Authors of a new study conclude that artificial intelligence could improve the selection of biosimilars for therapy.
Could machine learning (artificial intelligence) be used to help predict outcomes for specific patients treated with biosimilars? Authors of a review of studies on machine learning speculated on that potentiality and noted many other ways in which an adaptive computerized study of biosimilar use in the real world might be used to elucidate the value and qualities of individual biosimilars and inform providers to improve practice.
Because of the complex nature of biosimilars, they wrote, there are minute differences between them and their reference products that need to be more fully understood to build a practical understanding of these agents and their effects on patients. For example, patients might have immediate reactions to specific biosimilars or experience adverse events only after repeated treatments. Machine learning could track and interpret this information, although they said this relies on being able to “teach” computerized systems to interpret data collected across diverse health care systems with different protocols for recording patient data.
“Unwanted immune reactions may be induced for numerous reasons, including product variations. However, it is challenging to assess these unwanted immune reactions because of the multiplicity of causes and potential delays before any reaction,” the authors wrote.
It has long been accepted that the analysis of drugs should not end with the clinical assessments done to achieve FDA approval for commercialization. Follow-up should continue long afterward, especially with biosimilars, regulators and patient advocates have said. Due to their complex manufacturing and derivation from living organisms, biologic drugs are mutable and batches of drug will vary over time.
The FDA has recently championed the use of artificial intelligence technologies, which it says ”have the potential to transform health care by deriving new and important insights from the vast amount of data generated during the delivery of health care every day.”
Authors of the review said they hope to spur a broader discussion about machine learning with respect to biosimilars and include insight into maintaining patient safety amid the economic incentives that play a large role in which biosimilars get used. Biosimilars may offer discounts from originator products as deep as 30%; however, “economic incentives to use biosimilars and questions from the clinical community regarding unwanted immunogenicity can be incompatible,” the authors wrote.
To Learn and Adapt
“Machine learning, which is one type of artificial intelligence method, refers to programs that can adapt the instructions being used to produce the outcome. Here, the human aspect is its ability to learn and adapt the program’s code to experience,” the authors wrote. In particular, they suggested machine learning be used to identify risk factors for immune reactions, which they said could ensure safe biosimilar use while supporting their associated cost savings.
Machine learning, the authors explained, has the potential to detect important patterns or predict outcomes, but machine learning models require a learning process, just as humans do, to employ data in a way that enhances the accuracy of their predictions. The authors identified recent articles discussing machine learning applications in health care, but they noted that none have applied to biosimilar safety. Machine learning algorithms are a better fit for assessing immunogenicity than traditional statistical models, the authors suggested, because of the complexity and volume of the data involved. Immunogenicity assessments involve the characterization of human immune responses between biosimilars and reference products, in order to understand how patients might react to these agents.
The authors advocated for the use of real-world data to supplement clinical trial data on safety and immunogenicity. Although clinical trials measure safety metrics and antidrug antibodies, which may impair biosimilar performance, they “cannot provide a full picture of the safety profile,” the authors wrote, “because they are performed on a standardized population during a limited period.”
For purposes of informing and programming artificial intelligence systems, real-world data on biologic and biosimilar use and patient outcomes would be drawn from multiple sources, such as hospital systems and payers. Machine learning systems are best equipped to manage such a flood of data and help to make accurate and useful interpretations, the authors said. They suggested that “neural networks” would be capable of such functionality.
Using immune reactions as an example, the authors outlined a hypothetical neural network model for identifying risk factors for developing an immune reaction after a switch from a biologic reference product to a biosimilar. For a patient treated with reference biologic and then switched to a biosimilar, the model would consider the spectrum of medications the patient has received, the biosimilar prescribed, the reference biologic previously prescribed, drug batch information, and other data. The neural model would classify patients according to predefined categories of immune reactions and hone its power to predict an immune response for an individual hypothetical patient.
Ethical and Legal Issues
Proper handling of patient information is also a challenge of making a system like this work, the authors wrote: “Any machine learning application needs to comply with the ethical guidelines laid down by national ethics committees.” Complying with data privacy regulations in each country could be challenging, they said, but using aggregated and anonymized data would reduce the risk of violating those regulations.
One problem with neural networks, the authors wrote, is a lack of transparency, which may raise questions about how the machine learning systems arrived at their conclusions during the sifting of patient data. Neural networks use protocols that are layered one on top of another, and each of these folds involves a leap to a new level of interpretation whose workings may be hidden from clinicians trying to understand the reasoning process. “As a result, the black box might create conflict between processing by the neural network model and human interpretation,” the authors wrote.
In addition to immunogenicity studies, machine learning models could be used to automate pharmacovigilance processes, detect safety signals, make risk assessments to facilitate decision-making on dose regimens or switching, and compare switching initiated by a clinician vs substitution at the pharmacy, which could inform policy on substitution, the authors suggested. They concluded that machine learning technology “should be seen as a tool to learn more about biologic drugs after clinical trials and the medical practices during their uses.”
Perpoil A, Grimandi G, Birklé S, et al. Public health impact of using biosimilars, is automated follow up relevant? Int J Environ Res Public Health. 2020;18(1):E186. doi:10.3390/ijerph18010186