FDA in the Age of Machine Learning: Big Data, Big Risks?

Advances in machine learning and artificial intelligence may now enable patients and providers to detects life-threatening illnesses earlier than ever. But much remains unknown about these rapidly evolving diagnostics. Can the FDA keep up while protecting public health in the era of big data?

Deep Learning, Medical Devices, and the FDA


The U.S. Food and Drug Administration (FDA) is charged with assuring the safety and effectiveness of medical devices. Although the FDA has substantial historical experience evaluating the risk of traditional devices (e.g., cardiac stents or knee implants), recent advancements in the field of machine learning and artificial intelligence have posed new challenges to the agency as it seeks to safely usher patients into the modern era of medicine.


The IDx-DR

            In April 2018, the FDA granted first-of-a-kind marketing clearance for the IDx-DR (IDx LLC, Coralville, IA, USA) device, a software program that uses an artificial intelligence algorithm to detect evidence of diabetic retinopathy, a disease that occurs when high levels of blood sugar lead to damage of blood vessels within the eye1. Diabetic retinopathy is the most common complication of diabetes and the leading cause of vision impairment and blindness in working adults. Physicians typically diagnose diabetic retinopathy by direct visual examination, though many providers may often lack the time or skill to discern manifestations of the disease.

The IDx-DR is intended to address this diagnostic gap by serving as a clinical decision support tool for primary care physicians. Based on analysis of retinal camera images obtained in the clinic, the device offers primary care physicians one of two recommendations: (1) refer to an eye professional due to concern for moderate-to-severe diabetic retinopathy or (2) repeat screening in 12 months.1 The FDA cleared the IDx-DR for marketing on the basis of a large clinical study (9963 retinal images from 4997 patients) benchmarking detection by the automated deep learning algorithm against manual grading by ophthalmologists.2 In this high-profile study, the algorithm had a greater than 96% sensitivity and greater than 93% specificity of detection.


FDA Regulation

            The FDA cleared the IDx-DR through the seldom-used De Novo regulatory pathway,1 which is reserved for medical devices without precedent. Given the compelling public health interest of commercialization, the FDA granted the IDx-DR “breakthrough device” status, which entitles the manufacturer to additional regulatory resources as a means to expedite clearance. As a result, the agency approved the manufacturer application for marketing within 3 months of submission. Building upon this early success, the FDA has recently designed several programs to help facilitate the evaluation of machine learning/artificial intelligence-based devices in the near and long-term:3

  1. Medical Device Tools Development Program ­– this program is intended to qualify the software tools that medical device manufacturers adopt in designing machine learning/artificial intelligence algorithms. By validating commonly used development methods, the FDA aims to accelerate regulatory review of such devices. 
  2. Real-World Data Analytics – the FDA is now seeking to leverage data captured within clinical data registries to validate and monitor the development of machine learning/artificial intelligence algorithms. The FDA is now piloting this capability in partnership with the American College of Radiology machine learning/artificial intelligence algorithms designed to detect cancerous lung nodules. 
  3. Software Precertification Program – the FDA recently designated nine health software companies – including “big data” firms such as a Phosphorus and Verily Life Sciences – pilot participants in the Software Precertification Program. This program is intended to reduce premarket requirements for qualified digital health manufacturers, thereby expediting device approval.


Predictive Pitfalls

            In recent months, the FDA has cleared additional machine learning/artificial intelligence devices, such as algorithms designed to detect breast cancer and intracranial bleeding. However, despite the undoubted potential of such technologies to advance patient care, the FDA should be mindful to safeguard against potential risks to patients, which may arise as the unintended consequence of biases introduced through missing data, insufficient sample sizes, and misclassification or measurement error.4 For instance, machine learning algorithms based on data from electronic health records may exacerbate disparities in healthcare by incorrectly inferring real-world inequities (e.g., less frequent prescription of cholesterol-lower agents in women with heart disease) to represent standards of care.

To mitigate against such risks, the FDA will need to develop the analytical capabilities necessary to avoid such pitfalls while navigating rapidly evolving technology. Doing so will require a substantial investment in internal capabilities through the recruitment and support of data scientists and other subject matter experts. Furthermore, the FDA should consider partnerships with private companies focused on specific therapeutic areas or technological approaches, all of which fall under the diverse purview of the agency. Medical specialty societies – which often collect rich clinical data – will likely prove valuable as well, as evidenced by the recent initiative with the American College of Radiology.


The Way Forward

            As the FDA looks ahead to the future of big data in medicine, it must grapple with several existential questions. How should the agency balance speed of access with certainty of patient benefit? How will regulation affect medicolegal liability for providers and manufacturers? How much can the agency collaborate with for-profit manufacturers while protecting the interests of public health? (word count 794)




  1. U.S. Food and Drug Administration. FDA permits marketing of artificial intelligence-based device to detect certain diabetes-related eye problems. https://www.fda.gov/newsevents/newsroom/pressannouncements/ucm604357.htm. Published April 11, 2018. Accessed November 12, 2018.
  2. Gulshan V, Peng L, Coram M, et al. Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs. JAMA. 2016;316(22):2402-2410. doi:10.1001/jama.2016.17216
  3. Allen B. The Role of the FDA in Ensuring the Safety and Efficacy of Artificial Intelligence Software and Devices. J Am Coll Radiol. October 2018. doi:10.1016/j.jacr.2018.09.007
  4. Gianfrancesco MA, Tamang S, Yazdany J, Schmajuk G. Potential Biases in Machine Learning Algorithms Using Electronic Health Record Data. JAMA Intern Med. 2018;178(11):1544-1547. doi:10.1001/jamainternmed.2018.3763


The future in cement manufacturing is here: Transforming Cementos Molins’ production process through the application of AI


Challenge.Gov – A Model for Government Crowdsourcing

Student comments on FDA in the Age of Machine Learning: Big Data, Big Risks?

  1. As part of the FDA’s mission statement, they claim “FDA is responsible for advancing the public health by helping to speed innovations that make medical products more effective, safer, and more affordable and by helping the public get the accurate, science-based information they need to use medical products and foods to maintain and improve their health.” (https://www.fda.gov/aboutfda/whatwedo/default.htm) However, they are often viewed by many in the medical industry as a deterrent to iterative medicine. With the mandatory 10 month review cycle, plus the obligatory pre-meetings before the meetings, it can often seem like the FDA acts as a filibuster, not “helping to speed innovation”.

    Although mobile medical applications (MMA apps for phones) are not new, the FDA still has not issued regulations for this space, but merely general guidance issued back in 2015. For example, if an MMA simply wants to push out a weekly update of its application, does it need to get FDA approval for a minor change to the algorithm? Does it need to prove this out through clinical trials that the update didn’t change the intent of the original algorithm? (https://www.fda.gov/downloads/MedicalDevices/DeviceRegulationandGuidance/GuidanceDocuments/UCM263366.pdf)

    I believe the FDA is trying to do right by its overall mission of protecting the public health, but I do think it is at a cost of speeding innovation in the digital world. I appreciate the speediness that they approved IDx-DR, but wonder if the same is possible for more complex digital applications, such as dosing insulin in Diabetes patients, or personalized treatment for cancer.

  2. Vinay,

    I enjoy reading your article, it features both sides – the huge opportunity and the potential risk – in an elegantly balanced way.

    I think one area that FDA needs to act more quickly on issuing guidance is companion diagnostics for immuno-oncology therapies. One interesting company is called Gaurdant Health, they use machine learning algorithm to solve the noise problem in ctDNA. They are able to identify specific gene mutation (EGFR, KRAS, NRAS etc) that are the leading cause of a number of cancers (non-small lung cancer in particular). They are not yet in early detection or prediction, but more focused on companion diagnostic for therapy choice.

    FDA is still no where in issuing guidance for relatively safe application of machine learning in diagnostics.

  3. I really enjoyed your commentary on the rapidly changing healthcare industry. I think it’s interesting to see that the FDA has a Digital Health Action Plan now, which outlines guidelines and collaboration opportunities for manufacturers.

    As a previous employee of a for-profit pharma manufacturer, I think there are opportunities to collaborate with the agency and operate with speed on implementing machine learning in healthcare solutions. My previous company recently struck a deal with Rimidi, a diabetes healthcare solutions company that currently has a portfolio of tech-smart diabetes products in development. While the patient unmet need has been quantified, the approval timeline could be significantly impacted if the FDA does not approve the data from the clinical trials. To answer your questions about how the agency can be involved with manufacturers of these products, I believe that agency representatives need to be brought in early to the process, and the manufacturer needs to provide a thorough education on the unmet need and potential risks of this project. Unfortunately, I think there is a long way to go before both sides feel comfortable approving these new digital products. My belief is that once there is a critical mass of digital products in the marketplace and a clear pathway for approval, then the process to protect public health through digital means will be accelerated.

    Source: http://www.pmlive.com/blogs/digital_intelligence/archive/2017/december/lilly_boosts_its_forthcoming_diabetes_digital_health_offering_1214767

Leave a comment