AI and Recent Regulatory, Policy and Funding Updates

While COVID-19 has been dominating the healthcare headlines so far in 2020 there have been a number of recent policy and regulatory decisions bearing on the AI/ML market in healthcare. We have been reading quite a bit about the broader slowdown in the market for AI/ML developers since the pandemic began that may be a temporary disruption to some parts of the market. On the other hand, CMS and the FDA have made some moves that may be part of building a stronger foundation for the market in AI applications, at least in clinical decision support and radiology/medical imaging segments of the market. We have also seen some major bets placed on AI-based companies that are riding on the wave of interest in telehealth since the pandemic began. So what might these signals mean for the overall market and adoption of AI/ML in the healthcare space? This post will try to tease out some patterns in the signals of late.

Key Takeaways

• CMS and the FDA have initiated policies over the last 6 months with potential to boost AI/ML market. The introduction of reimbursement codes and draft guidelines from the FDA that signal flexibility are creating more incentives for AI adoption in healthcare. The first reimbursement code for an AI application for screening for stroke care was introduced.
• Significant investments in AI-based remote monitoring applications coupled with new approaches for Software-as-a-Medical Device (SAMD) have emerged in the wake of the pandemic.  Virtual health adoption is driving utilization of some AI-based RPM devices and the FDA has demonstrated some flexibility since the beginning of the pandemic in how they approach regulation of bots for triage and algorithms used in remote monitoring. Automation of backend operations for providers is another area receiving a great deal of attention from investors as well.
• AI in healthcare is going to require new approaches to regulation in contrast to devices. Broader ecosystem understandings of use cases and initial approvals are only the beginning. Regulators are likely going to want continuous evidence of impact on outcomes and understandings of how users engage with devices which will demand a broader ecosystem approach to regulation.

Recent Policy Developments and Investments

A number of data points have emerged in recent months that are notable and include:

• In early August the Centers for Medicare and Medicaid Services (CMS) released notice of a proposed rule for coverage of CPT code 9225x that enables reimbursement for safe, efficient, and equitable use of autonomous AI in primary care settings
• In February the FDA released a white paper with their proposed framework for software as a medical device (SAMD) and a critical aspect of this is how initial approval of software is only a beginning and monitoring of outcomes will be ongoing
• Viz.ai received approval in September by CMS for a reimbursement fee of up to $1040 per use of their Viz LVO deep learning model that identifies signs of stroke on brain CTs and automatically contacts neurointerventionalists to speed up the pathway to effective treatment
• SoftBank Vision Fund 2 recently announced a series C investment of $100 million in AI startup Biofourmis. Biofourmis is in the digital therapeutics space and couples sensors with AI to monitor patient progress and effectiveness of various drug treatments. The investment is focused on supporting growth in digital therapeutic solutions for pain management, cardiology, oncology and respiratory health. The company now is split into two divisions: Biofourmis Therapeutics for partnerships with pharma, and Biofourmis Health, focused on remote monitoring in the home.

These developments along with a growing number of FDA approvals for AI/ML algorithms and software over the past two years marks a new stage in the AI/ML space, undoubtedly, but does it mean that the field is ready yet for experiencing dramatic growth as much of the hype indicates?

FDA History with AI and Legacy Issues

In 1998 the FDA approved computer-aided detection software for breast imaging and cancer detection (CAD). Shortly after the approval CMS increased reimbursement and the impact of the software was to increase costs by over $400 million/year without any substantial improvement in outcomes. In fact, cancer detection rates were actually worse at hospitals using CAD heavily. The experience of CAD is now weighing heavily on the approach to regulation that the FDA is considering for the current generation of AI products. There is growing recognition and pressure to adopt an approach that views regulation along a continuum where products are monitored after approval to see how they work on different populations and with different levels of experience across providers. The risk that some clinicians may lean too heavily on algorithms rather than clinical judgement as well as potential bias in original training data are all potential issues in downstream deployment. 

This is the context in which Viz.ai recently received approval for their blockage detection deep learning system for CT scans, ContaCT. In their documentation to CMS they noted the following impacts:

• Faster times for referrals to neurointerventionalist and to central hospitals for procedures
• The faster times mean less damage to brain cells affected by blockage and better outcomes
• For the outcomes they utilized Rankin scores, NIH stroke score, mRS at day 90 metrics that are standards in clinical care

The caveat here is that the sample size was only 43 patients with an additional 80 patients who experienced improved referral ties. This was out of a database with nearly 5000 patients where ContaCT was used. Most will find this sample size on the very low side and will definitely raise flags for going the route that the FDA is suggesting with utilizing continual observation to see if these outcomes hold across more diverse populations and more patients.

The Floodgates are Open Thesis

Writing in the Healthcare Blog recently, Luke Oakden-Rayner made a number of points on the meaning of the Viz.AI reimbursement of over $1000 per patient. His key points for the reason why the reimbursement is going to incentivize much more rapid adoption of AI are as follows:

• Oakden-Rayner notes that the $1000 reimbursement fee is a cap on reimbursement per patient and is for one year. It will likely be revised down in following years. ContaCT costs $25,000 per year so hospitals will need to scan at least 25 cases per year to recoup the cost of the software.
• The key point is that CMS tried to work with Viz.ai’s business model where AI ventures are very different business models from traditional devices. The reimbursement policy was a surprise for many analysts covering AI in healthcare and marks a new era with slowly emerging incentives for adoption.

I’m not convinced that we are anywhere near floodgates opening based on this decision. Yes, this is a welcomed incentive and will build confidence in AI business models but far more is needed for this sector to really take off in the current COVID-19 context where hospitals really need to see an ROI within a short time horizon.

AI in Virtual Health and RPM

Now to the SoftBank investment in Biofourmis. We have been covering the rapidly growing virtual care market since the onset of the pandemic and one of the accompanying tools is RPM, often backed by AI software to make clinical sense of the data collected remotely. AI is becoming an important tool in de-professionalizing some medical devices. That is, devices that once required a physician or nurse to operate can now be utilized by a patient remotely. AI plays the role of assessing the quality of the user engagement (eg. whether an image is good enough to analyze or an ultrasound image passes muster for a clinician). It also provides directions to the user to ensure the device is used properly.

As we’ve discussed in earlier blogposts, AI’s most common use is with symptom checkers and triage bots, particularly with the onset of the pandemic. The FDA has given greater latitude for some forms of remote assessment in ophthalmology, for example. A growing number of companies in this space couple wearables and AI for monitoring cardiopulmonary health for example (eg. Biofourmis). The FDA appears to be demonstrating flexibility when it comes to how they plan to regulate AI in this growing number of devices and apps that can be used in remote monitoring.

Automation of Backend Operations is A Growing Area in AI Arena

Another important investment data point is the rapid growth of robotic automation company, Oliveai. A rather young company in the AI and healthcare space that was established in 2017 has already raised $220m to date with a $51m investment right before the pandemic became headline news (from Drive Capital, Oak HC/FT, Ascension Ventures, and General Catalyst). Oliveai currently has over 500 hospitals as customers and focuses on automation of administrative tasks including revenue cycle management, supply chain management, clinical administration and human resources.

Going after the core administrative task burdens that are viewed as a deadweight on healthcare costs is becoming even more important in the post-COVID19 context as many providers and hospitals have had to cut administrative staff, yet these functions are vital to financial success. One of the critical components of their offering is to identify mistakes in billing codes that are a major cause of delays or lost revenue for providers and hospitals. 

Conclusion

Taken together the emerging market dynamics around virtual health, the introduction of the first CMS reimbursement codes for AI applications and the FDA’s flexibility are promising signs for the overall AI/ML market in healthcare. Nevertheless, it is still going to be a steep hill to climb for many of the vendors in the market that do not have robust evidence on outcomes and limited risks in the near future. While many have been expecting an impending “AI Winter”, the risks of this are still present, but some market indicators are providing support to the sector and we are seeing more investors coming to the table where AI vendors can also ride the wave of virtual health adoption.

See the following for more details on FDA guidance in the examples cited https://www.ebglaw.com/news/fdas-support-of-ai-in-telemedicine/