AI-ENABLED RAPID EVIDENCE SYNTHESIS FOR EMERGENCY OUTBREAK RESPONSE: A PROSPECTIVE TIME-AND-COST EVALUATION USING ANDES VIRUS AS A CASE STUDY
Author(s)
Hariharan Murugadoss, PharmD1, Madhumitha Haridoss, PhD2, Lavanya Ayyasamy, MSc1.
1SRM Institute of Science and Technology, Chennai, India, 2Assistant Professor, SRM Institute of Science and Technology, Chennai, India.
1SRM Institute of Science and Technology, Chennai, India, 2Assistant Professor, SRM Institute of Science and Technology, Chennai, India.
OBJECTIVES: Rapid evidence synthesis is critical during infectious disease emergencies, but conventional systematic review workflows are resource-intensive. This study compared the time, cost efficiency as well as verification burden of an AI-enabled, human-in-the-loop rapid evidence synthesis workflow with a reviewer-led conventional workflow, using Andes virus outbreak evidence synthesis as a case study.
METHODS: A time-and-micro-costing evaluation used a common PubMed, Embase, and Scopus search corpus. The endpoint was review preparation to a verified evidence table, including question framing, protocol drafting, search, deduplication, screening, full-text assessment, extraction, risk-of-bias/certainty appraisal, and evidence-table synthesis. Manuscript drafting and formatting were excluded. AI-enabled times were prospectively recorded and included human checking. Conventional times used average stage-specific values from available resource-assessment responses. Costs included human labor at INR600/hour and full AI/tool subscription costs. Outcomes included person-hours, workflow cost, productivity, eligible-study recall, extraction verification, correction rate, and unsupported outputs.
RESULTS: The search retrieved 5,732 records; 1,821 duplicates were removed, leaving 3,911 for screening. The conventional workflow assessed 137 full texts and included 19 studies; the AI-enabled workflow selected 25 full texts, initially identified 15 eligible studies, and included 17 after human-in-the-loop recovery. Person-time decreased from 86.05 to 19.68 hours, saving 66.37 hours (77.1%). Workflow cost decreased from INR51,630 (US$543.42) to INR21,650 (US$227.87), saving INR29,980 (US$315.55; 58.1%). Screening productivity increased from 187 to 2,607 records/hour. Initial AI recall was 78.9% (15/19), improving to 89.5% (17/19) after verification. Among 748 AI-extracted fields, 204 (27.3%) required verification and 6 (0.8%) correction, yielding 99.2% field-level accuracy. No unsupported/hallucinated outputs were identified.
CONCLUSIONS: AI-enabled rapid evidence synthesis reduced person-time and cost in this case study. However, residual missed studies after verification indicate that AI should support, not replace, reviewer-led oversight in emergency evidence generation.
METHODS: A time-and-micro-costing evaluation used a common PubMed, Embase, and Scopus search corpus. The endpoint was review preparation to a verified evidence table, including question framing, protocol drafting, search, deduplication, screening, full-text assessment, extraction, risk-of-bias/certainty appraisal, and evidence-table synthesis. Manuscript drafting and formatting were excluded. AI-enabled times were prospectively recorded and included human checking. Conventional times used average stage-specific values from available resource-assessment responses. Costs included human labor at INR600/hour and full AI/tool subscription costs. Outcomes included person-hours, workflow cost, productivity, eligible-study recall, extraction verification, correction rate, and unsupported outputs.
RESULTS: The search retrieved 5,732 records; 1,821 duplicates were removed, leaving 3,911 for screening. The conventional workflow assessed 137 full texts and included 19 studies; the AI-enabled workflow selected 25 full texts, initially identified 15 eligible studies, and included 17 after human-in-the-loop recovery. Person-time decreased from 86.05 to 19.68 hours, saving 66.37 hours (77.1%). Workflow cost decreased from INR51,630 (US$543.42) to INR21,650 (US$227.87), saving INR29,980 (US$315.55; 58.1%). Screening productivity increased from 187 to 2,607 records/hour. Initial AI recall was 78.9% (15/19), improving to 89.5% (17/19) after verification. Among 748 AI-extracted fields, 204 (27.3%) required verification and 6 (0.8%) correction, yielding 99.2% field-level accuracy. No unsupported/hallucinated outputs were identified.
CONCLUSIONS: AI-enabled rapid evidence synthesis reduced person-time and cost in this case study. However, residual missed studies after verification indicate that AI should support, not replace, reviewer-led oversight in emergency evidence generation.
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
SA2
Topic
Study Approaches
Topic Subcategory
Literature Review & Synthesis
Disease
No Additional Disease & Conditions/Specialized Treatment Areas