Drug safety risk analysis – solving the puzzle of unstructured data

Historically, adverse event (AE) reporting has occurred directly through patients and HCPs. Traditional AE reporting has relied on manual and fractured systems and protocols that lead to an incredibly time-consuming process. This reporting has relied on structured data capture networks in a pre-approval landscape or from HCPs in a post-approval landscape, creating a gap in information sharing and potentially hindering early detection of safety concerns.

Adding insult to injury, not only do AE detection and reporting methods create a challenge in structuring accurate results, but data reveals that experts still lack a strong understanding of the current AE landscape.

The true number of AEs that occur throughout the drug life cycle is quite elusive. In fact, experts agree that the life sciences industry experiences chronic underreporting of AEs, and the estimates are staggering. As of today, the EudraVigilance database contains over 34 million individual case safety reports (ICSRs) for suspected AEs across Europe, and the volume of ICSRs continues to grow annually. However, even with such a large volume of suspected safety events, experts suggest that potentially 90% of AEs go unreported.  With so many potential AE cases, it is important that pharmaceutical organisations maximise detection strategies and sources to properly identify AEs.

Increasing the detection and reporting of drug safety events through expanded and improved monitoring aims to enhance the understanding and performance of pharmaceutical products. By improving the means and outcomes of AE detection, life sciences professionals and organisations can better understand the current drug safety landscape.

Expanding the sourcing and framing of AE detection to include new, non-traditional channels of information carries the potential to improve insights regarding patient safety and drug performance. For example, leveraging and interpreting data sourced from PSPs, digitised medical records, social media platforms, online forums, discussion groups and surveys provides a deeper comprehension of patient experiences and allows for greater detection of AEs that might otherwise remain unreported.

While unstructured data from these non-traditional channels poses challenges in operationalisation and interpretation due to their diverse formats and languages, leveraging AI, natural language processing (NLP) and optical character recognition (OCR) has provided the PV landscape with the tools to identify, analyse and sort various forms of unstructured data. Employing these technologies not only drastically expands the pool of patient information but also decreases the amount of manual labour required to identify AEs.

The challenge in interpreting language across unconventional channels is translating the use of social media text, emojis, slang and colloquialisms into structured, organised data. Through NLP, organisations can leverage technology to construct a dictionary or word bank of specific terms or patterns in order to identify safety-specific events. Through automated technology and ontologies, conceptual models connecting medical terminologies, organisations can identify safety-specific language to pinpoint patterns or word proximity within sentences along with misspellings and colloquialisms.

The use of this technology was heavily utilised following the rollout of coronavirus vaccinations. As HCPs had little direct contact with patients to assess the safety and potential side effects of the vaccination, PV experts turned to social media. By leveraging AI algorithms and pattern recognition, companies and experts were able to identify signals of suspected myocarditis in conversations across social media platforms. This is just one example of how non-traditional channels can provide further insight into AEs and patient safety.

Ingesting data from multiple structured and unstructured sources, leveraging technologies to perform semantic searches, product sentiment and syndicated ontologies can drastically streamline detection and identification of potential AEs. The success of these technologies’ ability to identify safety events and reduce manual labour involved in the safety risk identification is evidenced below.

In recent studies, PV efforts involving technology and automation, such as AI and NLP, were able to successfully identify 78% of AE data from virtual agents, such as chatbots and customer relationship management (CRM) systems. One use case involving a chatbot database used this technology to process over 292,000 messages and identify 445 AEs. Although the number of safety events identified appears small, the ability to sort through such a large number of potential events and leverage a new stream of information is incredibly valuable.

The advantages of this technology extend to PSPs, which can present a challenge as data often exists in multiple formats, both structured and unstructured, and in multiple languages. The difference between analysing PSPs and other data sources is the type of automation technology used. Through AI, NLP and OCR, PV technology can effectively review tens of thousands of PSP records while simultaneously decreasing reliance on manual effort.

Due to the unstructured nature of data sourced through PSPs, these efforts require a combination of multiple automation techniques in order to yield significant results. However, in one case, the automation process provided 90% efficiency, leaving just 10% of records for human review, delivering reviews of 45,000 records per month.

In another use case involving a top pharmaceutical company, the use of AI and NLP enabled the review of more than 7.7 million social media posts across 300 data sources, 91 languages and 38 countries. As a result of this implementation, the organisation was able to identify over 100,000 events relevant to the safety reporting process. This technology allows PV technology to sort through the immense amount of information across social media to validate safety events.

By increasing access to information while reducing the volume of irrelevant data entering safety databases, organisations can increase the validation of safety events. The value proposition of these technologies is that they increase efficiency by freeing up professional resources and shortening cycle times. Employing these technologies across novel, developing channels and instructing them to identify specific language also reduces the amount of irrelevant information entering the PV workflow. It also expedites the safety review process by identifying records required for human intervention, improving early detection and compliance.

The benefits of these technologies improve the breadth of sourcing for AE detection while also improving efficiency by eliminating time-consuming manual review processes.  Additionally, they enhance patient engagement and safety by incorporating direct patient feedback into drug safety analysis. This will ultimately lead to standardised processes resulting in higher compliance and reduced operational costs. AI and NLP technologies can also assist in validating signals of potential AEs by cross-referencing patient support programme data with other sources, thereby improving confidence in the validity of identified signals, reducing the likelihood of false positives and ensuring that resources are focused on addressing genuine safety concerns.

These non-traditional, evolving data sources represent novel avenues for uncovering previously unnoticed patterns and trends in patient experiences and safety event reporting. As the patient experience evolves and means of communication continue to diversify and become more consumer-centric, the life sciences industry can expect automation and analytics through new technologies to become increasingly beneficial.  Taking a holistic view of medication safety and considering diverse data streams enhance the overall understanding of AEs and potential risk factors. Learning to organise and analyse unstructured data across various sources will allow organisations to improve safety event detection and drug performance in the future.

Deepanshu Saini is Director of Program Management at IQVIA