This special issue on machine learning (ML) in Drug Safety illustrates the extent to which the excitement around ML in broader society is now pervading pharmacovigilance (PV). Pharmacovigilance is in a time of great change [1] and there is much discussion of the role that newer technologies including ML can and will play in driving this necessary change. This excitement is not of course unique to PV and we see widespread scientific research and discussion of ML and artificial intelligence (AI) in the broader healthcare field. Machine learning used routinely in many applications, for example, voice recognition for automated clinical visit scribing, and visual pattern recognition for medical imaging such as in retinopathy [2]. Given the complexity of medicine and healthcare delivery, rule-based systems that typically involve human-curated rule sets are necessarily limited in capability both for recognizing large numbers of complex patterns and for automation in data ingestion, pre-processing, and dissemination. One would theoretically anticipate expert-crafted rule-based systems will be upper bounded by human capacity, where with sufficient rich well-labeled training data and generalizability the potential for ML seems much more promising.

While the use of ML is not new in safety, see for example [3] nor even in application to safety reports, for example [4], there has been limited routine use in PV and there are many reasons for this [5]. However, there are signs this is changing and some of these barriers are beginning to be overcome, particularly in the area of natural language processing, which is finding extensive use in the extraction of information from free-text clinical notes in electronic health records [6]. In this issue, there is a scoping review [7] showing the breadth of research from data ingestion to signal detection. There clearly remains much confusion and lack of clarity around the scope of ML and AI and the usage as discussed in a systematic review [8] shows there is clearly a huge increase in published research on AI-based ML [7]. This is further illustrated by the wide range of examples of original research covering applications as diverse as predicting drug approvals [9], automated patient-reported adverse event [10] and drug coding [11], and adverse event report causality assessment [12], and disease prediction [13] and its role in supporting decision making by safety experts during signal validation [14]. This issue also contains perspectives from different stakeholders and data networks [15,16,17], insights and challenges into how ML can help facilitate identifying the completely unexpected ‘black swan events’ [18] and insights into how ML is making inroads into causal inference and telehealth and in resource-limited settings [19,20,21].

Despite the range of articles, it would however be a mistake to believe that all the challenges for effective, trusted, routine production ML have been resolved—we are still some way from ubiquitous ML yet! Many thorny issues remain for the use of ML in PV. Consider a few examples. How important is contemporaneous explainability in the broadest sense of the term? Clearly, the ability to explain so another understands the reasoning behind an output boosts trust in the system, but is it essential? Does this depend on the application or even the choice of the algorithm? For example, should we prefer deterministic over non-deterministic algorithms? A requirement on contemporaneous explainability may limit performance especially if we require the ML to only do what a human can do or at least comprehend the value in the immediacy of a ML-based suggestion. The example of the board game ‘Go’ and move 37 during the second game of a series defeat of the human champion by an AI system springs to mind, which at the time of the move was not readily appreciated, for example, “that’s a very strange move” and “I thought it was a mistake” [22], and was since seen as brilliant through the lens of retrospect. Yet if one was always to require some sort of retrospective comprehensibility to a human as a condition for trust, how would one define the length that a retrospective evaluation would apply? If this was a necessary cause, this could clearly preclude the timely use of some ML outputs for decision making in certain circumstances. Would AlphaGo have lost that second game of Go had a human-required approval been needed for proposed move 37 and denied?

Similarly, performance of ML is a contested issue, in safety, we have a responsibility to show we are doing all that we can to ensure the safe use of medicines and continuously strive to improve. How do we show improvement with new technologies? Much has been said that strong performance of ML in a subtask is promising for future PV, but it must also improve performance across the overall PV lifecycle [18], rather than just creating work or inefficiencies or delays elsewhere that overwhelm improvement at the particular steps or tasks. A compelling evaluation of safety system performance on a holistic level is notoriously difficult, however, with much discussion about reference sets for method evaluation [23]: progress in this area will be needed to show the future value of ML.

Much data of relevance to PV cannot be shared readily because of the broader ethical and privacy concerns around sharing healthcare data. Arguably, sharing code has limited value unless it is simple to follow and therefore can easily be accurately implemented and adapted with minimal effort to run on a data source or run on public domain data. In other parts of healthcare, many articles rely on data sets available in the public domain [24]. Data sources that have been used to train ML models more generally are increasingly being released so that results can be fully reproduced by other researchers. For example, the use of ML on real-world data for PV might accelerate if we see more publications with open source code that runs on public domain real-world data sources such as for example, MIMIC-III [25] or the US Centers for Disease Control and Prevention NHANES data set, and this could similarly foster reproducibility as well as address confidence in performance.

So what to the future? We expect to see more efforts to integrate ML holistically across the entire PV lifecycle, and as the need for rapid and effective learning from emerging data for decision making has become even more evident during the COVID-19 pandemic, we anticipate also evolutions in workflows [26]. For example, performance of disproportionality analyses of individual case safety reports is more limited in identifying high-order dependences such as drug–drug interactions because of the increased impact on quantitative scores of coincidentally similar reports, which may be artefactual. Machine learning to cluster and quantify similarity of reports [27] has been used to downweigh likely duplicative reports to make such higher order dependency signal detection more effective [28], therefore linking ML in data ingestion to a subsequent data analysis. 

In addition to more clear links between ML for data ingestion and analysis, we expect to see more evidence of the implementation and practical impact of routine use of ML, not as sole ML solutions but embedded in overall production systems including together with rule-based systems. We hope this special issue provides the reader with a clear perspective of ML in evolution in PV, and we expect it to herald ever more interesting, informative, and important articles on the use of ML in PV. Please see video 1 (online only) for authors’ views on this theme issue of Drug Safety.