Artificial Intelligence has been around for a while, but a relatively recent openness exhibited by the drug and device regulators has opened doors...

Dec 13, 2018:Artificial Intelligence has been around for a while, but a relatively recent openness exhibited by the drug and device regulators has opened doors for exploring the cognitive concepts and robotics across the pharma value chain. The FDA had recently embraced devices for detecting diabetes and predicting chances of a stroke both based on application of cognitive technology. WEB-RADR, an EU regulatory agency sponsored initiative, is coming out with recommendations on the use of Social Media to monitor drug related adverse events. This agency is also promoting Adverse Event (AE) reporting and information sharing using a new platform. To this effect, the FDA has indicated openness to using Artificial Intelligence (AI) driven tools and technologies to support adoption of new platforms.

Patient safety can significantly benefit from direct and indirect applications of AI. Both clinical and post-market pharmacovigilance (PV) present a unique set of challenges that traditional tools and technology struggle to meet. The increasing case volumes from traditional and new (real world sources), such as social media, has encouraged pharma companies to look for innovative solutions such as robotics and automation techniques to process large data sets - accurately and cost effectively. In this article, we discuss some of the practical approaches to implement Robotic and intelligent process automation, including AI and machine learning in PV.

Two thirds1 of the data from new sources (social media, electronic health records, claims etc.) used in PV is unstructured. Since ICSR processing contributes to 50%2 of PV spend, companies are exploring the use of automation technologies to reduce costs. Over the last 10 years, within Big Pharma, the core enterprise Safety Database (SDB) has been monopolized with just two major vendors (Oracle and ARIS Global). Companies are grappling with lack of real-time data ingestion and insights from these closed systems and also internally built systems to address ever-changing requirements from patients, physicians, providers and regulators. They are exploring new ways to maximize value and return on their investments.

Transactional aspects of PV - case receipt, case intake, data entry, literature surveillance and quality control (QC) provide opportunities for smarter, newer technologies to drive additional benefits. Aspiration is to explore the use of AI technology to gain operational efficiencies in AE processing, but also for scientific activities such as medical review, causality assessments, signal and risk management.

Current approach to PV system implementation

PV safety database implementations is typically driven by adoption of new regulatory requirement(s) or mandatory upgrades to align with supported versions. Though these efforts have largely been standardized, they continue to be costly affairs with projects running for a year, or more, and costing upwards of $3M3. Another reality of PV safety database is frequency of system upgrades, which are in part driven by changing regulatory requirements and in some cases M&A activities. Based on pvnetworks® benchmark data, pharma companies undertake a SDB upgrade or a major SDB implementation effort every 3 to 4 years. Frequent upgrades allow for the main vendors to maintain significant control over the industry and lack of competition makes it difficult for pharma companies to assert control over customizations. Moreover, transactional tasks like AE processing is a heavily outsourced activity with nearly 40% being outsourced. Outsourcing costs have been commoditized and there is no real room for driving cost down.

Given the above landscape, pharma companies are looking to innovate. They are taking a closer look at their current IT landscape, smarter tools and technology to drive the next generation of efficiency gains.

Near Term Automation of PV (2-year Horizon)

In the last two years we have seen significant push from large and midsized pharma to make strategic investments in new technology and partnerships4. Implementation objectives may differ based on data complexity, product portfolio, volumes, and overall vision of the PV function, but overall approach seems to be the same.

These initiatives start with defining a roadmap based on vision of a PV function, followed by a defined objective of identification and prioritization of use cases which are, subsequently, prioritized for pilots. The Pilots are short, 6-12 weeks engagements, and the success is measured based on pre-determined criteria defined for each use case. This assessment helps develop an understanding of benefits and challenges of automation and reaching an understanding of a stepwise approach for PV processes to automate using technology.

Figure 1 PV Automation landscape on 5-year horizon

The key technologies that are on trial currently are pattern recognition, process automation, and rule-based engines:

Pattern Recognition

Natural language processing (NLP), Image Processing, Optical Character Recognition (OCR) and simple rule-based prioritization techniques help process the unstructured ICSR data set to convert into machine readable format. This has several use cases across the AE process from data intake, translation, triage and prioritization, to removing duplicates. Apart from reduction in manual touchpoints, further value is derived by improving turnaround times by aiding real time follow ups and feedback loops.

Identifying AEs in real world unstructured data like literature, social media, field reports, claims data, and healthcare data holds the original and long-term promise for future of patient safety and, investments made today, help in creating a robust platform

Robotic Process Automation

Robotic Process Automation (RPA) is, arguably, the most established in terms of its capability and value proposition. Repetitive tasks performed by the bots can significantly help in reducing the dependency on humans with direct efficiency gains but also improvement in safety, compliance and overall quality (“First Time Right”). Some successful use cases include acknowledgement emails, downloading and uploading of source documents, searching and mapping WHO products, data entry etc.

Rule Based Automation

Rule-based automation is nothing new. For a long time, before the COTS SDB became a norm, rule-based tools were frequently used for satisfying specific elements of the business process during the era of bespoke SDB. The rule-based engines still provide a credible, but simple, approach to help data prioritization and filtering based on scenarios which can be plugged into the existing SDB workflows for helping with triage and AE prioritization. These tools could be high maintenance as they are bolted on top of existing applications and needs to be managed else they can become a compliance risk.

An example is smart auto-narrative utilizing NLP and rule-based engines can bring significant improvement over existing template-based narratives which are standard with COTS products but are difficult to use and maintain.

Long term view of Automation in PV (3 to 5-year Horizon)

Zero-touch AE processing may not become a reality for all case types any time soon but as per industry projections, by 2021, proactive PV will start seeing significant investments to ensure that medical assessments and case processing can benefit from aggregate data from the downstream processes, like signals detection and trending. The result would help in supporting medical assessment of cases in real time. Case categories like Literature cases, Non-serious cases could be processed by no, or minimum, touch. Major platform vendors are incorporating some of the basic automation approaches in their upgrades that will improve operational efficiency Vs current manual data entry and processing activities.

PV analytics will likely see quantum improvements, with the tools providing significant depth and context to the data being analyzed, with ability to look across structured and unstructured data sets. Smart searches, easy to use interfaces with heightened data aggregation ability.

Natural Language Generation (NLG) will also provide capabilities to generate aggregate report, narratives and outputs for signal summaries and risk analysis and tasks currently found to be most effort intensive.

In next 5 years, learning systems with capability to perform determination of causality relationship between drugs and events based on existing data and trends will begin to realize. Though the technology is already there, it is a long way from being operationalized for PV. Significant challenge exists as the tools need substantial training data to deliver the proof.

It is also expected that the maintenance and validation of these smart technologies will become standardized with regulators who are independently exploring some of these smart technologies and looking at simplifying the current paradigm of system validation

Some of the leaders in the space are evaluating AI based approaches and tools that are used in the clinical space for application in the PV space that will allow the rapid integration of data and information – both structured and unstructured – for near real time analytics of signals and drive “proactive safety”. This will also help some of the RWE and HEOR approaches for analytics and application of these insights into the design of clinical trial protocols for the future.

Mostly importantly - the pilots, partnerships, smart technologies are being developed at a faster rate by entrepreneurial companies. Agile software development methodologies are being employed by nimble start-ups to address the challenges traditional established vendors are unable to address. While the PV software market is owned by the top two vendors, newer smaller vendors are disrupting the market - on the back-end and front-end of the PV life cycle.

Approach to implement Identify, Define, Design and Operationalize AI and robotics in PV

Working with our clients we have learned, that these initiatives are based on their product portfolio, case volumes, and technology infrastructure. It is important to develop a long-term strategic vision, but also equally important to start small! Quick wins go a long way in providing proof of concept and rationalizing big investments in technology. The steps to initiate a PV automation initiative may look like this:

Figure 2 Approach to Identify, Define, Design and Operationalize AI and robotics in PV


Compared to other industries and even some of the other functions such as in clinical, PV is at very early stages of the AI journey. Given the highly regulated environment the PV stakeholders operate in, risk aversion is normal. But, technology has matured enough and offers tools that can drive operational benefits and strategic value for PV. The technology space is buzzing with point solutions based on matured technology proven in other industries. Early adopters can benefit from experimenting with pilots and industry partnerships.

Stakeholders – Pharma companies, PV technology vendors, implementation partners, regulators are investing in new and collaborative ways of working. This is the time to dust off the vision for your PV and IT department and develop an early adoption strategy and a practical implementation approach with tangible milestones and outcomes for your organization.

1 The roughly 2/3 is from the fact that the 66% of cases are spontaneous is from pvnetworks benchmark data
250% spend is from pvnetworks benchmark data
3For a major version upgrade at a large pharma
4Source: pvnetworks®


Mayank Raizada
Managing Consulting, Navitas Life Sciences


Click here to read the original post


Latest News & Articles


Digital Solutions in Drug Development

It is vital that we prepare our workforces with the necessary skillsets to...


R Programming in clinical trial data analysis

It is vital that we prepare our workforces with the necessary skillsets to...