Authored and contributed to PharmaLeaders by Michelle Marlborough, Chief Product Officer, AiCure, LLC
Machine learning (ML) innovation for clinical research holds great promise for improving drug development and patient care. But, like with all new technology, increasing prevalence can be accompanied by growing skepticism. Pharmaceutical sponsors and clinicians are understandably suspicious about the value of ML – they are used to working in a highly regulated environment that requires robust documentation around a drug’s development and its efficacy. They are wary of using ML-powered tools that provide little to no evidence as to how they draw conclusions, particularly when those conclusions may shape the future of a clinical trial or affect a patient’s treatment plan. To overcome this distrust, we need to provide the transparency needed to help end-users feel confident in ML’s ability to augment their work and believe in its transformative power in healthcare.
To address the increasing buzz around the need for transparency and standardization of ML, the FDA, in partnership with international regulatory agencies, developed Good Machine Learning Practices (GLMP) to assist with the standardization of safe and ethical ML models. While these recommendations are a promising start toward increasing adoption and trust in these solutions, there is an increasing need to take this a few steps further, and shift from guidelines to actionable processes and requirements that improve the visibility into how models work. However, before we can do so, we need to level our understanding – Why does transparency matter, what exactly do we mean when talking about transparency, and how can developers provide vigilant proof that sound processes were used to develop a model?
The need to enforce transparency to build trust
The ability of ML to analyze increasingly complex data points with stunning accuracy can fundamentally transform how we garner new insights and approach clinical research. This power relies on the assumption that a model is developed in a scientifically-sound, ethical manner – if not, the validity of its output is not only put into question but can also pose significant safety concerns. Just like we rely on the “Nutritional Facts” label on our food to make decisions about what we eat, sponsors and clinicians need visibility into the ingredients of ML. With very little oversight and visibility into how data is collected, how a model is trained and tested, what generates a specific output, and more, it is difficult for end-users to establish a basic understanding of how this technology may affect their patient population.
Transparency starts with redefining ML
Traceability should be an integral part of ML development to encourage transparency. This requires defining the basics – what aspects of a model can and should be traced? Contrary to popular belief, the proprietary nature of an algorithm only accounts for a small part of the structural foundation. Behind the curtain, there’s an entire system built around that small piece, which includes how data is chosen and collected, how that data is tested for relevance and accuracy, and how all that information works together to create a meaningful output that can be used to inform future research and care. In sum, the processes which make up the entire model are much more indicative of its success than simply reviewing an algorithm’s code. If we can better understand how a model works by gaining insight into how it functions, sponsors and clinicians may be less hesitant to integrate the technology into their research and practice.
Traceability in practice
Regulations around ML will continue to evolve as regulators balance the fine line between stunting innovation with enforcing practices for the greater good of the industry. We need to anticipate that one-day regulation may go beyond the advisory nature of GMLP, and developers will have to demonstrate the model’s effectiveness beyond a reasonable doubt, including that it was developed safely for the intended patient population. This could mean a “pedigree” or audit trail that illustrates the workflow of a system and the impact of each component. Additionally, as part of its ongoing maintenance once deployed in the real world, the technology must be continuously monitored and refined to continue meeting the intended population’s needs while also ensuring the developers are held accountable and its true performance aligns with the “evidence” provided. It is important to keep in mind that the average user of this technology is unlikely to have a deep understanding of data science. Whatever the reporting format, we need to ensure that the information put forward is easily understood.
Advancing responsible innovation
Beyond GLMP’s ten recommendations to help improve both the safety and transparency of ML in clinical settings, there needs to be more accountability and even incentive to accelerate ML innovation and adoption. Despite their lack of collaboration to-date, ML developers and regulatory agencies have a common goal – to improve patient care, and ease the burdens of sponsors and clinicians. By standardizing ML’s development and improving visibility into its innerworkings, we all get closer to achieving that goal.
About the author:
Michelle Marlborough, CHIEF PRODUCT OFFICER
As Chief Product Officer at AiCure, Michelle Marlborough is responsible for the product direction, definition and delivery of the company’s award-winning artificial intelligence platform. With over 24 years of experience in life sciences and software development, Michelle has been at the forefront of transforming clinical trials through innovative technology and analytics and previously held the positions of Vice President of Product Management at Veeva and Vice President of Product Strategy at Medidata Solutions. Before making the shift to the technology sector, Michelle worked in data management roles at GlaxoSmithKline and AstraZeneca. There she experienced firsthand the industry’s struggle with outdated processes and systems. Michelle earned her bachelors of science degree in biology and mathematics from Coventry University.