Artificial Intelligence Based on Machine Learning in Pharmacovigilance: A Scoping Review

Benjamin Kompa^1,2, Joe B Hakim³, Anil Palepu³

¹Department of Biomedical Informatics, Harvard Medical School, Boston, MA, USA.

|May 17, 2022

Summary

This review examines how machine learning is being used to monitor drug safety. While many studies still rely on older statistical methods, newer approaches like deep learning are becoming more common. The authors highlight a need for better adherence to best practices in the field.

Area of Science:

Pharmacovigilance research within clinical pharmacology
Artificial intelligence applications in health informatics

Background:

The integration of advanced computational tools into drug safety monitoring remains poorly defined. Prior research has shown that automated systems transform various medical sectors. That uncertainty drove a need to evaluate current progress in this specific domain. No prior work had resolved the extent of modern algorithmic adoption. This gap motivated a comprehensive assessment of existing literature. Investigators sought to clarify how these digital innovations influence safety reporting. Previous efforts often lacked a systematic overview of diverse methodological approaches. Researchers required a clear synthesis to understand the current landscape of safety surveillance.

Purpose Of The Study:

The aim of this work is to evaluate the application of automated computational tools within drug safety monitoring. Researchers sought to understand how these technologies are currently utilized for surveillance tasks. The team wanted to characterize how these efforts differ from other scientific domains. They also intended to identify specific opportunities for improving safety reporting through advanced modeling. This study addresses the need for a systematic overview of existing literature. The authors focused on identifying which methodologies show the most promise for future implementation. They aimed to provide a clear picture of how modern algorithms influence safety workflows. This investigation serves to bridge the gap between technical potential and practical application in the field.

Main Methods:

The review approach involved searching four major databases for relevant publications. Investigators queried PubMed, Embase, Web of Science, and IEEE Xplore. The search covered articles published between the year 2000 and September 2021. Experts performed a manual screening of over seven thousand initial abstracts. A final set of three hundred ninety-three papers met the inclusion criteria. The team extracted specific details regarding study design and sample size. They also documented the technical methodologies employed in each selected work. Finally, the group evaluated whether these papers adhered to established quality standards.

Main Results:

Key findings from the literature reveal that fifty-three percent of papers rely on traditional statistical approaches for signal detection. Among studies utilizing modern algorithms, sixty-one percent employ off-the-shelf techniques with minimal changes. The analysis shows that only forty-two papers, representing ten percent of the total, follow current best practices. Regarding data intake, thirty out of one hundred fifty-four studies successfully incorporated these rigorous standards. Temporal trends indicate that deep learning methods are seeing increased usage in recent years. The data demonstrate that widespread adoption of advanced computational tools remains limited. Most research currently lacks the depth required for optimal safety monitoring. These results quantify the gap between potential innovation and actual implementation in the field.

Conclusions:

The authors suggest that advanced computational integration into safety monitoring remains in early stages. Their synthesis indicates that widespread adoption of rigorous technical standards is currently lacking. Researchers propose that future progress depends on aligning studies with established best practices. The review highlights that while traditional statistics dominate, deep learning is gaining traction. The team notes that data ingestion processes frequently miss opportunities for modern algorithmic optimization. They conclude that current trends show a slow but positive shift toward sophisticated modeling. The evidence implies that consistent methodological quality is necessary for meaningful improvements. The study provides a framework for evaluating future developments in this evolving field.

Keywords:

drug safety deep learning clinical informatics signal detection

Frequently Asked Questions

The researchers propose that while traditional statistical methods remain dominant, there is a growing trend toward using deep learning. This shift indicates a gradual transition toward more sophisticated computational approaches in drug safety monitoring.

The authors identify off-the-shelf techniques as a common tool, noting that sixty-one percent of studies utilizing modern methods rely on these pre-existing models with only minor adjustments.

The researchers indicate that adherence to rigorous best practices is necessary for valid results, noting that only ten percent of the total papers reviewed met these high standards.

The authors highlight that data intake and ingestion represent a critical area, where only nineteen percent of the one hundred fifty-four papers focusing on this task incorporated current best practices.

The team measured the adoption of modern algorithms by comparing the prevalence of traditional statistical methods against newer deep learning models across the identified literature.

The researchers propose that although advanced technology has not yet fully transformed the field, recent evidence suggests a potential shift toward greater integration of these tools.

Related Concept Videos