In the past few years, the domain of Pharmacovigilance (PV) (incl. Cosmetovigilance and Materiovigilance) has moved from a point where automation was being explored to automation becoming an imperative.
The primary reasons for this change are simply the numbers. It is estimated that the number of Adverse Events (AEs) being reported has been increasing by 30 % on an average since 2019. Large pharma companies, especially the ones with vaccines (which is almost all of them), are having to process AEs upwards of 500k per year, costing the companies in tens of millions of dollars. Even Small and Medium size pharma and biotechs are having to spend significant amounts in buying and maintaining the software and the case processing costs.
Automation in PV has been introduced over the years through rule-based automation or in some cases, robotic process automation (RPA). However, these automations have hardly made a dent in the case processing costs where drug safety associates perform the data entry, quality check and medical review steps. With the rapid evolution of Artificial Intelligence (AI) driven automation in the last five years, companies are hoping for complete automation of data entry, quality check and to some extent medical review of non-serious safety cases. Touch-less case processing has become the holy grail in this business.
The challenges are many. Firstly, the number of sources for safety case collection have increased. Earlier, it used to be post cards, Emails, Web forms, Clinical Trial systems, Faxes and Call Centre transcripts. The safety case forms would flow into the Case Processing centres managed by in-house or CRO teams, and dozens of case entry personnel would enter the information in Safety Databases.
These sources still remain. In addition, the new sources such as Literature articles, and Real-World Data (Social Media and Electronic Health Records) are throwing up an explosion of potential safety cases from which one needs to identify valid safety cases for data entry and processing.
For example, even after using a refined search strategy, one would still find 100s of articles that talk about a particular drug and AE combination. The reviewer has to read through all these articles to find at best 4 or 5 articles that yield a valid safety case. With thousands of articles being published every week globally in dozens of languages, imagine the number of people and effort required, for a company with 100+ drugs in the market across geographies.
To perform Causality Assessment for an Individual Case Safety Report (ICSR), more often than not, for each valid case, between two to three follow-ups will be required with the reporter (could be the patient or the Health Care Worker). The challenge is exacerbated due to the stringent timelines for reporting a suspected serious adverse event and the regulatory expectation around signal detection and risk management so that an up-to-date benefit-risk profile of the drug in placed in the public domain for the use of agencies, providers, physicians and patients.
With many big pharma exploring additional uses for their drugs, the conduct of observational trials for such drugs has increased manifold. These trials depend a lot on medical literature, RWD and Patient Reported Outcomes to establish the benefit-risk profile of the drug for the specified additional use. Safety case data collection and medical review of the same needs to be efficient to help companies make decisions on R&D, Sales and Manufacturing.
The above background makes it imperative that the business process and technology of PV needs to be reimagined to meet the evolving challenges and opportunities.
Broadly speaking, automation would yield the best gains in the following areas of PV:
- ➣ Case Intake and Triage – Automated Case Intake and Case Creation from multiple sources and languages and triaging on the basis of a) Valid Case information and b) Seriousness
- ➣ Quality Review and Follow-ups – Automated quality checks based on field level validations, and auto generation of follow-up queries and resolution.
- ➣ Medical Coding – Automated medical coding, Expectedness/Listedness
- ➣ Medical Review – Causality Assessment and Narrative Generation
- ➣ Safety Case Submission – Automated generation of Safety Case Reports and other documents in prescribed formats based on submission rules of different agencies, and integration with submission gateways for submission in E2B R2/R3 Formats
The AI/ML techniques that can help automate the above process steps, could be listed as below:
| AI/ML Technique | Process Steps | Maturity Level |
| Computer Vision (OCR) | ➣ Case intake from scanned PDFs, Faxes, electronic PDFs, Emails and Office documents ➣ Extraction of Safety information from Labelling documents | Evolving – reasonably mature; Will require refining/retraining standard models with company-specific data |
| NLP – Named Entity Recognition, Relationship Extraction, Natural Language Generation and Speech to Text transcription | ➣ Case creation of OCR output based on mapping with E2B data elements ➣ Extraction of safety case information from Literature articles and case creation ➣ Medical Coding ➣ Narrative generation ➣ Case creation from Social Media feed ➣ Case creation from Call Centre voice transcripts | Reasonably mature in some use cases (Eg: Medical coding and Narrative generation) and evolving in other use cases |
| ML Models – Random Forest, Statistical Methods including Neural Network models | ➣ Seriousness Prediction ➣ Quantitative Signal Detection ➣ Causality Assessment | Mature in case of classification (random forest) and quantitative signal detection |
| Generative AI – Large Language Models with Retrieval Augmented Generation (RAG) | ➣ Extraction of Safety Case information from articles and other sources ➣Q&A on large safety case data sets for insights ➣ Narrative generation ➣ Generation of draft documents in prescribed formats such as CIOMS, MedWatch, Signal Management Report, PSURs, Safety data Exchange Agreements ➣ Safety Case analytics ➣ Machine translation | Early stages – Some good use cases established as POCs. Generative AI has the best potential, when mature, to transform PV completely. |
| Rule-based Automation – no AI involved. | ➣ Rule-based Automation – no AI involved. ➣ Case triage based on country and seriousness ➣ Quality checks based on field level validations ➣ Seriousness prediction ➣ Causality Assessment based on WHO or COLIPA scales/decision trees ➣ Medical Coding via browser ➣ Submission Rules configuration ➣Configurable Workflows ➣Template-based case creation and narrative generation ➣ Case report generation and Operational metrics ➣ De-duplication of safety cases or literature articles based on parameters | Mature level in multiple use cases |
| Semantic Search | Search large data sets of articles or documents to extract relevant articles based on configurable parameters | Mature |
An important aspect that is usually not discussed much when it comes to implementation of AI/ML models in PV automation is the quality of data. Companies expect that AI models will be ready-to-go as part of the system implementation. The truth is that any pre-trained model by the safety technology vendor will only be as good as the similarity between its training data set and the live data that comes in through the safety system. A certain amount of retraining and refinement will be required as part of the system implementation phase, and with proper planning, the ML algorithms should be able to provide acceptable levels of accuracy in the production environment (98%+). However, it is critical that the data quality be good for this retraining to work well. The data quality will also impact the ability of the RAG model to set the context for the LLM and thereby reduce the hallucination and/or bias.
Industry analysts expect the cost and effort savings from PV automation to be upwards of 45% when compared to the current implementations. That is very much a possibility and we already have a few good use cases that help achieve significant cost and time savings on specific process steps such as case intake and quality checks.
Please watch this space for more articles that will delve into the aspect of reimagining Pharmacovigilance using the available and evolving AI/ML technologies. If you want a specific use case to be discussed, please feel free to comment below or write to: narasimha.k@datafoundry.ai
The author is Narasimha Kumar – Chief Product Officer at Datafoundry. AI, where we have infused the AI technologies discussed in this article in the DF mSafety AI Platform, to provide and end-to-end automated and intuitive user experience for professionals in the PV space.