Testing cycles can be accelerated with automation
Automation in PV testing cycles brings forth a significant opportunity to formulate faster hypothesis. A 2018 pilot study conducted by the American Society for Clinical Pharmacology and Therapeutics (ASCPT)2 takes cognizance of the same. The said study tested the feasibility of using Robotic Process Automation and Artificial Intelligence (AI), among other new-age technologies, to automate the processing of AE reports. Solutions proposed by 3 commercial vendors participating in this pilot study were simultaneously tallied against Pfizer. The outcomes confirmed that AI-enabled automotive tools expedite the deduction of AEs from source documents.
Further advancement of technology in this domain, including leveraging Natural Language Processing (NLP) technology, Big Data Analytics, and data extracted from social media, could pave the way for complete migration of manually-intensive methods of source document annotation, which could effectively be replaced by total reliance on data fields of safety databases. While the benefits of such an overhaul can only be studied in the long term, it has been accepted universally by PV professionals that the ability to differentiate between vendor capabilities and identify the best candidate in a testing cycle for the discovery phase analysis is an assured benefit of process automation.
AI methods such as NLP, Machine Learning (ML), and Neural Networks have proven to improve the current workflow and provide support in the medical judgment and decision-making process.
Following are some of the categories where technology can be applied in ICSR processing:
Ingestion of structured and unstructured content Early introduction of technology to convert case information into a structured and digital format can eliminate the cumbersome task of case data relevance and data capture. Before data entry, cases need to be verified for validity and routed appropriately. A technology-driven approach to convert a case to digital includes reading incoming case intake information in the form of text, images, and text embedded in forms/tables from XML, Docx, and PDF files. Optical Character Recognition (OCR) can be used to extract relevant information from the files. Medical ontology-enabled NLP/ML maps relevant ICSR information to International Conference on Harmonisation (ICH) E2B data fields in a regulatory compliant manner, therefore eliminating a large portion of data capture efforts.
Case management and processing With 40% to 80% of allocated PV budgets spent on case processing and case volumes growing at a rate of 10% to 15% per year, driving costs out of case processing is the primary goal of 90% of the ASCPT survey respondents.2 Low-cost leaders are outsourcing, taking advantage of scale, and moving aggressively to automate case processing. Gaining cost control over this process while maintaining compliance and enhancing patient safety is entirely dependent on companies' ability to automate most of these activities.
Figure 2: PV automation focus areas (Data source: Deloitte survey)
Reducing the cost burden through simplification
90% Focused on reducing case processing cost through automation
~35% Expected cost reduction due to automation
>80% Utilization of managed services for case processing
Investing in signaling
>70% Identified maturity gap in signal processing
>90% Focused on expanding capabilities in product benefit and risk management
100% Investing in advanced visualization technologies
The highest impact opportunities in case processing come from effective triaging, duplicate checking, data entry, medical assessment, medical coding, narrative creation, and in-line quality control (QC) activities. Case content can be ingested and structured using OCR and NLP, following which business rules and AI classifiers can help identify correct case types for triaging to corresponding functions. More complex algorithms can additionally determine the case type, ensure data handling, identify duplication, route cases for additional information required, and also create follow-up mechanisms till resolution.
The inclusion of AI to automatically identify case type from a highly structured and known case profile can reduce manual review efforts and streamline the medical review process. AE case processing can be made more efficient for common low-risk cases by creating profiles of highly common non-serious and on-label case types. Approaches to automate narrative creation, ability to determine the listing of an AE against labels, medical coding, and detection of discrepant data through complex algorithms and business rules can reduce medical review efforts effectively. Lastly, ML can be applied to determine the machine generated confidence in classification to ensure intervention when necessary. Periodically using this confidence data as training sets helps the AI algorithms to achieve 100% confidence ultimately eliminating the overall requirement for QC.
Signal detection Most pharmaceutical companies continue to use traditional signal detection and investigation methods (e.g., medical assessment of individual spontaneous reports of AEs, interventional clinical trials, and database mining), a few are leveraging real-world evidence, and almost none are progressing using social media channels. This is consistent with current PV system capabilities. As pharma companies continue to drive toward true safety management, short-term signalling investments are likely to focus on visualization and long-term efforts on data integration tools and process investments. Using safety information to feed knowledge back into the discovery process in real-time remains a gap due to limitations with existing signal detection and management systems. The additional opportunity to make use of real-time signal intelligence integrate into the case processing information that can prompt detailed review on specific case types also remains an opportunity where AI can play an effective role. The ultimate goal is predictive signalling that applies knowledge of existing signals to strengthen the ICSR process to flag case types that need further investigations for signal management.
Decision-making Sometimes, the quality of information available in ICSR is poor. In such scenarios, semi-supervised or unsupervised learning plays a major role in devising hypotheses. For example, while building unlisted events and drugs correlation, and causality classifiers, specific types of Neural Networks are built and improvised with training over a period of time to enhance the medical assessment of specific patient data parameters, lab analytes, AEs, and corresponding medications. In some cases, technology can pull insights from precedent information that allows decisions on the case based on relevant reference information. This AI-based predictive capability provides reference precedent data for the relevant ICSR presented, enabling teams to evaluate the case and determine if it classifies as a known corresponding signal or as a potentially new signal.