrastructure, and their experiment results.

In addition, Aidoc’s AI team can use MissingLink to view and control multiple experiments aggregated from multiple cloud servers via a single web page. Aidoc’s team can also queue and prioritize experiments, while MissingLink automatically manages deployment.

Manage Experiments with MissingLink

Running deep learning models to simulate the audits done by the FDA AI regulatory framework and in healthcare in general almost always involves intensive tasks such as training CNN models to scan and analyze large amounts of data from medical imaging. Running these models requires powerful hardware such as high-end Graphics Processing Units (GPUs), whether in-house with clusters or via GPU clouding services.

Running these models is not only challenging but also be expensive and time-consuming, especially at production scale. However, with a deep learning platform like MissingLink, you can easily manage, automate, schedule and record multiple experiments to save time and money.

Request a demo and see how easy it is to use deep learning in healthcare with MissingLink.

Train Deep Learning Models 20X Faster

Let us show you how you can:

  • Run experiments across hundreds of machines
  • Easily collaborate with your team on experiments
  • Reproduce experiments with one click
  • Save time and immediately understand what works and what doesn’t

MissingLink is the most comprehensive deep learning platform to manage experiments, data, and resources more frequently, at scale and with greater confidence.

Request your personal demo to start training models faster

    Thank you!
    We will be in touch with more information in one business day.
    In the meantime, why not check out how Nanit is using MissingLink to streamline deep learning training and accelerate time to Market.