Resources

Resources for R and Health Economics

• Signposting Health Economic Packages in R for Decision Modelling (SHEPRD): Introduction | SHEPRD (hermes-sheprd.netlify.app)
• Decision Analysis in R for Technologies in Health: DARTH – Decision Analysis in R for Technologies in Health (darthworkgroup.com)
• R for HTA: R for HTA (r-hta.org)

Resources for AI in Health

• AI & Pharma: Generative AI in the pharmaceutical industry | McKinsey
• AI & Healthcare systems: Harnessing artificial intelligence for health (who.int)

Takeaways from the Talks:

All of the talks genuinely warrant a read through if you are interested in this topic (see resources provided at the end of this piece). I won’t give an overview of every talk, instead I will summarise some of my personal highlights and takeaways:

• HTA Agencies are Taking the Leap: with talks from HTA Agencies such as UK’s NICE and the Netherlend’s ZIN, their academic collaborators, and pharma company BMS’s submission experience, it feels like we are getting closer to R being accepted in certain HTA settings. More to do though as a collective to help see this through across settings, leading me to…
• The importance of Quality Assurance: something easier to do and more common in R models than MS Excel, but something we tend not to do beyond the usual (do the transition probabilities to 1 at any given stage? Does the outcome make sense with extreme values? Do the output plots from sensitivity analyses and scenario analyses make sense visually?).  A new package could help with this, even linking to GPT 4.0 to help visualise your R model in the process (“assertHE” with a youtube tutorial and all – R HTA 24 AssertHE RobSmith TomWard (youtube.com)). Also, keep an eye out for packages you use being maintained on CRAN as some may get decommissioned! (e.g. the “dampack” package).
• Object orientation: Linked to the above, and also based maybe more on discussions I had with attendees beyond the talks and more of a personal learning goal, there’s a movement towards more object oriented coding. This can help with quality assurance, linking with other software developing methods & something I’m going to try and do more of. Although, if you’re starting out, I would still “stick to the scripts”, i.e. functional programming. If this is all starting to read like another language in itself – it is discussed in Hadley Wickham’s Advanced R book: Welcome | Advanced R (hadley.nz) . If you are new – a plug to our previous papers on health economics in R: Health Economic Evaluation Using Markov Models in R for Microsoft Excel Users: A Tutorial | PharmacoEconomics (springer.com) & Extensions of Health Economic Evaluations in R for Microsoft Excel Users: A Tutorial for Incorporating Heterogeneity and Conducting Value of Information Analyses | PharmacoEconomics (springer.com) .
• Making Health Economics Sparkle using Shiny Apps: Technical tips: 1 or 2 speakers mentioned moving beyond hosting their apps on shinyapps.io, instead using AWS (amazon web service), css code in your Shiny App can be adapted to make it look nicer & use shades of a few colours rather than the rainbow. Beyond the technical: talk to stakeholders early and often, even create a specific fake person e.g. “Alice who works in marketing” and imagine her using your app, BUT draw the line somewhere as there will always be comments and sometimes they will not help the user interface, a balance needs to be struck between functionality and features.
• Machine Learning is Making Moves: The conference ended with a fantastic panel discussion on AI for health economics and outcome research. Technological advances are much quicker than the guidance generation so it’s difficult to publish consensus on how to use this, but the presenters had experience in linking R code for HTA with GPT models for various purposes. I asked the panel if there was a risk relying on private company ML models (upping price? stability?) & whether other MLs beyond GPT could/should be used? Answers – “yes” and “yes”. Claude (better) and Gemini were also given as MLs we could link our code to, and luckily the structure of these APIs is similar, so code which calls one into your code should in theory be easily adaptable across them. There we also brief discussion around how Python can now be executed in MS Excel & is what many AI programmes use, so the “R for HTA” umbrella should perhaps merge with a “Python for HTA” umbrella to form a “coding for HTA” canopy ? It reminded me of last years discussion by Devin Incerti on using Python for HTA – R for HTA 2023 Workshop – Devin Incerti (youtube.com) . Whatever you use, the advice was to break it down as much as possible (e.g. feed in specific functions rather than whole scripts), and give context, when using such ML models.

Equity Considerations:

Figure 2. Examples of Tech Discussed at LSX

Equity Considerations:

Across many of the sessions there was an underlying hum of “we need people to trust the tech” and how potential bias and ethical considerations need to be thought about in order to enhance this trust, across the tech pipeline. There needs to be consideration of potential model bias and ethics in training models, in patient groups used in trials, and then also equity of access to that tech (for example, not all patients may have smartphones). In terms of equity of access, there was also discussion about different reimbursement pathways – for example if there are higher out of pocket costs for consumers, this will likely be inequitable for those with less disposable income. There may also be legal and ethical issues linked to health data, such as abortion and associated women’s health data in the US, these need to be carefully considered.

There was discussion of synthetic control groups, and their potential benefits. This is where researchers identify a group receiving the new treatment (treatment group), analyse historical data from a large pool of similar individuals (potential control group), and using statistical methods, they create a synthetic control group that closely resembles the treatment group in all relevant aspects except receiving the new treatment. And while this approach does offer advantages, there are equity considerations; the quality and representativeness of the data used to build the control group are crucial and if the historical data reflects existing inequalities in access to healthcare, the synthetic control group might not accurately represent how the new treatment would affect a broader population.

So what can us future researchers do? Pfizer mentioned the use of “Ethicara” a bias detection service (an “AI risk assessment for responsible AI model development” – Program Schedule (goeshow.com) ), we also can strive to create and share more diverse data sets, for example (where possible) share data from different trials (with good GDPR practices, with “fair and proportional” data flows, of course). There was an interesting discussion about how people are more willingly now sharing banking data through apps etc., and what can we learn from this. The solutions proposed – promotion of health data literacy and the understanding of potential benefits to the data holders themselves. We need to make it easier for data to be collected in healthcare settings, but also to start building more holistic pictures by bringing in social determinants data where possible and relevant.

Tech can be used to help make it easier for patients to enroll, such as a Roche cancer trial app or Altoida giving patients tech to take home to take measurements outside of site. However, equity considerations on ease and ability of use need to be considered for the intended populations. Tech can also help reduce participant burden, for example Merck uses smart filters for their trial protocols to see what can be reduced/is really needed, and Roche has tech which helps employees feel like they are a patient going through that journey, and be more aware of the whole experience from the patient perspective.

Now, onto a theme I noticed throughout the conference – VALUE PROPOSITIONS FOR STAKEHOLDERS. I know this might read as self-serving (because it is) but I felt as though health economics could fill a huge hole in the health/med tech value proposition pieces. Often speaker answers to participant questions had some element of “you need to demonstrate value”, but value goes beyond initial estimates of ROI for investors, in order to get a good ROI, you need stakeholders (including patients and payers) to understand the efficiency implications on investing and/or up taking the tech…

This is where health economics can come in…

Health economic models can (i) highlight underserved populations and their current cost burden to healthcare systems that could be targeted and (ii) determine cost-effective pricing and potential populations, therefore providing key evidence in potential reimbursement returns. It can do this with uncertain data from literature reviews, trials and real world evidence. Value of information analyses can be done to show which parameters most impact cost-effectiveness results, and such parameters could even then be highlighted as key ones to collect in future observational trials. Beyond this, we need to start understanding the economic value of disruption, change management and wider behaviour changes as a result of tech implementation to have a wider value proposition for use in healthcare.

There were a lot of sessions on market entry of tech products, which of course is a key step in the journey of adoption. However, market access (as noted earlier), is key in having an ongoing ability to operate and sell the tech in a way that benefits all stakeholders, in that new market. Cost-effectiveness from the payer/provider perspective can be a key factor in a successful market access strategy, whilst understanding cost-effectiveness of scalability from the industry perspective can help ensure long term adoption.

A lot of tech companies tout the ability of increasing efficiency in the healthcare system from their product, but then do not investigate what this actually means in terms of money saved and increases in population health. Understanding these links, both in terms of direct cost savings (e.g. reduced unnecessary operations or drugs) and opportunity cost savings (e.g. more free hospital beds), should help further the adoption of truly efficient tech in healthcare.

In terms of incorporating equity impacts, extended cost-effectiveness can also be done, where the distributional effects of tech adoption and use, as well as financial risk protection, can be explored.

My advice to tech companies and payers would be to think about what the overall objectives are, and how these translate to the outcomes you want for any economic analyses. Stakeholders need to talk to each other in each different market. Some markets might be quality-of-life focused, some budget impact, some clinical improvements, some all of the above, which will require different data and different modelling approaches. The earlier health economics is integrated into tech creation, entry, adoption, integration, optimization and scaling, the more efficient and equitable these processes can be across healthcare systems.