COVID has brought a new wave of investments products into healthcare, to springboard innovations leveraging advances in technology, powered by computational biology, artificial intelligence, and mobile platforms.
Industries’ investment into advancing healthcare technology is being met with parallel programs across FDA Centers, to support the advancement of innovative technologies. In our recent conversations with FDA, we spoke about their efforts to spearhead advancements on numerous fronts to enable smart pharmaceuticals and devices with an Ironman Jarvis like interface:
- Between medical products and the body,
- Modeling the body to accelerate product development, and
- Modernizing Total Product Lifecycle to monitor and accelerate Next Generation innovations.
Realistically – while synergistic – these three different future paths are intertwined and reinforce each other along the way. Each Center has invested in various public-private relationships and Collaborative Communities to work side-by-side on various initiatives to develop both voluntary and mandated programs and bring each of the paths from idea to reality. They are building programs and infrastructure to improve product quality, develop policies for novel technologies, modernize the Next Generation Computer Systems Validation guidance, explore continuous Artificial Intelligence, and accelerate solutions across products lines.
The keystream that flows between all three paths and the various collaborative initiatives is the digital data about the human body, to drive data driven ecosystem for science-based solutions.
Between medical products and the body
While FDA has written about how it would evaluate Medical Devices with self-adapting software capabilities (AI that “learns” the more patients it sees) – the type of AI it is currently comfortable approving is one where the model is trained during development, and then “locked into” the device before approval. The device may collect data from patients – but that data would need to be analyzed by the Manufacturer outside the individual patients. And any “modifications” to the artificial intelligence model on the devices, would likely need some form of efficacy and safety evaluation prior to release.
This means that any artificial intelligence/machine learning (AI/ML) solution – even if it is not controlling the primary components of the device (lasers, scanners etc.) must undergo rigorous clinical trial testing because the AI model fundamentally is layers of statistical analysis, which adapts based on the data, and thus must be tested to provide safety and efficacy and demonstrate error case handling, rather than just reasoning about error cases from the underlying physics or electronic simulations
Modeling the body to accelerate product development
Virtual human models allow for more effective product design, development, and testing. For example, FDA has utilized human models to test the impact of a next generation MRI safety on patients with defibrillators. As the number of patients with defibrillators increases, hospitals have a need for MRIs which will not displace the patient’s defibrillator. Prior to clinical trials virtual patient models were leveraged to assess safety and deem reasonable to proceed to clinical trials. While Virtual Testing is not a full substitute for actual In-Patient trials, it dramatically reduces both cost and the testing time since computer simulations can be scaled very quickly, time can compress seconds to milliseconds, and controls can simulate diverse anatomies to test a product impact across populations.
While the development of virtual human models is challenging, several models have been developed by the larger pharmaceutical and medical device companies. Development of these models is needed to accelerate innovation from start-ups and reduce their product development costs and timelines.
Modernizing Total Product Lifecycle to monitor and accelerate Next-Generation innovations
To-date AI/ML has been used to analyze customer and market data – This can range from something as simple as identifying which marketing messages resonate best, to detecting and identifying fraud in both the supply chain and sales chains for the devices.
Expanding the data set analysis to link customer and market data to events which occurred in operations such as a temperature change during shipment, or manufacturing events such as a change in materials supplier, to the design and development of the product, can improve safety and accelerate product adoption, and can enable continuous learning AI to become a reality in medical products.
Such initial work has been proven with the use of medical society registries. For example, the SVS VQI registry was used to accelerate FDA approval to expand stent Indications for Use. The ability to collect patient data from hundreds of hospitals allowed FDA to successfully monitor safety and efficacy and allowed the manufactures to gain their product data from production.
Expanding the data sets to collect the complex data from a products TPLC, can support product approval, allow manufactures to learn about their product performance with which to innovate next generation products, utilize the data to expand Indications For User without always requiring additional clinical trials, and reduce the cost for next generation product to the market.
The ability for innovators to interconnect the data from design and manufacturing, and demonstrate with evidence the ability to successfully monitor Real World Performance, requires a higher level of scrutiny, with the right level of data, to address both the model and error rates verification. It requires greater level of understanding AI/ML models and under what conditions to use a neural network model versus a vector model and what evidence is needed to ‘train’ and ‘validate’ the AI/ML model, let alone, to understand under what conditions ‘learning’ algorithm could be considered.
As the Computer System Validation next generation guidance moves through approval, FDA will enable companies to further develop and adopt the ability to generate, collect and analyze data across product development, manufacturing, and operations.
One example of AI/ML accelerating a portion of the TPLC, are the investments in manufacturing by investors such as Stanley-TechStar1, where investors are seeking innovate manufacturing. As the large manufactures are seeking to data tie across facilities for efficiencies, it could also create a more level playing field between small and large manufacturing companies.
Summary
All of these pathways require an understanding of AI/ML “On-Board” algorithms – Taking large amounts of examples of a particular pathology or symptoms or events, and “training” an algorithm to correctly
categorize. While currently FDA approves products that are ‘locking’ the algorithm, or the traditional methods of physics and rules-based software algorithm, the advantages of AI/ML to remove bias in the data processing and gain insights improves the scientific integrity with product development.
The future for both industry and FDA are the continuous learning AI/ML. Since continuous learning AI removes the direct data controls and proofs innovators need support to determine the mechanism to collect the right amount of evidence and the ability to describe AI development.
We recommend innovators work with Subject Matter Experts (SMEs) who have experience not just with AI/ML or with existing FDA regulatory frameworks, but specifically seek SMEs that are actively involved in the evolutions happening in both domains of expertise.
References:
https://www.techstars.com/accelerators/stanley
