Extracting Data from Biomedical and Scientific Charts

The Meta-Analysis Problem

Biomedical research advances through synthesis. Systematic reviews and meta-analyses combine results from multiple independent studies to arrive at stronger conclusions than any single experiment can provide. Organizations like the Cochrane Collaboration produce thousands of these reviews each year, covering topics from drug efficacy to surgical outcomes to public health interventions. But there is a persistent bottleneck in this process — getting the actual data out of published figures.

The problem is straightforward: many published papers present their key results exclusively as charts and graphs, without providing the underlying numerical data in tables, supplementary files, or public repositories. A 2019 analysis found that fewer than 20% of biomedical studies deposited raw data in accessible formats. When a research team conducting a Cochrane review needs to include a study in their quantitative synthesis, they often face a choice: contact the original authors and wait weeks or months for a response that may never come, or extract the data directly from the published figures.

Contacting authors has a notoriously low success rate. Studies report response rates ranging from 10% to 40%, and even when authors respond, the data they provide may be incomplete or in an incompatible format. For older studies — particularly those published before data-sharing norms became common — the original datasets may no longer exist. The researchers who generated the data may have retired, changed institutions, or simply lost track of files from years or decades ago.

This makes chart digitization an essential skill in evidence-based medicine. The Cochrane Handbook explicitly acknowledges that extracting data from figures is sometimes the only viable approach, and provides guidance on how to do it reliably. Tools that speed up this process without sacrificing accuracy have a direct impact on the quality and completeness of systematic reviews. Every study that can be included rather than excluded strengthens the evidence base that clinicians use to make treatment decisions. For a general overview of chart data extraction methods, see our complete guide to extracting data from charts.

Common Chart Types in Biomedical Research

Biomedical and scientific publications rely on a specific set of chart types, each optimized for the kind of data and comparisons that researchers need to communicate. Understanding what each chart encodes is the first step toward extracting data from it accurately.

Error bar plots

Error bar plots are arguably the most common chart type in experimental biomedical research. They display a central measure — typically a mean — along with bars that extend above and below to represent variability or uncertainty. These charts appear in nearly every paper that compares experimental conditions: drug concentrations, treatment groups, time points, or cell lines. The bars encode critical statistical information, but their meaning varies depending on the study. They might represent standard deviation (SD), standard error of the mean (SEM), or 95% confidence intervals (CI). This distinction matters enormously for meta-analysis, because each measure carries different implications for how the data should be pooled.

Box plots

Box plots (also called box-and-whisker plots) summarize entire distributions in a compact visual form. The box spans from the first quartile (Q1) to the third quartile (Q3), with a line at the median. Whiskers extend to the minimum and maximum values, or to 1.5 times the interquartile range, with individual outlier points plotted beyond. In clinical research, box plots are widely used for comparing distributions across treatment groups — for example, blood pressure readings before and after intervention, tumor sizes across different dosage cohorts, or biomarker levels in patients versus healthy controls.

Scatter plots

Scatter plots map the relationship between two continuous variables. In biomedical research, they are standard for dose-response studies, correlation analyses, and pharmacokinetic modeling. Each point represents a single observation or subject. Researchers often overlay trend lines, regression curves, or confidence bands to summarize the relationship. Extracting data from scatter plots is particularly valuable when the original study reports only the correlation coefficient but the meta-analyst needs the raw paired observations.

Heatmaps

Heatmaps use color intensity to represent values in a matrix, with rows and columns corresponding to different variables. In genomics and proteomics, they are ubiquitous — gene expression heatmaps show thousands of genes across multiple experimental conditions, with red typically indicating upregulation and blue indicating downregulation. Protein interaction maps, metabolomics profiles, and correlation matrices between clinical variables all use the heatmap format. The challenge for data extraction is translating color values back into numbers, which requires understanding the color scale and its mapping to the data range.

Survival curves (Kaplan-Meier)

Kaplan-Meier survival curves are the standard way to present time-to-event data in clinical trials and epidemiological studies. They show the proportion of patients surviving (or remaining event-free) over time, with characteristic step-wise decreases at each event. Tick marks indicate censored observations — patients who were lost to follow-up or whose study period ended. Extracting coordinate data from survival curves allows meta-analysts to reconstruct individual patient data, which enables more sophisticated pooled analyses than simply combining hazard ratios. For more examples of chart types and what data they contain, visit our use cases page.

Extracting Data with Error Bars

Error bars carry the uncertainty information that makes biomedical data meaningful. A mean value alone tells you very little — it is the variability around that mean, combined with the sample size, that determines whether a result is statistically significant and how much weight it should receive in a pooled analysis. When extracting data from error bar plots, capturing the bar endpoints is just as important as capturing the central values.

Plot2Data's error bars extraction feature is designed specifically for this task. When enabled, the AI identifies not only the central value (mean or median) for each data point but also the upper and lower bounds of the error bars. The extracted data includes three values per point: the central measure, the upper limit, and the lower limit. This gives you everything you need to calculate the standard deviation, standard error, or confidence interval width depending on what the error bars represent.

Interpreting what the error bars mean

Before using extracted error bar data in a meta-analysis, you must determine what measure of uncertainty the bars represent. This information should be stated in the figure legend or methods section of the paper. Here is why it matters:

• Standard deviation (SD) describes the spread of the raw data. It does not change with sample size. If the error bars represent ±1 SD, the extracted range captures approximately 68% of the data distribution.
• Standard error of the mean (SEM) describes the precision of the estimated mean. It decreases as sample size increases (SEM = SD / √n). SEM bars are always smaller than SD bars for the same data, which is why some authors prefer them — they make results look more precise. For meta-analysis, you can convert SEM back to SD if you know the sample size: SD = SEM × √n.
• 95% confidence intervals provide a range within which the true population mean likely falls. For normally distributed data, the 95% CI is approximately ±1.96 × SEM. CIs are the most directly useful format for meta-analysis because they convey both the estimate and its uncertainty in a single measure.

Practical tips for error bar extraction

• Always enable the error bars option in Plot2Data when processing charts that display uncertainty measures
• Check whether the error bars are symmetric — asymmetric bars often indicate log-transformed data or non-parametric statistics
• Cross-reference extracted bar heights against any values reported in the paper's text or tables to validate accuracy
• Record the type of error bar (SD, SEM, or CI) alongside your extracted data so that downstream analyses use the correct formulas

Working with Box Plots in Clinical Data

Box plots encode five key summary statistics: the minimum, first quartile (Q1), median, third quartile (Q3), and maximum. Some box plots also display the mean as a separate symbol (often a diamond or cross) and show individual outlier points beyond the whiskers. Extracting these values from published figures provides a rich description of each group's distribution, which is far more informative than a simple mean and standard deviation.

In clinical research, box plots frequently appear in studies comparing biomarker distributions across patient groups, treatment outcomes across dosage levels, or quality-of-life scores before and after intervention. The median and interquartile range (IQR = Q3 – Q1) from extracted box plot data can be used directly in meta-analyses that handle non-normally distributed outcomes, or they can be converted to estimated means and standard deviations using published formulas when a meta-analysis requires those statistics.

Extracting the five-number summary

When digitizing a box plot, each box yields up to five data points per group:

• Minimum — the lower end of the bottom whisker (or the lowest non-outlier observation)
• Q1 — the bottom edge of the box
• Median — the horizontal line inside the box
• Q3 — the top edge of the box
• Maximum — the upper end of the top whisker (or the highest non-outlier observation)

Handling outlier points

Many box plots display individual outlier points as dots or circles beyond the whiskers. These represent observations that fall outside 1.5 × IQR from the edges of the box. When extracting data, record these outlier values separately. They affect the true minimum and maximum of the dataset and may indicate important clinical phenomena — for example, patients who responded unusually well or poorly to a treatment. For meta-analysis purposes, note whether the reported whiskers extend to the true min/max or only to the 1.5 × IQR boundary, as this affects how you interpret the range of the data.

From box plots to meta-analysis

Several well-established methods exist for converting box plot statistics into the means and standard deviations that most meta-analysis software requires. Wan et al. (2014) and Luo et al. (2018) provide formulas that use the median, Q1, Q3, minimum, maximum, and sample size to estimate the mean and SD. Having accurately extracted box plot data is the essential input for these conversions. Even small errors in reading the median or quartile positions can propagate through the formulas, so using an AI-powered tool that can read these positions precisely from high-resolution images offers a meaningful advantage over manual estimation.

Scatter Plots in Biomedical Research

Scatter plots are the workhorse of correlation and regression analysis in the life sciences. They appear in pharmacology (dose-response relationships), epidemiology (exposure versus outcome), diagnostics (comparing two measurement methods via Bland-Altman plots), and genetics (genome-wide association studies where effect sizes are plotted against chromosomal position). Each point represents a paired observation, and the spatial distribution of points reveals the strength and shape of the relationship between variables.

Extracting trend line data alongside raw observations

Many biomedical scatter plots include fitted curves: linear regression lines, polynomial fits, logistic curves, or locally weighted smoothing (LOESS). These trend lines represent the model that the original authors fitted to their data. When extracting data from scatter plots, you may want both the raw data points and a set of coordinates sampled along the trend line. The raw points allow you to re-fit models using your own methods, while the trend line coordinates provide a quick reference for the original authors' interpretation of the relationship.

Handling dense data sets with many overlapping points

A common challenge in biomedical scatter plots is overplotting — when multiple data points occupy the same or nearly the same position on the chart. This is especially prevalent in studies with large sample sizes, such as clinical trials with hundreds of patients or genomics experiments with thousands of genes. When points overlap, even the most careful digitization can only recover the visible points, not those hidden beneath.

Several strategies help mitigate this issue. First, look for the figure in the paper's supplementary materials, which sometimes includes versions with jitter (small random offsets) applied to reduce overlap. Second, check whether the paper reports the total sample size — if your extracted point count is significantly lower than N, overlap is likely hiding data. Third, AI-powered tools like Plot2Data can sometimes detect and separate partially overlapping points that would be indistinguishable to the human eye, particularly when points are semi-transparent or slightly offset. For a deeper comparison of manual versus AI approaches to these challenges, see our guide on manual vs AI chart digitization.

Heatmaps in Genomics and Proteomics

Heatmaps are among the most information-dense charts in biomedical research. A single gene expression heatmap might encode tens of thousands of values — each cell in the matrix representing the expression level of one gene under one experimental condition. Extracting numerical data from heatmaps is fundamentally different from extracting data from scatter plots or bar charts because the data is encoded in color rather than position.

Gene expression matrices

In transcriptomics studies, heatmaps display normalized expression values (such as log2 fold changes or Z-scores) across rows of genes and columns of samples or conditions. The color scale typically runs from blue (low expression or downregulation) through white (baseline) to red (high expression or upregulation). Extracting these values requires the AI to interpret the color of each cell, map it to the color scale legend, and assign the corresponding numerical value. For heatmaps with a clearly visible color bar and well-separated cells, Plot2Data can recover the matrix values with reasonable accuracy.

Protein interaction maps and correlation matrices

Protein-protein interaction data is often displayed as symmetric heatmaps where both axes represent the same set of proteins, and the color at each intersection indicates the strength of interaction (measured by co-immunoprecipitation, yeast two-hybrid assays, or computational prediction scores). Similarly, correlation matrices use heatmaps to show pairwise Pearson or Spearman correlations between clinical variables, metabolites, or gene expression profiles. These matrices are symmetric, which provides a useful consistency check — values above and below the diagonal should mirror each other.

Challenges specific to heatmap extraction

• Color scale resolution: Subtle color differences can represent meaningful data distinctions. Image compression (especially JPEG) can blur color boundaries and reduce extraction accuracy. Always use the highest-quality image available, preferably from a PDF or SVG source.
• Row and column labels: Heatmaps in genomics papers often have tiny, densely packed gene names or sample identifiers along the margins. These labels are essential for interpreting the extracted matrix, but they may be too small to read in a low-resolution screenshot.
• Hierarchical clustering dendrograms: Many heatmaps include dendrograms (tree diagrams) along the margins that show how rows and columns were clustered. These visual elements do not contain numerical data but can confuse extraction if the tool interprets them as part of the data matrix. Cropping the dendrogram before extraction can improve results.

Best Practices for Biomedical Chart Extraction

Whether you are building a Cochrane review, replicating a published analysis, or assembling training data for a machine learning model, following consistent best practices ensures that your digitized data is accurate, traceable, and defensible.

Maximize image quality

• Download figures directly from the journal's website rather than copying from a PDF viewer, as PDFs sometimes render charts at reduced resolution
• Use the "download high-resolution figure" option that many journals provide alongside each figure
• Check supplementary materials — they often contain the same figures at higher resolution or in vector format (SVG, EPS)
• When taking screenshots, zoom in to the chart first to capture more pixels per data point

Handle composite figures correctly

Biomedical papers frequently use composite figures with multiple panels labeled A, B, C, and so on. Each panel is typically a separate chart with its own axes, scales, and data. For best extraction results, crop each sub-panel into a separate image before uploading. This prevents the AI from conflating data across panels and ensures that axis labels from one panel are not misattributed to another. Most image editors and screenshot tools make it easy to select and save rectangular sub-regions.

Verify extracted data against reported statistics

Published papers almost always report some numerical values in the text or tables — sample sizes, group means, p-values, or summary statistics. Use these as ground truth checkpoints. After extracting data from a figure, compare your digitized values against any numbers reported elsewhere in the paper. For example, if the paper states that the treatment group had a mean blood pressure of 128 mmHg with SD of 14, and your extracted value from the error bar chart gives 126 ± 13, you can be confident that the extraction is accurate within a reasonable margin. Discrepancies of more than 5–10% warrant re-examination of the extraction or the image quality.

Document and cite digitized data properly

When using digitized data in your own publications, transparency is essential. Report that data was extracted from published figures, name the tool used (e.g., Plot2Data), and note any assumptions made during the process. Many systematic review guidelines, including PRISMA, require authors to state how data was obtained from each included study. A typical methods section statement might read: "Data not reported numerically were extracted from published figures using Plot2Data (www.plot2data.com), an AI-powered chart digitization tool. Extracted values were cross-checked against statistics reported in the text."

Maintain an audit trail

• Save the original chart image alongside your extracted data for reproducibility
• Record the source (DOI, figure number, panel label) for every digitized chart
• Note whether error bars represent SD, SEM, or CI, and include the sample size (N) for each group
• Keep a log of any discrepancies between extracted and reported values, along with how they were resolved
• Store extracted data in a structured format (CSV or spreadsheet) with clearly labeled columns