Visual Data Extraction in SLRs: Can You Trust the Data?

Data Extraction from Charts and Graphs: An Overview

While extracting structured data from published text has become increasingly mature, extracting equivalent precision from visual data remains a significant challenge. Extracting the same precision from figures is not. In systematic literature reviews (SLRs), critical numbers often live only inside forest plots, Kaplan-Meier curves, Sankey diagrams, or patient flow charts. They never appear as explicit text. Platforms that handle text extraction well still struggle, or fail, to infer those values.

This gap matters significantly across modern research. Systematic reviewers, health technology assessment (HTA) teams, and evidence synthesis groups routinely need point estimates, confidence intervals, survival probabilities at specific time points, flow proportions, and attrition counts that exist solely as visual elements. Without reliable, traceable extraction, those values either go unused or require slow, error-prone manual digitization. Specialized AI-powered literature review services help teams overcome these limitations by automating the capture of critical figures.

A recent expert validation study examined whether modern graph-inference systems can close that gap under real conditions. Subject-matter experts reviewed AI-inferred values drawn from published articles across oncology, dermatology, ophthalmology, and other therapeutic areas. The review assessed seven common chart types and extracted insights from more than 80 graph-inferred data points that text-based methods could not capture. The findings provide concrete evidence on accuracy, traceability requirements, and remaining limitations in automated evidence workflows.

Visual Data Extraction: Bridging the Gap in Evidence Synthesis

Integrating robust Visual Data Extraction techniques ensures that evidence synthesis teams do not miss crucial data locked within complex clinical graphics. By combining advanced vision models with precise geometric reasoning, researchers can extract quantitative estimates directly from figures with minimal error.

Challenges in Visual Data in Evidence Synthesis

Published clinical literature encodes large volumes of quantitative results in figures. Forest plots carry standardized mean differences, odds ratios, and heterogeneity statistics. Kaplan-Meier curves show survival or response probabilities at exact time points or the inverse time at which a probability threshold is crossed. Patient flow diagrams and Sankey diagrams record enrolment, discontinuation, and treatment sequencing. Bar, line, and pie charts present subgroup rates, measurement values, and proportional breakdowns. For organizations operating in Global medical affairs, rapidly converting these visual elements into structured data is essential for timely decision-making and regulatory reporting.

Evidence teams spend significant time manually extracting data from these figures, and the process often introduces variability. Tools such as Plot Digitizer and specialized algorithms for Kaplan-Meier digitization improve consistency over pure visual estimation. However, they still require significant human effort and introduce their own sources of error. Recent evaluations of multimodal large language models confirm that general-purpose systems reconstruct table structure reasonably well but frequently fail at precise numerical recovery when data labels are absent from the image. To explore how automated methods are changing this landscape, read our analysis on the future of systematic literature reviews.

In SLR workflows, this creates a practical bottleneck. Reviewers either omit figure-only data, spend hours calibrating axes and reading points, or accept low-confidence estimates. Neither approach scales when hundreds of studies must be processed under tight timelines for regulatory or market-access submissions. Effective visual data extraction mitigates these bottlenecks by automating point recovery with full provenance.

Industry Landscape and Technical Considerations 

Evidence synthesis teams in pharmaceutical, medical device, and health-technology assessment settings operate under strict reproducibility standards. Every extracted value must be auditable. Teams must trace every extracted number to the original figure, caption, and supporting sentence to ensure it meets quality standards. Deploying a comprehensive Life sciences research platform ensures that all visual data extractions remain fully traceable across regulatory filings.

General vision-language models face three recurring difficulties:

• Geometric reasoning under visual noise (overlapping curves, dense multi-arm plots, unlabeled scatter points).
• Accurate scale interpretation when axis ticks or legends are partially obscured or non-linear.
• Hallucination risk when the model fills in plausible but incorrect values instead of declaring uncertainty.

Specialized graph-inference systems address these challenges by combining computer vision for element detection, geometric mapping of visual coordinates to data values, and explicit linkage to the source document context. The design goal is not open-ended chart question answering but constrained, high-precision recovery of requested numeric fields with full provenance. Leveraging a dedicated AI platform for Life Sciences allows organizations to enforce these strict verification guardrails seamlessly.

How Traceable Graph Inference Works in Practice

A production-grade approach follows a clear sequence:

Traceable extraction of clinical figures into structured, source-linked data for human verification.

1. Locate the relevant figure within the full-text article or PDF.
2. Parse visual structure (axes, legends, data series, flow branches, or pie segments).
3. Map the requested field (for example, “survival probability at month 30 for the PSA50 responder subgroup”) onto the geometric elements.
4. Compute the numeric value from the visual representation.
5. Attach the exact figure identifier, caption text, and any supporting sentence used to disambiguate the request.
6. Surface the result with that provenance so a human reviewer can verify it against the original image in seconds.

This workflow differs from end-to-end multimodal generation. The system does not produce free-form answers; it returns structured fields that can be checked independently. Traceability is not an optional audit log. It is the mechanism that makes the extraction usable in regulated evidence work. Modern Life science AI architectures prioritize this deterministic auditability over generative freedom, ensuring complete confidence in extracted figures.

Validation Design and Scope

Independent subject-matter experts with extensive experience in SLRs  performed the validation. They had no role in developing the extraction technology. Reviewers scored each AI-inferred value directly against the source figure.

The scope included data points that appear nowhere in the article text and can be obtained only by reading the figure. Seven chart types were examined:

• Forest plots present point estimates, confidence intervals, standardized mean differences, odds ratios, and heterogeneity statistics.
• Bar charts compare treatment-line usage rates and subgroup response percentages.
• Patient flow diagrams track enrolment, discontinuation, and attrition counts.
• Sankey diagrams illustrate flow proportions and treatment sequencing, including non-mutually exclusive groups.
• Kaplan-Meier curves show survival or response probabilities at specific time points and indicate when a probability threshold is crossed.
• Line charts display measurement values across groups and quadrants.
• Pie charts provide proportional breakdowns of categories and mutation status.

The review covered more than 20 published articles and approximately 81 individual graph-inferred data points. Therapeutic areas included oncology, dermatology, ophthalmology, and related fields. Every value carried full traceability to its source figure and supporting context, proving that automated visual data extraction can achieve clinical-grade reliability.

Results Across Chart Types

Overall accuracy on the reviewed graph-inferred values reached approximately 98 percent. Performance held across therapeutic areas and study designs rather than succeeding only on a narrow subset of clean figures.

Breakdown by format (approximate counts from the validation set):

Concrete examples illustrate the nature of the task. The system recovered a standardized mean difference of 0.18 from a forest plot. It identified 48 percent second-line enzalutamide use from a bar chart and extracted 553 patients assessed for eligibility from a patient flow diagram. From a Kaplan-Meier curve, it estimated an approximate 50 percent probability of response at day 29 for a specific dose arm. The workflow links every extracted value to the exact figure and caption so reviewers can independently verify the result.

Optimizing Visual Data Extraction in Complex Charts

While visual data extraction handles most standardized charts seamlessly, dense multi-arm graphics present unique visual challenges that require refined engineering solutions.

The validation deliberately documented failure modes rather than presenting only successes. Two conditions currently limit reliability:

• Multi-arm Kaplan-Meier curves with closely grouped lines. Single-arm and well-separated two-arm curves are read reliably. When more than two arms cluster tightly at the queried time point, visual separation shrinks and accuracy declines.
• Funnel plots. These require identification of discrete, often unlabeled scatter points rather than continuous lines or bars. Current performance on this format does not yet match the seven validated types.

These are treated as active engineering targets rather than permanent constraints. Additional risks common to any visual extraction pipeline include degraded image quality in older PDFs, non-standard axis scaling, and figures that combine multiple chart types without clear separation. Human review remains essential for high-stakes use; the system’s value lies in accelerating the extraction step while preserving an auditable trail. Solutions like MadeAi provide the infrastructure needed to maintain human oversight while scaling review speed.

Benefits and Use Cases

When accuracy and traceability are both high, several concrete benefits appear:

• Reduced manual digitization time for figure-heavy reviews.
• Inclusion of previously inaccessible quantitative results in meta-analyses and network meta-analyses.
• Consistent handling of common clinical chart types across large literature sets.
• Audit-ready provenance that supports regulatory and internal quality processes.

Primary use cases include:

• Oncology SLRs that rely heavily on forest plots and Kaplan-Meier curves for overall survival and progression-free survival.
• Real-world evidence reviews that use Sankey diagrams and patient flow figures to map treatment sequences in metastatic castration-resistant prostate cancer and similar populations.
• Dermatology and ophthalmology reviews that extract response probabilities or quantitative imaging measurements from line and bar charts.
• Any evidence synthesis workflow that must combine text-extracted and figure-inferred data under a single quality framework. For insights into AI interactive query capabilities, see our article on conversational AI in literature reviews.

Best Practices for Reliable Adoption

Teams integrating graph inference into SLR pipelines should follow several practices:

• Require full provenance for every numeric field. Reject any extraction that does not link to the exact figure and supporting text.
• Limit automated extraction to validated chart types and conditions. Route dense multi-arm Kaplan-Meier curves and funnel plots to human digitization.
• Maintain human-in-the-loop review for all inferred values used in formal analyses.
• Version extraction models and validation datasets to regularly evaluate performance as literature formats evolve.
• Reconcile figure-inferred values with text-extracted values against the original source document before combining them.

These steps convert a promising technical capability into a controlled process that meets the reproducibility expectations of evidence-based decision-making.

About the Future 

Research on multimodal models continues to improve geometric reasoning and value recovery. Specialized training regimes that emphasize progressive chart reading (structure first, then precise values) show measurable gains over general vision-language training. Domain-specific fine-tuning on clinical figure corpora is likely to further reduce error rates on multi-arm survival curves and scatter-based plots.

Parallel progress in standardized benchmarks for chart data extraction will allow more transparent comparison across systems. As these benchmarks mature, the gap between laboratory demonstrations and production reliability in regulated workflows should narrow. The practical requirement will remain the same: every extracted number must still carry an independent path back to its visual source.

Conclusion

AI systems can now read many of the charts and graphs that previously required manual digitization. The critical question is not whether the technology can produce a number, but whether that number can be trusted and independently verified. Expert validation across seven common clinical chart types demonstrates that high accuracy is achievable when the system is designed for constrained inference and full traceability.

Limitations remain, particularly for dense multi-arm survival curves and funnel plots. Transparent reporting of those boundaries is what makes the successful results usable. For SLR teams, the practical path forward is selective automation of validated chart types, mandatory provenance, and continued human oversight. That combination turns visual data from a persistent bottleneck into a reliable component of evidence synthesis.

Author’s Note: This article was supported by AI-based research and writing, with Claude 5 assisting in the creation of text and images.