Server-side Stats Agent, extraction QC, and review adjudication
Statistical analysis
Faster, more reliable meta-analysis execution — results appear in seconds with no cold-start delays.
Sensitivity analysis tool: leave-one-out, comparison of 8 alternative τ² estimators (REML, DL, HS, SJ, PM, HE, ML, EB), and Knapp-Hartung adjustment.
Publication bias tool: Egger’s regression, Begg’s rank correlation, Duval & Tweedie trim-and-fill, and funnel plot generation with automatic power caveats for k < 10.
Forest plot images are now saved and available directly in evidence previews.
Extraction quality control
Automated QC triggers inspect every extraction for five common issues: missing effect sizes, impossible values, implausible magnitudes, subgroup-vs-overall confusion, and HR-in-OR leakage.
When QC detects an issue the system re-runs extraction with targeted repair hints before surfacing the result for review.
Improved study design classification for more accurate arm identification.
Review and adjudication
Per-outcome Accept / Reject decisions replace batch approval — reviewers make granular choices about which extracted outcomes to keep.
Full QC history visible in the review panel: which triggers fired, retry status, and score changes.
Agent-assisted adjudication evaluates conflicting extraction candidates and recommends the most likely correct value with a confidence score.
Reviewers can provide rerun instructions to guide a new extraction attempt for specific outcomes.
Navigation and bug fixes
Fixed broken Analysis tab and sidebar links.
Fixed a login crash when entering an unrecognised email address.
Batch extraction pipeline, supplement support, and Search Agent
Batch extraction
Auto-Extract All Outcomes processes every eligible study in parallel instead of one at a time.
Confidence badges (green / amber / red) on each extracted value give instant reliability signals.
Quality dashboard scores each extraction on evidence adequacy, arm alignment, schema completeness, and provenance quality.
Review Queue collects low-confidence or conflicting extractions for human validation with approve / reject / override actions.
Conflict side-by-side view lets you compare and accept competing values with one click.
Companion documents and supplements
Supplementary materials (PDF appendices, Excel tables, Word documents) are automatically fetched from PMC, publisher sites, and ClinicalTrials.gov.
Non-PDF supplements (DOCX, XLSX) are automatically read and included in the extraction process.
The extraction pipeline now discovers companion publications (secondary analyses, follow-ups) and includes them for a complete trial narrative.
Supplement sub-rows appear in the Full Text tab with descriptive labels.
Full text improvements
Bulk Upload lets you drag-and-drop up to 50 PDFs at once with automatic study matching by PMID, author-year, or fuzzy title.
A fallback search automatically checks PMC, CrossRef, and publisher sites when the primary PDF source is unavailable.
Upload PDFs button added to the Full Text toolbar for quick manual supplement uploads.
Search Agent
AI-powered Search Agent builds and executes optimized PubMed and ClinicalTrials.gov queries from your PICO/PEO criteria.
Real-time precision feedback screens a sample of scouted results against your PICO criteria before committing, preventing low-quality queries.
Newly discovered trials are automatically screened through your PICO/PEO criteria before entering the pipeline.
Extraction math engine
Arm swap detection identifies when treatment and control arms may have been mislabelled.
Cross-outcome hazard ratio deduplication prevents the same HR from being counted for multiple endpoints.
Events derivation from percentages when only percentage-based reporting is available.