Revolutionising Drug Safety with AI
Introduction
Advances in the life sciences have brought about a transformative impact on healthcare, with life span and quality of life dramatically improving world-wide. Improvements in manufacturing now mean that the pharmaceutical sector has the capacity to produce vast quantities of effective drugs. However, this capacity brings about the task of distilling the vast amounts of data relating to pharmaceutical products into an accurate and understandable form, so that patients can use treatments effectively. Even common over the counter medicines such as paracetamol (a.k.a. acetaminophen) can cause serious health issues if the information relating to the indications, directions, dosage, side-effects and other advisory and safety information is not accurate.
At the other end of the pharmaceutical spectrum, more advanced treatments such as immunotherapy can be ineffective or counterproductive if the latest information relating to their optimal use is not up to date. Once a medicine is in the market it must undergo constant surveillance, through clinical trials and the monitoring of physician and patient reported adverse events. Any new information on the safety and efficacy of the drug must be rapidly propagated to the consumer. Failure to do so can result in compromises of patient safety, recall of products and large fines from regulators.
Pharmaceutical Drug Labelling
In the pharmaceutical sector the management of such information is carried out by a labelling process, where teams of medical legal and regulatory specialists track and coordinate the latest data on medicinal products, so that accurate material ultimately ends up on the consumer packaging and information leaflets. These leaflets and packaging are known in the industry as drug labels however, while they are a major source of information for both physicians and patients, there have been issues relating to the accuracy and timeliness of relevant information in drug labels. Information such as adverse drug-drug interactions (DDIs) with the medication is a source of critically important clinical data and regulators mandate that label information for approved drugs should include observed and predicted clinically significant DDIs[i]. Drug-drug interactions can lead to a range of preventable adverse events, which are reported as the eighth leading cause of death in the United States[ii]. Drug product labels can also lag behind emerging drug knowledge as it may be a number of years since the drug was first released on the market[iii].
There are also issues relating to multiple versions of drugs. In a review study of 9105 drug product labels it was found that significant numbers of multi-manufacturer drugs and generic drugs had discrepancies in the adverse drug reaction (ADR) sections of the labels, due to missing and outdated information and formatting issues[iv]. ADRs are among the leading causes of death and are estimated to cost approximately $136 billion annually[v].
A number of other studies have found that drug product labels failed to keep up to date with the latest finding from scientific research and clinical trials
● deficits in the pharmacokinetic data listed in product labels[vi]
● omissions in age-related product label information in antidepressants[vii]
● quantitative information on renal clearance changes in elderly patients was omitted in 92% of 50 products[viii]
● deficits in drug-drug interaction information 15% of the product labels for drugs that interact with warfarin[ix]
Software integration and structured document authoring plays a role in addressing these issues by maintaining the integrity of information across the labelling process, however there is still an enormous level of manual effort involved. Therefore, human and system failures can lead to inaccurate safety information[x].
AI to the Rescue
However, artificial intelligence (AI) has recognised potential for processing and managing drug related information[xi], with numerous case studies already demonstrated, including the detection of adverse events[xii] and for drug-drug interaction extraction[xiii]. At The Dock, Accenture’s global innovation center, we are also addressing these challenges by creating AI technology that can automatically review the highly specialised medical language used in drug product information. Our AI can assist humans in validating the veracity and consistency of the content of drug label information on cartons and patient information leaflets by using deep learning based natural language processing (NLP) that can extract clinical terms in drug labels and detect the relationships between them, essentially putting meaningful structure on unstructured text. We then use this structured output to build a knowledge graph of interacting substances using Ampligraph, a machine learning tool developed by Accenture’s Tech Labs that creates neural network-based graphs. These graphs aggregate related drug label knowledge into a data structure that serves as a basis for detecting missing drug information and recommending suitable drug interactions, see the visualisation of such a graph below:
This graph allows our software to reason over the meaning of the language in drug labels, for example identifying the drug of interest and ensuring that interacting drug mentions are correct and recommending additional drug classes. Such deep learning models also have the ability to be trained with new knowledge and improve over-time in terms of coverage and accuracy. Our AI pipeline can also validate the AI’s understanding of the semantics against business rules to ensure that the drug information corresponds to the guidelines for the medication in question. This approach can also be extended to other industries that must comply with regulations, such as finance, legal, transport and supply chain.
While we can quickly assemble an AI pipeline to help review pharmaceutical artwork, we need to be cognisant of the limitations and pitfalls of AI. Solely relying on AI to review such crucial text would be foolhardy. Rather we aim to leverage AI to assist humans in accelerating and enhancing such reviews, so that the cost of medication and time to market can be lowered. Human oversight should be in the loop and existing medical, legal and regulatory processes should be used in the final approval process. Moreover, human expertise can in turn be used to refine and enhance the statistical models that underpin the AI behind our system, ensuring that humans and AI work in a symbiotic relationship.
This is just one of the ways in which The Dock, where I work, is using AI and NLP across a wide variety of problems.
https://www.accenture.com/ie-en/company-dublin-innovation-centre-the-dock
References
[i] Boyce, Richard, Gregory Gardner, and Henk Harkema. “Using natural language processing to identify pharmacokinetic drug-drug interactions described in drug package inserts.” In Proceedings of the 2012 Workshop on Biomedical Natural Language Processing, pp. 206–213. Association for Computational Linguistics, 2012.
[ii] . Goldstein, J., Jaradeh, I., Jhawar, P., Stair, T. ED Drug-Drug Interactions: Frequency & Type, Potential & Actual, Triage & Discharge. The Internet Journal of Emergency and Intensive Care Medicine 8(2), (2004)
[iii] Leveraging the semantic web and natural language processing to enhance drug-mechanism knowledge in drug product labels (November 2010) https://dl.acm.org/doi/10.1145/1882992.1883070 [See slides at http://www.pitt.edu/~rdb20/data/presentation-of-the-PI-mashup-use-case-11032011.pdf] [Contact Boyce?]
[iv] Consistency in the safety labeling of bioequivalent medications (October 2012), https://onlinelibrary.wiley.com/doi/pdf/10.1002/pds.3351
[v] Preventable Adverse Drug Reactions: A Focus on Drug Interactions,
[vi] Marroum PJ, Gobburu J: The product label: how pharmacokinetics and pharmacodynamics reach the prescriber. Clin Pharmacokinetics 2002, 41(3):161–169. [http://www.ncbi.nlm.nih.gov/pubmed/11929317]. [PMID: 11929317]
[vii] Boyce RD, Handler SM, Karp JF, Hanlon JT: Age-related changes in antidepressant pharmacokinetics and potential drug-drug interactions: a comparison of evidence-based literature and package insert information. Am J Geriatric Pharmacother 2012, 10(2):139–150. [PMID: 22285509].
[viii] Steinmetz KL, Coley KC, Pollock BG: Assessment of geriatric information on the drug label for commonly prescribed drugs in older people. J AmGeriatrics Soc 2005, 53(5):891–894. [http://www.ncbi. nlm.nih.gov/pubmed/15877571]. [PMID: 15877571]
[ix] Hines L, Ceron-Cabrera D, Romero K, Anthony M, Woosley R, Armstrong E, Malone D: Evaluation of warfarin drug interaction listings in US product information for warfarin and interacting drugs. Clin Ther 2011, 33:36–45. [http://www.ncbi.nlm.nih.gov/pubmed/21397772]. [PMID: 21397772]
[x] https://www.fda.gov/safety/recalls-market-withdrawals-safety-alerts/novis-pr-llc-issues-voluntary-nationwide-recall-pecgen-dmx-due-labeling-error
[xi] Machine learning-based identification and rule-based normalization of adverse drug reactions in drug labels, (March 2019) https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-019-3195-5
[xii] Adverse Event extraction from Structured Product Labels using the Event-based Text-mining of Health Electronic Records (ETHER) system (2019), https://journals.sagepub.com/doi/pdf/10.1177/1460458217749883
[xiii] Mahajan, Diwakar, Ananya Poddar, and Yen-Ting Lin. “A Hybrid Model for Drug-Drug Interaction Extraction from Structured Product Labeling Documents.”, Proceedings of the 2019 TAC, https://tac.nist.gov/publications/2019/participant.papers/TAC2019.IBMResearch.proceedings.pdf
