The Value and Evolution of Knowledge Bases for Automating Pharmacovigilance.
Pharmacovigilance (PV) practices have proven incredibly useful for pharmaceutical companies who need to know which drugs may cause adverse drug reactions (ADRs) throughout a product’s lifecycle – and in what context.
But traditional PV activities have also received recent criticism for relying too much on information from spontaneous reporting systems (SRSs), which are sometimes incomplete, incorrect, and even biased. Critics say traditional SRS-based PV often ignores growing pools of additional relevant information such as product labels, peer-reviewed medical literature, and data from social media and other online sources.
However, harnessing all that information has been an enormous challenge for the research community, primarily due to the sheer (and growing) scale of medical literature, historically poor standardization, and a lack of interoperability and accessibility.
For these reasons, automated PV powered by machine learning (ML) for postmarket surveillance has become more common. It’s simply too costly and time-consuming to manually wade through all that data – especially when some regulators demand that companies do it weekly.
That’s largely why automated PV approaches have developed more or less in tandem with the growth of standardized ADR knowledge bases. These knowledge bases have helped facilitate the next generation of PV automation, while allowing individual researchers to more easily connect with global datasets, biomedical knowledge, and algorithms.
What is a drug knowledge base, and how does it help with PV automation?
Very generally, a knowledge base is where a computer system stores complex information (either structured or unstructured). Knowledge bases are primarily used in conjunction with artificial intelligence (AI)-based expert systems designed to solve particularly challenging problems but which typically need a lot of data to make decisions.
A knowledge base for ADRs, then, is essentially a computable database of standardized information derived from a broad range of biomedical information sources. It provides a machine-readable repository of information that can more easily facilitate AI- and machine learning (ML)-driven automation of PV activities.
Methods of standardizing ADR data include the World Health Organization Adverse Reaction Terminology (WHO-ART). That system is now no longer actively maintained, however, having largely been replaced by MedDRA, a standardized medical terminology “to facilitate sharing of regulatory information internationally for medical products used by humans.”
Such knowledge bases and standardization are essential for the large-scale automation of PV activities using ML and AI models, which have historically faced major challenges around scalability, data interoperability, data management, reproducibility, accessibility, and security.
Examples of drug knowledge bases
Many drug knowledge bases now exist and range from focused and relatively simple projects to larger and more comprehensive databases. “These curated drug knowledge bases have made significant contributions to the development of efficient and effective health information technologies for better health-care service delivery,” explain Zhu Et al.
The MediDrug Knowledge Base, for example, consists of four specialized databases: SFINX (a drug-drug interaction database), Pharao (adverse drug effects), Renbase (drug safety for patients with renal failure), and Gravbase/Lactbase (drug safety during pregnancy and lactation). Clinical recommendations are classified in MediDrug using the GRADE system.
Such a knowledge base allows users to more easily assess a drug’s risk profile based on previous ADRs and other risk signals.
Other popular drug knowledge bases include the following:
| Knowledge Base | Owner/administrator |
|---|---|
| PharmGKB | Stanford University |
| TTD | National University of Singapore |
| DailyMed | NLM |
| KEGG DRUG | Kyoto University |
| DrugBank | University of Alberta |
| SuperTarget | Charité – Universitätsmedizin Berlin |
| SIDER | European Molecular Biology Laboratory |
| DGIdb | Washington University in St. Louis |
| DrOn | University of Arkansas for Medical Sciences |
| DINTO | Universidad Carlos III de Madrid |
| Merged-PDDI | University of Pittsburgh |
| DID | Merck Research Laboratories |
The evolution of drug knowledge bases for PV automation
There hasn’t always been such an array of rich, openly-available knowledge bases available to pharmaceutical researchers. Thankfully, several studies over the past decade helped prove the viability of drug knowledge bases as a knowledge-sharing tool along with their usefulness for automating PV activities and ADR detection.
- Neubert Et al. (2013) built an ADR knowledge base (ADR-KB) using WHO-ART standardization for ADRs, ATC classification standardization for drugs, and LOINC standardization for lab test results, evaluating it within study populations of nearly 1,000 pediatric and internal medicine inpatients. The system achieved a specificity of 73 percent (up from seven) for internal patients and 91 percent (up from 19.6 percent) in pediatric patients.
“This study shows that contextual linkage of patients’ medication data with laboratory test results is a useful and reasonable instrument for computer-assisted ADR detection and a valuable step towards a systematic drug safety process,” wrote the study’s authors. “The system enables automated detection of ADRs during clinical practice with a quality close to intensive chart review.”
- Lopes Et al. (2013) created a distributed PV platform, including a cloud-based knowledge base derived from integrated and imported datasets, to help solve scalability, reproducibility, security, and interoperability issues. The system, known as the EU-ADR Web Platform, is still used today.
- Voss Et al. (2017) created the LAERTES (Large-scale Adverse Effects Related to Treatment Evidence Standardization) open-source and standardized knowledge base, which integrates several different evidence sources, including regulatory and clinical use cases combined with a predictive model. The predictive model demonstrated high predictive accuracy for models using all available evidence.
Automate your pharmacovigilance with CapeStart
Staying on the right side of industry regulators hungry for more and more compliance information isn’t easy. Companies need more efficient and cost-effective ways of conducting ongoing PV activities based on the latest medical knowledge and literature. CapeStart’s machine learning experts, natural language processing engineers, pharma subject matter experts, and pharmacovigilance specialists help you navigate and choose the right drug knowledge databases for your PV automation and other postmarket surveillance activities.
Contact us today to schedule a brief consultation with one of our experts.