The Enduring Challenge of Medical Device Identification: Is Artificial Intelligence Abstraction Ready for Broad Use and Acceptance?
A key barrier to generating robust real-world evidence (RWE) for medical devices and diagnostics is the inability to reliably identify specific products within real-world data sources. As discussed during an ISPOR 2026 Annual Meeting workshop, unlike pharmaceuticals, which are typically captured in structured data fields, devices and diagnostics are often recorded inconsistently or embedded within unstructured clinical documentation. This undercapture limits accurate exposure assessment, cohort identification, and outcome measurement, ultimately constraining the credibility of RWE for regulatory, payer, and health technology assessment (HTA) decision-making.
Traditional approaches to device identification remain limited. Structured data fields are often incomplete, keyword searches may miss indirect references, and manual chart abstraction is costly and difficult to scale. Even conventional natural language processing methods struggle with the variability of clinical language. A particular challenge is that device use is frequently implied rather than explicitly documented. For example, references to monitoring activities or clinical improvement may suggest device use without identifying a specific product. These limitations highlight the need for approaches capable of interpreting clinical context rather than relying solely on direct mentions.
Generative artificial intelligence (GenAI) offers a promising solution by extracting clinically relevant information from narrative text. By identifying both explicit and implicit signals, GenAI can infer device use and generate structured outputs, including device type, supporting evidence, and usage status. Performance is highly dependent on task design, with structured prompts improving both accuracy and consistency.
When implemented with appropriate validation, governance, and human oversight, GenAI can enable scalable transformation of clinical notes into analysis-ready data, helping address a critical evidence gap and strengthening RWE generation for devices and diagnostics. Although these advances suggest that AI-based abstraction is approaching readiness for broader implementation, widespread acceptance will depend on continued demonstration of accuracy, transparency, and reproducibility in real-world applications. Building trust among regulators, payers, providers, and other stakeholders will remain essential to adoption at scale.
Reference
Weiss L, Williams A, Royer J, Caños D. The Enduring Challenge of Medical Device Identification: Is Artificial Intelligence Abstraction Ready for Broad Use and Acceptance? Presented at: ISPOR 2026 Annual Meeting; May 17-20, 2026; Philadelphia, PA.
