Use of Artificial Intelligence in Diagnostics and the Early Stages of R&D
This is the third part in a series dealing with advancements of artificial intelligence (AI) in the health care industry. In Part I, we looked at general uses and advancements of AI inside and outside the health care industry, some terminology, and potential opportunities and concerns related to AI. In Part II, we looked at some uses of AI related to clinical trials. Here in Part III, we’ll explore the use of AI early in drug discovery and the early detection of safety signals.
Drug Discovery and Research
For years, it has been theorized that the power of AI could be used not only to drive faster and more cost-effective development of existing drug compounds, but also to design such compounds.
That day has arrived.
On Jan. 30, 2020, BBC News reported that for the first time, an AI-designed drug will be used on human clinical studies. The drug, called DSP-1181, will be used to treat patients with obsessive-compulsive disorder (OCD). The molecule was discovered by British start-up company Exscientia and Japanese pharmaceutical company Sumitomo Dainippon Pharma. According to the BBC story, the development of DSP-1181 took only 12 months from discovery to first-in-human trial, whereas the normal time span averages almost five years.
As Exscientia chief executive Prof. Andrew Hopkins told the BBC, “We have seen AI for diagnosing patients and for analyzing patient data and scans, but this is a direct use of AI in the creation of a new medicine." i
The molecule was designed through the use of algorithms that sifted through a number of potential compounds while comparing those compounds against a large database of parameters. The identification of this compound using AI makes the type of AI discovery that experts have predicted for some time a reality by use of machine learning along with processor speed and access to large amounts of data.
Prof. Hopkins explains, “There are billions of decisions needed to find the right molecules and it is a huge decision to precisely engineer a drug.” ii
The importance of using AI to find new therapeutic molecules is emphasized by the fact that 90% of drug candidates fail between Phase I and final approval by regulators. iii AI can sift through more potential compounds faster than has ever been done before. Machine learning also allows machines to narrow down potential compounds faster, resulting in teams being able to focus on the best drug candidates, additionally speeding the process.
In the same manner in which new drug compounds can be developed, additional indications can be identified for existing marketed drugs by using AI.
In September 2018, then-FDA Commissioner Scott Gottlieb described how regulatory authorities are now trying to change the regulatory paradigm by better allowing for advancing technologies such as these, including new drug development and genomics.
“Early stage drug development is one of the most high-risk endeavors, marked by lots of dead ends,” said Gottlieb. “Yet the risk and time are also the factors that we can most easily influence if we have clear, modern, efficient policies to govern how we develop new innovations.” iv
Early Safety Signal Detection
In Part II of this series, we mentioned the increasing opportunities provided by AI for pharmacovigilance (PV), including safety signal detection. One advantage of incorporating AI into the PV process is detecting safety signals as early in drug development as possible. If you’ve been involved in the health care industry for long, you probably have unpleasant memories of the development of a drug that was halted, perhaps late in the process, because it was determined to present an unacceptable risk. By catching potential issues with a drug as early as possible, it allows the company to shift time and money to other promising treatments, which is in the best interests of the company and potential patients alike. Machine learning, with its continually adjusting algorithms, has the potential advantage of identifying molecules that pose an unacceptably high risk, as well as identifying compounds that may be better candidates for development.
Because pharmaceutical companies or CROs have finite resources to dedicate to a given clinical trial, AI has advantages of putting those resources to best use. As Dr. Michael Levy puts it:
“Instead of focusing on the resource-intensive manual and repetitive tasks of processing adverse event reports, Pharmacovigilance can now focus on more analytic tasks with greater potential to improve the lives of patients through improved benefit-risk assessment and risk management programs. In the future, we expect to analyze patient safety data more quickly and detect trends within larger volumes of data.” v
Dr. Levy emphasizes:
“Faster– ideally real-time – detection of relevant safety signals will contribute towards the optimum use of therapies and enhanced patient safety.” vi
In addition to faster results, is accuracy in diagnosis an advantage of AI?
According to some new information online, it is, however the full picture is still developing.
One particular comparison of diagnostic accuracy involving deep learning algorithms against readings by health care professionals in classifying diseases using medical imaging was recently reported. The comparison, with results published in “Lancet Digital” in September 2019, assessed the diagnostic performance of 31,587 studies between January 1, 2012 and June 6, 2019. The team conducting the comparison “found the diagnostic performance of deep learning models to be equivalent to that of health care professionals.” However, the comparison determined that the number of studies is limited in which results were externally validated or that compared the accuracy of deep learning models with that of health care professionals using the same sample. vii
As Part II of our AI blog stated, the opportunity exists for AI to mitigate an expected shortage of specialists in the coming years. According to the comparison described above in “Lancet Digital,” this applies to areas of diagnostics:
“The need for, and availability of, diagnostic images is rapidly exceeding the capacity of available specialists, particularly in low-income and middle-income countries. Automated diagnosis from medical imaging through AI, especially in the subfield of deep learning, might be able to address this problem.” viii
In addition, the comparison encouraged a clear-thinking assessment of the potential users, saying:
“Reports of deep learning models matching or exceeding humans in diagnostic performance has generated considerable excitement, but this enthusiasm should not overrule the need for critical appraisal.” ix
The comparison essentially concluded that, as with most undertakings in our industry, performing an analysis of risks against benefits when planning and executing health care projects is warranted.
To address AI-driven medical devices, including those used in diagnostics, FDA published the discussion paper “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD)” in April 2019. The paper provides some examples for “high-value uses,” including “detection, more accurate diagnosis, identification of new observations or patterns on human physiology, and development of personalized diagnostics and therapeutics.” x
Our industry is experiencing increasingly rapid changes in technology, and these changes can be used for the benefit of our patients as well as our industry. AI has numerous applications in the health care field, including the subprocesses of drug discovery, early safety signal detection, and diagnostics, particularly for those tasks that are otherwise repetitive and labor-intensive.
The recent use of AI to actually create a new medicine has charted new ground, and it’s very likely that we’ll see AI used to create more new medicines in the near future. It’s also important to remember that this new medicine was developed in about one-fifth the time as most medicines. The increased speed to market and cost savings that should be realized over time should be an advantage to our industry and our patients alike.
About the Author: Tony Ettwein, Former Sr. Manager at Pfizer, and Managing Expert for YourEncore’s Quality Center of Excellence, is one of YourEncore’s premier subject matter experts in IT Quality, increasing the effectiveness and efficiency of quality auditing by managing enterprise-wide computer validation quality and compliance teams.
i. “Wakefield, J. “Artificial intelligence-created medicine to be used on humans for first time,” BBC News online. Jan. 30, 2020. https://www.bbc.com/news/technology-51315462
iii. Fleming, N. “How artificial intelligence is changing drug discovery.” Gale Academic Onefile. May 2018. https://go.gale.com/ps/anonymous?id=GALE%7CA572639347&sid=googleScholar&v=2.1&it=r&linkaccess=abs&issn=00280836&p=AONE&sw=w
iv. Mockute, Ruda et al. “Gottlieb, S. “Harnessing the Curative Potential of Genomic Technologies” (speech). Sept. 28, 2018. https://www.fda.gov/news-events/speeches-fda-officials/harnessing-curative-potential-genomic-technologies-09282018
v. Levy, M. “How AI is Transforming Pharmacovigilance.” Pharma Boardroom online. https://pharmaboardroom.com/articles/how-ai-is-transforming-pharmacovigilance/
vi. Levy, M.
vii. Liu, X. et al, “A comparison of deep learning performance against health-care professionals in detecting diseases from medical imaging: a systematic review and meta-analysis,” Lancet Digital Health, Oct. 9, 2019, p. e271, https://www.thelancet.com/journals/landig/article/PIIS2589-7500(19)30123-2/fulltext
viii. Liu, X., et al, pp. e271-e272
ix. Liu, X., et al, p. e272
x. FDA. “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) - Discussion Paper and Request for Feedback.” April 2019. p. 2. https://www.fda.gov/media/122535/download