October 2024 Update: Why Insurers are Dropping Medicare Advantage, AI as a Medical Device, and the Truth about Developer Productivity Using AI

Who is leaving Medicare Advantage (MA) and why?

If you’ve tracked MA plans this year, you’ll know that quite a few insurers have been dropping or cutting back on MA plans. Examples of insurers exiting or scaling back include: Premera Blue Cross (full exit), BCBS Arizona (full exit), BCBS Kansas City (full exit), Cigna (major reductions), Centene (exiting markets), Aetna (exiting unprofitable counties), Humana (exiting 13 MA markets), and Clear Spring Health (exiting state 2 of 5 states). Over 1.8 million people are in plans in 2024 that won’t exist in 2025. More than half of that member impact will come from Humana and Aetna. Reasons for exiting often relate to financial and regulatory pressures including: STARS calculations adjustments by the CMS resulting in lower ratings overall ratings for plans, which translate to lower bonus payments (e.g., 7 plans are 5-star rated in 2025, down from 38 in 2024), decreased MA benchmark payments for 2025, increased government scrutiny and audits seeking to recover overpayments to insurers, increased in-patient utilization, inflation, provider consolidation (which most agree drives up costs), and the sharp increases in utilization post COVID-19, which occurred in 2023 and continues. This all culminates in less money for insurers to operate their companies and draw a profit (i.e., higher medical loss ratios of 90% or greater).

• See this link for the official CMS changes to payments from 2024 to 2025:
  https://www.cms.gov/newsroom/fact-sheets/2025-medicare-advantage-and-part-d-rate-announcement

• See this article for a list of insurers exiting MA:
  https://www.beckerspayer.com/payer/5-payers-exiting-medicare-advantage-markets-in-2025.html

Is my AI or mobile app a medical device?

If you don’t want to read the paragraph below, the punch line is that the FDA offers guidance on AI and software as a medical device, the links are below, and AI product creators in healthcare and life sciences should be aware of the requirements…
Many are familiar with the story of 23andMe. They offered genomic testing for consumers and were shut down by the FDA for 2 years for offering unapproved tests. Emerging AI solutions often target healthcare, and some I’ve spoken to aren’t immediately thinking of the regulatory implications, so here’s a recap and links for more information. Government regulation of medical devices (and AI) is risk based. In the U.S. and Europe, for example, devices are classified in three levels, from low risk to high risk. Class 1, low risk, includes items like bandages and health information apps, and must meet a few requirements such as appropriate labeling and basic quality management. Class II, moderate risk, includes items like motorized wheelchairs and pregnancy tests (and Apple Watch), which could cause minor injuries, and therefore normally require registration with the FDA and full quality management systems, etc. Class III, high risk, includes items like artificial joints and pacemakers, which could result in major harm or death, and therefore require extensive controls including full premarket trials/testing and FDA/government approvals. The same classification scheme is used for AI. The FDA has provided specific guidance for Software as a Medical Device, that is, standalone software not part of a physical device but impacting health and safety. See below for guidance specific to AI. The good news is that the FDA classification is fair and reasonable. That being said, if your AI solution can have an adverse effect on health and safety, take the time to be sure you understand its classification and requirements.

• Brief read from the FDA with good examples of software regulated by the FDA: https://www.fda.gov/medical-devices/device-software-functions-including-mobile-medical-applications/examples-device-software-functions-fda-regulates

• FDA resources for AI and machine learning applied to Software as a Medical Device: https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-software-medical-device

• This next link is for the UK, but offers clear and simple examples and required steps for software medical devices (e.g., such as symptom checkers)
  https://www.gov.uk/government/publications/medical-devices-software-applications-apps

How much of a productivity increase should I expect when programmers use generative AI?

Of course, the answer is, “it depends,” but since that’s not helpful, following are details to give a sense of what’s possible. To begin with, as tools like GitHub Copilot became available at end of last year, most tech leaders I know attained about a 12%-25% productivity gain in pilots. Findings from more recent research varies, some examples include: a UC San Diego study found 48% and 63% less time spent on high and low complexity changes respectively. A GitHub study found developers using CoPilot were 55% faster in completing a programming assignment. CIO magazine reported that the Planview company achieved a 5%-20% productivity improvement and also a study of 100 engineers found a 70% reduction in time to update existing code and 32% reduction in writing new code. In contrast, a later CIO article claimed that an analysis of about 800 developers found no improvement in productivity from AI tools and rather found an increase in bugs. What gives?

The answer is that no universal productivity improvement range exists. Furthermore, performance measures vary between studies, some attempt to estimate total time in building a solution, some estimate time per code change, some count how many changes (i.e., pull request) are made in a day, and some how many AI code recommendations are accepted. Additionally, team and technical factors can wildly affect results. The following factors, if present, will reduce AI’s benefit: very large code bases requiring contextual knowledge to make changes, programming languages less commonly represented in the LLM training data, integration with proprietary products or industry-specific solutions, lack of developer training on effective use of AI tools, limitations of the specific AI tool selected, lack of prompt standards to generate production-level code, lack of proper LLM use to validate code and test, etc. To summarize, AI coding assistants do improve developer productivity when implemented effectively and used for coding tasks well-represented in LLM training data. It’s less a question of do they work and more a question of how effectively teams can put them to work. A 15-20% productivity improvement is certainly attainable in many cases.

• Following are links for the two CIO magazine articles referenced in case you’d like to contrast the results:
  https://www.cio.com/article/1312721/3-areas-where-gen-ai-improves-productivity-until-its-limits-are-exceeded.html

  https://www.cio.com/article/3540579/devs-gaining-little-if-anything-from-ai-coding-assistants.html

Next month I plan to dive deeper into how emerging AI tools are changing the way code is developed.

I hope you found this post informative. Reach out to me if you have questions or feedback.

Contact us

Whether you have a request, a query, or want to work with us, use the form below to get in touch with our team.