According to the FDA, precision medicine is an innovative approach to tailoring disease prevention and treatment that takes into account differences in people's genes, environments, and lifestyles. This approach allows to accurately predict which treatment will work in a particular group of people and leads to a massive shift in clinical development.
To successfully adopt precision medicine, businesses need to consider several factors, the three crucial ones being the complexity of data, the use of AI and ML, and regulatory compliance.
DataArt held a webinar, “Three Pillars of Clinical Development in Precision Medicine,” to discuss the key areas that early adopters should consider.
The panel of experts was moderated by Gregg Kravatz, VP of Precision Medicine at DataArt, and included:
- Slava Akmaev, Chief Technology Officer, Scipher Medicine
- Doug Shaw, Principal, DShaw Consulting LLC
- Yuri Gubin, Chief Innovation Officer, DataArt
In this article, we summarized the highlights of their conversation.
Complex Data
Gregg Kravatz: What are the most common challenges with data complexity and data quality companies face in precision medicine?
Slava Akmaev: To be successful in this field, you have to be able to acquire data, process it, put it into a centralized repository, and then maintain data governance and compliance. You need to make sure to become a reliable source of discovery, which means that you can acquire data, store it, and link it to the database in an efficient and seamless way.
We have been working on creating our own data lake for the last few years, and I think we are making tremendous progress there. This infrastructure allows us to be more efficient and innovative in identifying and developing novel algorithms and classifiers for predicting patient response.
Maintaining data quality is an enormous challenge. We acquire data through our clinical trials from partners, data brokers, EHRs, and EMRs, and the quality of the data varies. I would say a third of my team's effort is focused on data quality.
AI/ML Applications
Gregg Kravatz: How are AI and machine learning used in precision medicine?
Yrii Gubin: I would start by addressing three different questions: why we use AI, what we can do with AI, and how we do it.
Why We Use AI
AI is a tool that allows you to quickly tackle large data sets and run a number of experiments. We can tap into genomics, images, and data from clinical trials. We can analyze the sensitivity of the data to different parameters and run experiments around the segmentation of the user base, and now there are prediction techniques that allow you to improve the screening process.
AI is a tool to not necessarily find the only right answer but the tool that allows you to eliminate wrong answers faster because you can run multiple experiments simultaneously.
What We Can Do with AI
There are conventional capabilities like classification, clusterization, regression, and statistical analysis. Some techniques are typically under the hood of Generative AI, for example, vectorization of data, creating embeddings, and then building RAC, retrieval augmented generation, when you use subsets of the data grouped by different parameters. We can also do summarization or tokenization and even use generic abilities to do data mining.
The landscape is changing quickly; every other week, there is a new model or concept out there. This is why it is very important to continuously educate the team on best practices, blueprints, and use cases.
How We Use AI
There are studies by Harvard on a model that can predict lung cancer with a certain precision based on just one image. Some models can do the same from the prognosis standpoint. There is also an FDA-approved model that can determine the prognosis of kidney diseases. These are fairly mature products powered and driven by AI.
Regulatory Compliance
Gregg Kravatz: For these types of tasks, how do we handle regulatory issues and speed up the compliance process?
Doug Shaw: When we talk about the FDA, the old rule of thumb is if it is not documented, it did not happen. FDA is documentation-heavy, and they are evolving certainly in that regard.
One of the things I see my clients and colleagues need is a data strategy document. This document would determine what is needed from the data, including the target population, the features, the sort of cleansing required, how you split the training data and test data, and other details.
Another area I think we must consider in the life science industry is moving towards a proof of concept and doing it early on. What is the problem we are trying to solve? We need to develop correct match metrics to evaluate the model, think about data governance, and certainly start with risk management early.
Transparency is very important. You want to be able to see what is in the so-called “black box”: how the algorithm is working, how the data is being processed, and what parameters are being used.
You cannot independently review the algorithm and its capabilities, which ties to having a human AI team; the FDA mentions this in multiple documents. It is essential to have the skill sets we need today on that AI team. Existing skill sets include quality regulations, and new ones would include data science proficiency. Eliminating bias, reproducibility, security, privacy, and feature engineering are all important cross-industry concepts. We just have to figure out how to build this into the process within life science organizations.
Gregg Kravatz: How would we reduce the risk of not getting FDA approval based on the fact that AI can be viewed as a black box?
Doug Shaw: It is difficult to completely define something like an AI model in a specification, but I think it has to be explainable. In fact, this term is used more and more throughout. I see that in EMA guidance documentation and the FDA guidance documents as well: explainable AI.
There is a big focus on this. You need to make sure it is transparent, and your team understands what is going on. Maybe you cannot explain it fully, but you have to provide some documentation.
If you do due diligence to explain it the best you can, address and mitigate the risks, have the right team in place for a review along the way, and have the right data to train the model on, then you can achieve a certain level of comfort with what you have developed and the accuracy of your model.
Conclusion
Precision medicine is a dynamically evolving area that transforms clinical development while getting a boost from the latest technologies. Businesses that want to take full advantage of it must address data complexity, master AI and ML applications, and ensure regulatory compliance.
To get more expert tips on this booming field, request the recording here.
