OR WAIT null SECS
© 2023 MJH Life Sciences™ and Pharmaceutical Technology. All rights reserved.
The ability of artificial intelligence to process large amounts of diverse data cleanly creates more accurate reports for improved drug safety.
The World Health Organization defines pharmacovigilance as “the science and activities relating to the detection, assessment, understanding, and prevention of adverse effects or any other medicine/vaccine related problem” (1). Drug safety includes quality control during the development and manufacture of a drug and monitoring the safety of the drug post-approval when it is in the market and being used by patients.
The European Medicines Agency (EMA) states that European Law requires a pharmacovigilance system be operated by EMA, the national competent authority, and the market authorization holder to monitor medicines throughout their use for adverse side effects that may occur (2). In the United States, FDA’s Adverse Event Reporting System database support’s the agency’s post-marketing safety surveillance program and contains adverse event reports, medication error reports, and quality complaints. Reports sent to FDA by drug manufacturers and healthcare professionals are used by the Center of Drug Evaluation and Research and the Center for Biologics Evaluation and Research to track the safety of a drug (3).
“Pharmacovigilance activities start with clinical trials to provide information on the safety and therapeutic benefits of a drug. The aim of [pharmacovigilance] monitoring during the clinical trials is to demonstrate that the therapeutic benefits outweigh the risk (severe side effects), and if proven, the respective data is submitted to regulatory agency (i.e., FDA) to gain approval to commercialize the drug,” says Dr. Foram Vaishnav, head of the North American pharmacovigilance and risk management team at Dr. Reddy’s Laboratories.
Gathering and reporting safety data to regulators is part of the responsibility of drug sponsors to ensure the safety of the products they manufacture and distribute. “Each sponsor is legally required to manage adverse events based on specific guidelines by territory. A sponsor with authorization to run clinical trials for a product will need to collect and report all adverse events per study protocol, with a subset requiring expedited reporting,” says Sharmila Sabaratnam, senior director, Vault Safety strategy at Veeva. “The most considerable volume of adverse event reports occurs post-market and can present challenges if sponsors don’t have the right systems, processes, and talent in place. In pharmacovigilance, you need this foundation to ensure patient safety, scientific rigor, and data integrity.”
Use of artificial intelligence (AI) in the monitoring and reporting of drug safety data is a tool companies can utilize to assist them in this process. “AI assists companies by automating information capture, yielding quick output, limiting errors, and helps to understand and classify clinical documentation,” says Aloka Srinivasan, PhD, principal and managing partner at RAAHA LLC.
“With the evolution of safety technology, AI is more commonly leveraged in case processing for intake, validation, and coding to assist case processors or automatically process cases,” says Sabaratnam. “AI can help extract and aggregate large data sets, applying natural language processing (NLP) to automate case intake. Accelerating the capture of cases drives earlier analysis so companies can report issues and make preventative changes. These AI technologies also significantly reduce data entry associated with case intake, lowering overhead costs.”
“Given the short time frame in which adverse events must be collected, evaluated, and reported to regulatory agencies, pharmaceutical safety teams are forced to deal with increasingly large volumes of data (due to an increase in safety data sources, including social media, online forums, and physical documents) that must be processed quickly,” says Updesh Dosanjh, practice leader, Pharmacovigilance Technology Solutions, IQVIA. “Not only does AI-driven automation expedite the process by supporting or replacing manual activities; it often does a more thorough job than manual reviewers facing deadline pressure in far shorter timescales, minutes vs. hours … AI tools can analyze both structured and unstructured data in real-time. [Pharmacovigilance] query tools automate the reporting and management of case documentation, while NLP tools analyze complex narratives, including patient charts, social media posts, articles, and other unstructured data.”
The importance of AI in drug development is seen in the way it can run multiple analytical methods in real-time and evaluate data from different angles, according to Dosanjh, giving an organization a “richer picture of the data.”
AI serves a variety of purposes in safety evaluations during drug development, confirms Dr. Vaishnav, including clinical, administrative, and research areas. However, Dr. Vaishnav warns, use of AI comes with challenges. “AI-enabled products, for example, have sometimes resulted in inaccurate, even potentially harmful, recommendations for treatment. These errors can be caused by unanticipated sources of bias in the information used to build or train the AI, inappropriate weight given to certain data points analyzed by the tool.” Vaishnav continues that while AI cannot replace clinical trials, “the machine learning software can run faster and more accurately to analyze the data generated from these clinical trials, producing results that, again, are checked by experts to see if the software is properly evaluating the data to generate insights that help drug manufacturers make better-informed decisions.”
Mark Moorman, vice president of Artificial Intelligence, TCG Digital and Analytics Business Development, LabVantage Solutions, states that a lack of data is a concern in drug development. “A better AI/predictive model for drug safety and vigilance would be a quantitative model which uses expert guidance,” he says. “With qualitative models, you can guide the learning with an expert’s opinion on what could or should happen. This can provide more insight when real data is limited, and it would be a good way of carrying out pharmacovigilance in the early stages.”
AI can also help sponsors with drug development by using data to choose which products are worth developing further, according to Dr. Vaishnav. “From [a] pharmacovigilance perspective, if the drug does not meet efficacy criteria, or demonstrates unusual and unexpected side effects during drug development, using AI supported data analysis, drug manufacturers can focus on redirecting resources towards developing and delivering medications that could have a positive impact on patient life,” says Dr. Vaishnav.
Srinivasan agrees that AI can be used to improve decision making. “One can use AI for natural language understanding and image recognition to improve the quality of data received from drug studies,” she says. “This, combined with the current developments in big data analytics and cloud-based pharmacovigilance platforms will enable more sophisticated analysis of large datasets from real-world experiments.” AI can also reduce human error, speed up risk assessments, and identify patterns and trends in data, adds Srinivasan.
AI can help pharmaceutical companies study, learn, and predict needed changes for products already on the market based on safety information attained post approval to protect patients. “The findings can provide signals around the effects of long-term use of medicines that are unknown and help drive changes in either dosage or patient instruction,” says Sabaratnam.
When it comes to post-marketing safety data, Vaishnav states that AI and machine learning assist drug sponsors to collate data and develop solutions to practical adverse events. “Computer-assisted coding (CAC) of International Council for Harmonisation Medical Dictionary for Regulatory Activities (MedDRA) code adverse events mentioned in product labels [can be used]; however, they must be collected and stored in a way that streamline processes and introduce efficiencies into manual review processes and enable the tools to convert them to real-world data (RWD). This analysis of RWD produces real-world evidence (RWE), which is clinical evidence that can provide information about the usage and potential benefits or risks of a drug product,” says Vaishnav.
AI and computational linguistics methodologies are also used in the reporting of adverse events to regulators by using NLP techniques, according to Vaishnav. But there are limitations to the use of AI, Vaishnav warns, such as “difficulties with language learning, arising from the need to understand context and interpret ambiguities, particularly affecting translation, and inadequacies of databases. These limitations require careful curation of adverse event data for accurate submissions.”
“Qualitative models use expert opinions (often reflected in rules or significant factors) to aggregate events as meaningful or not. For post-approval changes, causal models may be a better fit because they help identify the root cause of events and whether it leads to meaningful events, such as side effects. A causal analysis run by a machine/AI can look at all the events that happened post-approval and potentially identify problems that need to be reported to the FDA,” says Moorman.
AI provides higher quality data for submission to regulatory agencies, according to Dosanjh, by asking smarter questions. “This allows cleaner data to be easily sent in an ingestible format to the internal teams allowing them to focus on analysis rather than data collection and extraction. AI can detect potential signals earlier in the reporting process, again giving analysis teams more time to make the right determination,” says Dosanjh. “The biggest risk organizations run today is the very real reality that a human processer has missed an event and the risk profile of my product is incomplete or consistent. AI offers a consistent analysis of data that significantly reduces these risks.”
AI applications used in pharmacovigilance include optical character recognition (OCR), which converts handwritten and typed text into machine-readable text; RPA; autonomous software; desktop automation; NLP; speech to text conversion; and Natural Language Understanding (NLU). According to Srinivasan, these are used to collect data on adverse drug reactions (ADRs) and improve accuracy, speed, and scalability, as well as reduce costs. Some of the neural networks and deep learning models used to create real-world data from ADRs, according to Srinivasan, include FastText, long short-term memory recurrent neural network (LSTM), and convolutional neural network (CNN). “By using different combinations and integrations ofthese available technologies, there is apotential to simplify and standardize the intake of ICSR [Individual Case Safety Report] data into [a pharmacovigilance] system,” says Srinivasan.
Sabaratnam points to the automation of pharmacovigilance tasks such as rule-based robotic process automation (RPA) or lookups, cognitive machine learning and chatbots, and orchestration of workflows or blockchain. “Both regulators and life sciences are on a learning curve to determine appropriate use cases, GxP validation, and quality assurance in a highly regulated environment,” Sabaratnam says. “Most of the AI technologies available now aren’t providing scientific judgement. They allow pharmacovigilance teams to filter information to identify trends and deliver signals. For example, AI is helping identify the right patient populations for treatments that perform within that demographic. In areas like oncology, this is a critical factor.”
While the application of AI in drug safety operations is evolving, according to Vaishnav, using AI for the analysis of big data can assist sponsors to make drug-event associations and help anticipate benefits or adverse reactions when performing risk-benefit assessments. “These signals provide valuable real-world intelligence which cannot be determined by mining data from controlled clinical trials. Similarly, [NLG] technologies can be applied to generate aggregate reports, or its basic framework, which can help with human experts freed up to provide further analysis and finalization.”
Natural language processing. NLP applications are used in pharmacovigilance to understand and classify post-marketing adverse events information from a variety of sources such as patients, healthcare providers, and clinical trials, explains Vaishnav. “NLP systems can analyze unstructured clinical notes on patients, giving incredible insight into understanding quality, improving methods, and better results for patients,” says Vaishnav.
Moorman stresses the benefits of NLP because of the textual or unstructured data involved in drug safety. “This area of AI continues to develop, but it is showing real-world usage and is ready to be applied to medical documents around adverse events. Nearest Neighbor models could also be used to categorize adverse effects and associate with either treatment, demographics, disease anomalies, personal information, etc. Also, the causal and qualitative models bring a lot of value—they are getting better and better every day.”
Additional tools. There are new modeling techniques for visual/image classification being developed as well, Moorman points out. “So, all in all, there’s been some incredible work on purifying natural language and image classification modeling that could prove very useful for drug development and safety.”
Post-marketing surveillance (PMS) or signal detection is used in the ongoing monitoring of adverse events. “PMS depends on analysis of data from sources such as medical assessment of spontaneous AEs, medical literature, regulatory agency databases, and clinical trial data, along with ‘real-world evidence’ such as electronic health records, medical device data, and data from social media and other online sources,” says Vaishnav.
The future of AI in pharmacovigilance may be found in new AI tools such as data hubs and medical evaluation support, confirms Dosanjh. Data hubs allow for the collation and sharing of information in real time across an organization, allowing for the combination of data across teams and a full view of the data. AI-driven medical evaluation support can also provide a more holistic view of the data and connect similar and relevant cases based on structured and unstructured data, according to Dosanjh. “An evaluator can compare their decision against an AI tool that has scanned the entire history of the product or similar events reported in all similar classes of products. This will bring a new dimension to evaluation and should support doctors in making better informed decisions,” says Dosanjh.
“As AI tools are used and have more data for the entire life cycle of an adverse event, we can extend data analytics capabilities and allow AI tools to not just identify potential issues, but also the mitigations that proved most successful, then suggesting those with smaller datasets in a far quicker manner than organizations can today,” says Dosanjh. “This will increase mitigations’ effectiveness and help organizations enact them quicker, reducing patients’ potential exposure to future [adverse events] or even preventing them.”
Susan Haigney is Managing Editor of Pharmaceutical Technology.
Vol. 47, No. 3
When referring to this article, please cite it as Haigney, S. Harnessing Technology to Ensure Drug Safety. Pharmaceutical Technology 2023 47 (3).