FDA Siege | Debiased AI | Inpatient Screening

“We cannot allow unregulated AI to make clinical decisions.”

Hardian Health’s Dr. Hugh Harvey in response to the FDA’s move to waive regulatory approval requirements for 84 medical device classes, including many imaging AI solutions.

Imaging Wire Sponsors

  • Arterys – Reinventing imaging so you can practice better and faster.
  • Bayer Radiology – Providing a portfolio of radiology products, solutions, and services that enable radiologists to get the clear answers they need.
  • GE Healthcare – Enabling clinicians to make faster, more informed decisions through intelligent devices, data analytics, applications and services.
  • Healthcare Administrative Partners – Empowering radiology groups through expert revenue cycle management, clinical analytics, practice support, and specialized coding.
  • Hitachi Healthcare Americas – Delivering best in class medical imaging technologies and value-based reporting.
  • Nuance – AI and cloud-powered technology solutions to help radiologists stay focused, move quickly, and work smarter.
  • Riverain Technologies – Offering artificial intelligence tools dedicated to the early, efficient detection of lung disease.
  • Siemens Healthineers – Shaping the digital transformation of imaging to improve patient care.
  • Zebra Medical Vision – Transforming patient care with the power of AI.

The Imaging Wire

An FDA Siege

Last week ended with what some are calling a “siege” on the FDA that’s intended to make the independent regulatory agency a lot less independent…. or to make FDA regulations a lot less cumbersome (or both).

The Trump Administration is reportedly “forcing through a steady stream of changes in its final days” including:

  • Inserting an HHS leader into the top FDA legal role.
  • Setting a 10yr expiration on all FDA regulations that haven’t been reviewed.
  • Eliminating FDA 510(k) / PMA review requirements for 84 Class II medical device types, including a list of imaging equipment and imaging AI.
  • Requiring the FDA to publish how long it takes to review new drug applications.
  • Transferring oversight of genetically modified organisms to the US Department of Agriculture.
  • Setting term limits on FDA and CDC career scientists.

These changes have alarmed many in the FDA and healthcare community, including former FDA commissioner Scott Gottlieb, while the imaging community is trying to wrap their heads around what happens if premarket reviews are eliminated.

The Class II devices that would be exempt from premarket reviews includes a relatively wide range of imaging AI categories (e.g. lung CT CAD, image-based triage & notification) and imaging systems (e.g. C-arms, tomographic x-rays, MRI coils).

The HHS justified ending regulatory reviews for these devices by citing their clean adverse event histories, and its belief that the time / money spent on regulatory approval is too long / expensive. However, folks from the AI space were quick to point out that AI is too young to use historical evidence as reason to let unregulated AI make clinical decisions.

Debiased AI

Weill Cornell Medicine researchers developed an imaging AI approach that could reduce racial disparities in knee pain assessments and treatment, while potentially revealing a new way to remove bias from imaging AI (and imaging diagnosis in general).

  • The Background – Poor and minority patients have historically experienced more severe pain than their imaging would indicate. This is true for knee osteoarthritis, noting that the current knee OA severity grading system was developed and validated in a predominantly white population (different knee physiology and external factors).
  • The Study – The researchers used a deep learning model trained on knee X-rays from a more diverse population (20% Black, low income, or low education) to measure osteoarthritis severity linked to pain.
  • The Results – The algorithm predicted 43% of the patients who experience more pain than their traditional severity measurements would indicate (patients with “pain disparities”), while study radiologists’ traditional osteoarthritis severity grades identified just 9% of these patients.
  • The Osteoarthritis Impact – An algorithm like this could help get many patients into OA treatment that they would have otherwise missed out on if they were evaluated with traditional severity measurements.
  • The AI Science Impact – This study gives new evidence of how bias affects diagnosis (with and without AI) and how more diverse data can improve imaging AI’s diagnostic accuracy and equity.

The Wire

  • Signify on Optum & Change: Signify Research shared its take on Optum’s recent Change Healthcare acquisition, suggesting that Change’s imaging business has an “apparent lack of synergies” with Optum’s strategy, and detailing potential outcomes for the imaging IT businesses. Although there’s upsides to Optum keeping its new imaging IT business (its relatively independent, high share, growable, could support Optum’s network), Signify believes it is more likely that Optum will sell Change’s imaging business to private equity or a major OEM (maybe: GE, Siemens, United Imaging).
  • Inpatient Mammograms: A new Annals of Family Medicine study detailed an MGH pilot program intended to explore whether / how performing mammograms on targeted ICU patients could improve breast cancer screening participation. Each day, the program team identified patients who met certain criteria (MGH-affiliated PCPs, overdue for screening, not scheduled for a mammogram, insured by Medicaid/Medicare) and sent up to three emails to each qualified patient’s inpatient clinician asking them to consider mammogram screening before discharge. Over the course of a year, 39 of 48 clinicians (81%) responded to these emails and 17 of 21 “appropriate” patients (81%) received mammograms during their inpatient stay.
  • Hydrocephalus DL: A team from Qure.ai and Beth Israel / Harvard Medical developed a CT-based deep learning method to measure ventricular and cranial vault volumes for hydrocephalus evaluations, representing the first AI model focused on these measurements. It also represents an interesting expansion for Qure.ai, which historically focused on detection and triage tools. The DL model achieved high fidelity against manually-segmented lateral ventricular and cranial vault volume measurements (DICE 0.878 and 0.983, respectively). The study represents the first population-scale (13,851 scans) measurement of these structures, allowing the team to establish reference ranges for their ‘normal’ size and shape.
  • Another No to Shielding: The shift away from routine gonadal shielding gained more steam last week when the NCRP formally recommended its retirement, while outlining how providers and local / state regulators can finally end the antiquated practice.
  • Keya Medical’s $46M: Chinese imaging AI company, Keya Medical, continued its funding momentum with a $46M Series D round ($110M in the last year) that it will use to support its product development and global expansion efforts. Keya Medical’s flagship DeepVessel FFR solution performs functional coronary assessments using CTA scans, although the company is developing solutions for a range of specialties (cardiology, neurology, pulmonology).
  • Appropriate Savings: The Choosing Wisely-based “R-SCAN” appropriate imaging program could bring major imaging cost reductions with widespread adoption. That’s from a JACR paper that compared imaging costs from 27 hospitals that adopted R-SCAN against CMS data from other hospitals. They found that the R-SCAN policies cut these hospitals’ imaging costs by $260k over 3.5 months, which would equal $433M in annual healthcare costs reductions if extrapolated across the entire Medicare population.
  • Hyperfine’s AI: The FDA approved Hyperfine’s new Advanced AI Applications, which automatically measure and return annotated and segmented brain images captured with its Swoop portable MRI system. The standard software supports a range of measurements, including ventricular volume, brain extraction, brain alignment, and midline shift.
  • Med Student CXR Accuracy: A new study out Jordan found that sixth year medical students who completed at least one radiology rotation (n = 530) can detect life-threatening conditions on CXR with decent accuracy. The students were each shown seven chest X-rays (6 abnormal) and asked to choose the most likely diagnosis, with all students providing at least six correct diagnoses and 139 (26.2%) diagnosing all seven CXRs correctly.
  • How AI Is Helping: There’s a lot of talk about healthcare AI’s benefits mainly coming in the “future,” but it’s helping in a lot of ways right now. That’s from a new Nature Medicine paper that identified the following ways ML is helping medicine today: 1) Improving how we understand disease; 2) Testing research hypotheses; 3) Identifying patients to recruit for clinical trials; 4) Making more data available for other research; 5) Supporting diagnoses – you knew that one; 6) Predicting patient outcomes and guiding management; 7) Monitoring treatment progress; 8) Increasing interdisciplinary collaboration.
  • A PE Warning: A Current Problems in Diagnostic Radiology paper warned radiology trainees of private equity’s growing role in radiology and the downsides of working at private equity-owned practices (e.g. lower base compensation, clinical conflicts of interest). The paper also encouraged trainees to study-up on the PE financial model, while warning that private equity growth could “disrupt the practice of radiology and the competitiveness of future talent pools.”
  • Facebook & NYU’s COVID AI: Facebook and NYU Langone released a trio of open-source ML models to help in the COVID fight, which they pretrained with two public CXR datasets using the MoCo self-supervised method, and then fine-tuned with NYU’s COVID-19 dataset. The release is highlighted by a model that uses a series of CXRs to predict if a patient might need more intensive care up to 96hrs in advance (avoiding premature discharge, supporting resource planning), and also includes models that use a single X-ray to predict patient deterioration risk or predict whether patients will require supplemental oxygen.
  • COVID’s Screening Impact: A Brigham and Women’s study detailed the hospital’s rapid drop in cancer screening volumes (LDCT, pap, colonoscopy, prostate screening, mammography) during the peak of the COVID emergency, causing them to “miss” 1,438 cancerous and precancerous lesion diagnoses. They found that far fewer patients attended screenings between March 2 and June 2, 2020 but these patients had higher positive diagnosis rates (15,453; 1,985 with a diagnosis), than the subsequent three months (51,944; 3,190 w/ diagnosis), the previous three months (64,269; 3,423 w/ diagnosis), and during the same period in 2019 (60,344; 2,961 w/ diagnosis).

The Resource Wire

– This is sponsored content.

  • Workflow improvement, remote access, high image quality, and some major time savings. These are just a few of the benefits that cardiologists and their teams experienced with Hitachi Healthcare’s VidiStar platform.
  • In this Nuance Q&A, Inference Analytics CEO Farrukh Khan discusses how its NLP analytics AI solution improves radiologists’ reporting productivity through the PowerScribe One platform.
  • “Work-flow validation should happen very early in the AI development cycle…” Read on to learn about “Building Radiology AI that Works” featuring Babak Rasolzadeh, Sr. Director of Product & Machine Learning at Arterys.
  • Watch this recorded webinar from Healthcare Administrative Partners where they examine how the 2021 Medicare Physician Fee Schedule (MPFS) and Quality Payment Program (QPP) final rules will impact radiologists.
  • With a post-COVID surge in orthopedic surgery looming, this Zebra Med blog details how its 3D imaging orthopedic solution can help surgeons improve their preoperative efficiencies.

You might also like

You might also like..

Select All

You're signed up!

It's great to have you as a reader. Check your inbox for a welcome email.

-- The Imaging Wire team

You're all set!