Turning the Tables: AI Halves Table Extraction Time in HEOR Systematic Reviews

Objectives

This study aimed to quantify the time savings achieved by using the Laser AI workflow for extracting tables, compared with the conventional manual extraction process in Excel during Health Economics and Outcomes Research (HEOR) systematic reviews.The study also aimed to assess the frequency and types of errors occuring during extraction with both methods .

Methods

Nine tables representing diverse HEOR-related data types: (costs, resource use, treatment patterns, utilities, cost-effectiveness results, epidemiology, population characteristics and transition probabilities for clinical and safety outcomes [4-12]) we repurpose fully sampled from published studies. Three reviewers of differing experience extracted each table once with Laser AI and, in a separate session, once with Microsoft Excel. Table-reviewer assignments were alternated to avoid learning effects.Table-level extraction time was recorded. Per-cell extraction speed (seconds/value) was calculated post-hoc by dividing table time by value count. To determine whether AI-supported extraction performance varied according to domain knowledge and tool experience we conducted a subgroup analysis based on two independent factors [Fig 2]: HEOR domain expertise (Senior vsJunior), Tool familiarity (Expert vs Junior).

After the extraction stage, quality assurance (QA) of the extracted data was also performed. For Laser AI, a dedicated QAmodule was used, while in Excel the results extracted to the sheet were compared with the data in the original publication.

Conclusion

This is the first known project to evaluate how AI can support data extraction from tables in HEOR reviews.

An AI-assisted extraction halved the time required to collect tabular HEOR data while completeness and error rates were comparable to the manual approach, demonstrating clear operational benefits for systematic review teams.

Future work will audit accuracy and explore learning-curves. As we transition toward a fully automated extraction process, we anticipate further improvements in time savings and efficiency