2023 Seminar, Princeton: Connected Natural Language Processing

Across drug discovery and development, scientists, clinicians, product teams and more need access to the information buried in unstructured text. AI technologies can reduce manual effort and provide effective access to the critical information for your decision support.  

In this presentation, we will provide an update of some of the latest innovations in AI. We will talk about the ability to bring machine learning models (such as BERT and GPT) into our technology platform, to allow for the combination of multiple NLP technologies. We will demonstrate IQVIA Human Assisted Review Tool, which enables effective document curation to speed up review and generate annotations for learning NLP. We will also showcase some recent use cases from customers, such as Genentech and AstraZeneca. 

The pharmaceutical industry is one of the world’s most heavily regulated industries, and the traditional document-centric approach poses challenges in terms of efficiency, cost and value. The ever-increasing volume of safety records and relevant regulatory data mean companies need tools and solutions to assist with safety and regulatory processes. Many companies are looking to innovative new technologies to transform document-driven processes to data-driven ones. In this session, we will discuss the potential that innovative NLP can bring, for digital transformation, automation, and human-in-the-loop augmented intelligence, presenting use cases from the FDA and top pharma. 

There is a huge amount of valuable text and document-based information available to medical and commercial teams, but it is often a challenge to work with. From notes captured by field teams that capture the unmet needs of physicians, patient perspective online, to the latest scientific literature and congress abstracts, this data can inform content development, medical and commercial strategy, competitive insights and more. Join us to hear about three examples of applied AI and NLP to unlock these invaluable data sources.

Unstructured text data from clinical and medical domains can generate substantial insights for healthcare organizations looking to drive innovation and improve outcomes for patients. In many cases, anonymization is the most practical way to leverage such data for secondary uses including research, analytics, and AI/ML.

However, anonymization of clinical text can be a difficult space to navigate, with numerous potential pitfalls to be aware of: Ensuring compliance with legal and regulatory requirements; dealing with dirty or non-standard data; and properly applying safe, privacy-protective techniques. 

While the use of machine learning (ML) brings its own inherent complexities, it also provides powerful tools that can bring scalability and increased efficiency to the anonymization process and help to navigate (or bypass) commonly encountered pitfalls.

In this talk, we will present case studies in which we successfully leveraged AI/ML technologies to scale up our unstructured text anonymization workflow to handle large and complex clinical texts. We will discuss the special considerations we made to handle this data in an ML context and enumerate the benefits we realized in terms of scalability and effort savings.

The growing document volume and varied data sources in drug development is a constant challenge. Natural language processing (NLP) technologies are key to efficiently keep up with the latest information. Generate insights 10-100x faster by combining new innovative large language models (LLMs) with established methods for a range of drug development tasks, including target identification, target prioritization, clinical trial design, and more. 

As part of a continued shift to greater transparency in clinical trials, sponsors must comply with increasingly demanding document publication requirements, such as those under EMA’s Clinical Data Publication Policy (Policy 0070), Health Canada’s Public Release of Clinical Information (PRCI) and more recently, the EU Clinical Trials Regulation. These initiatives require the sharing of anonymized versions of a broad range of document types, many of which contain dense participant data.

While rules-based redaction of this information is acceptable, even this simple solution can be difficult to achieve especially over tens of thousands of pages. Furthermore, regulators have been vocal and explicit in their preference for a robust, utility-preserving method such as statistical anonymization. This approach requires that participant data be transformed in such a way as to achieve defensible privacy protection. 

In this talk, we will outline some of the common challenges and pitfalls in handling statistical anonymization of clinical documents for public disclosure and set the scene for how AI/ML technologies can be successfully leveraged to scale up a workflow for even the largest and most complex clinical texts.