One of the principal objectives of clinical studies is to test the safety of drugs and ascertain potential Adverse Drug Reactions (ADRs) in the study population. However, it is only feasible to study a few hundred or thousands of participants at any given time. Furthermore, these studies are performed under controlled clinical conditions that may not represent every real-world situation or circumstance. The limitations of these traditional approaches do not allow for all ADRs to be identified before a drug is made available on the market, thus mandating the need for drugs to be constantly monitored post-launch. According to the World Health Organisation (WHO), “the long-term monitoring of existing drugs is crucial, because potential ADRs, interactions, and other risk factors, may only emerge many years or even decades after the drug initially received market authorization”.
Challenges to the Traditional Pharmacovigilance Landscape
Pharmacovigilance (PV) is dedicated to the long-term monitoring of drug safety beyond market authorization, and is defined by the WHO as “the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or any other medicine-related problem”. Every biopharmaceutical company has a legal responsibility to collect, process, and report details of Adverse Events (AEs), and other product safety information to healthcare regulators. It is common for patients to report the incidence of AEs to their healthcare practitioner, who then informs the relevant manufacturer, and subsequently, the regulator.
The COVID-19 pandemic raised public awareness of the importance of biopharmaceutical innovation and drug safety, and specifically the need for reporting the side effects of the newly developed vaccines. The US Food and Drug Administration (FDA), for example, received over one million COVID-19 vaccine AE reports in the first year, while the European Medicines Agency (EMA) received a million reports from the five vaccines used over two years (based on 868 million doses). Changes in public awareness, combined with changes in technology are demanding innovative approaches to drive forward PV processes and AE analysis.
The COVID-19 pandemic was undoubtedly a catalyst for change as it brought to the fore changes that were already beginning to impact the traditional PV landscape due to:
Evolving disease complexity: Growth in new product portfolios and therapy areas has led to a surge in global databases, leaving the biopharmaceutical industry saturated with data.
Complex regulatory requirements: Healthcare regulators are incorporating processes and standards that explore Real-World Evidence (RWE) for signals and drug safety information.
Increasing cost burdens due to rising safety reporting volumes: The processing of Individual Case Safety Reports (ICSR) is a time-consuming and costly process as it requires significant resources across manual workflows. In 2019, Pfizer processed approximately 1.4 million AE reports globally. In 2021, GSK, estimated a mean cost of US$33 per processed case report, while other pharmaceutical companies approximated their costs to be even higher.
The exponential rise in safety reporting volumes is placing an increasing cost and human resource burden on biopharmaceutical companies.
Exponentially rising number of data points: The International Society of Pharmacovigilance (ISoP) has highlighted the exponential rise in ADR-related data inflow from a variety of sources that include patient/healthcare professional AE reporting; RWE studies; data generated from connected medical devices; and information derived from social media platforms that are increasingly being used to share personal health-related insights.
Other sources such as spontaneous reports, chatbot interactions, Electronic Health Records (EHRs), emails, published literature, patient registries, claims data, and post-approval safety studies are also contributing to a surge in data. Data published by the US FDA in 2021 indicated that over 2.2 million AEs were submitted to the regulatory body’s Adverse Event Reporting System (FAERS), up from 500,000 in 2009. This highlights the burden not only shared by the regulatory agencies, but also by the industry in terms of rising ICSR processing.
Applying AI and ML to pharmacovigilance
PV is fundamentally a data-driven field that requires the collection, management, and analysis of data gathered from a wide range of disparate sources. The central challenge for PV is the ability to extract meaningful insights from large volumes of heterogeneous data, quickly and reliably to find safety signals that require attention. To aid this process, Artificial Intelligence (AI) powered by advancements in Machine Learning (ML) is being used to optimize processes and find solutions to many of the core problems confronting PV.
According to Gosh et al. (2020), current areas of AI activities in the field of PV include “digital media screening; extracting and classifying data from source documents; checking for duplicate reports; case validation, triage and initial assessment; data entry; medical assessment, including causality; narrative writing; and coding AE concepts into standardized terminology”.
In general terms, AI/ML has the potential to impact, and is being explored across three broad areas of PV activities that include data ingestion-related tasks, signal detection activities, and ‘other’ applications that include analysing social media posts and integrating data sources with multiple analytical approaches to gain richer insights.
Case intake or processing:
According to Ball et al. (2022), an ICSR contains “information on the patient, the AE, the suspect medical products, the reporter, and, for ICSRs submitted by industry, information on the company that holds the application or license for the drug”. Providing complete and accurate reporting of ICSRs is critical to the understanding of a drug’s safety profile, and as such, applying AI/ML to improve the efficiency and scientific value of ICSR case reporting is a major area of interest for the biopharmaceutical industry.
AI and ML applications are well suited to extracting and organizing the information contained in ICSRs that take the form of structured, semi-structured and unstructured data, all of which are needed to make a clinical assessment. Once extracted, this catalogued information can then be submitted to a drug safety specialist for review and confirmation, or correction. This ‘augmented intelligence’ serves as input for subsequent rounds of ML to be applied to AE case processing, which in turn, serves to improve the algorithms, increase consistency, and reduce errors. Such applications have the potential to significantly enhance the efficiency of the current PV operating model by automating the routine and manual tasks associated with PV, decreasing overall costs. With case volumes growing at a rate of 10–15 percent per year, a survey published by Deloitte (2019) noted that 90 percent of companies working with PV are looking for solutions to cut the costs of case processing through automation.
Signal Detection Activities: The WHO defines a safety signal as “reported information on a possible causal relationship between an adverse event and a drug, of which the relationship is unknown or incompletely documented previously”. The practice of signal detection is typically targeted at marketed drugs where a safety concern may be new in nature (having not been identified previously in pre-marketing phases), or is a novel pattern of already known ADRs (relating to factors such as severity, high-risk sub-populations, and frequency). Signal detection involves the analysis of a wide range of sources that include case-by-case assessment of ICSRs, or the mining of pharmacovigilance databases comprising Real World Data (RWD) derived from clinical trials, EHRs, biomedical literature, and spontaneous reporting, to name a few.
The databases may be owned by a pharmaceutical company, a drug regulatory authority, or a large healthcare provider. For example, VigiBase, the WHO global database of medicines-induced reported side effects contains reports of ADRs submitted by national pharmacovigilance centres participating in the WHO Programme for International Drug Monitoring (WHO PIDM) whose purpose is to ensure that early signs of previously unknown medicines-related safety problems are identified as rapidly as possible. Vigibase is the largest database of its kind in the world, with over 30 million reports of suspected medicinal adverse effects submitted by member countries of the WHO PIDM. Sophisticated ML techniques, such as Lasso shrinkage regression, Bayesian borrowing algorithms, and temporal scan statistics are being applied for signal detection.
Social Media Monitoring
With the emergence of Web 2.0 and social media platforms, there has been a significant rise in user-generated content in the form of blogs, forums, and social media posts that are published on the Internet. Concurrent with this, advances in AI technologies are giving rise to powerful methods and algorithms that led to the creation of Natural Language Processing (NLP) techniques, which enable the processing and understanding of human-generated text. Text mining, which is generally defined as the analysis of textual data (unstructured or semi-structured text), emerged from the need to analyse large amounts of text containing human language.
The principal task of signal extraction from social media “consists in the identification of drugs and symptoms (entity recognition) and of the relationship between them”. Since more patients use social media platforms to express their responses to medication regimens, this medium has become an important source of information for biopharmaceutical companies and regulators in recent years. Although colloquial terminology, including acronyms, emojis, hashtags, images and slang, together with duplicate reporting (such as parallel posting on multiple platforms), makes it challenging to mine social media data, NLP enables the analysis of multi-formatted AERs to extract meaningful health insights. Novartis, for example, developed AE Brain which analyses and processes social media posts to identify potential AEs, and based on the content and context, predict whether they contain reports of AEs. AE Brain processes around 15,000 messages per week, thereby improving the quality of safety information derived by the company, and reducing the burden of manual repetitive work.
Looking ahead
Advances in AI/ML applications are presenting biopharmaceutical companies with opportunities to improve their overall PV capabilities and achieve investigation maturity through investment in next-generation digital learning systems that can increase efficiency and generate richer insights. Given that risk minimization measures can be initiated faster, and with greater accuracy, the application of AI/ML in PV has the potential to further promote and protect the health and well-being of patients, which is the overall objective of a successful PV strategy.
