Plot Data Extractor: Methods and Validation

Why extract numerical data from figures?

Numerical results needed for evidence synthesis are sometimes available only in figures. The Cochrane Handbook advises review authors to consider software-based extraction when numerical data are unavailable elsewhere, while also emphasizing accurate, transparent and reproducible data collection.

Plot Data Extractor Pro is designed for this situation. It does not infer the underlying data automatically; the user identifies labelled axis references and places markers on the plotted values. The resulting calculations are deterministic from those inputs.

Open the Plot Data Extractor product page or read the Plot Data Extractor Pro User Guide.

PDF rendering and local processing

Research PDFs are rendered in the browser using a locally bundled copy of PDF.js. PDF.js is a Mozilla-led, web-standards-based platform for parsing and rendering PDFs. The bundled library is served from the Blazon Scientific plugin rather than loaded from a third-party CDN at runtime.

Source PDFs, calibration markers and extracted scientific values remain in the browser during the extraction workflow. See the Pro privacy information for the full data-handling description.

Axis calibration

Plot Data Extractor Pro stores marker positions in the intrinsic coordinate system of the selected plot image. Scientific values are derived from those raw positions and the active calibration, so changing a calibration can recalculate the scientific value without replacing the original selected position.

Linear calibration

Given two reference positions p₁ and p₂ corresponding to known values v₁ and v₂, a selected position p is converted using linear interpolation:

The same form applies to either numerical axis. Reversed axes and non-zero origins are supported because the direction and range are determined by the two user-defined references rather than assumed from screen orientation.

Log10 calibration

For a log10 axis, interpolation is performed in log10-transformed value space. If L₁ = log₁₀(v₁) and L₂ = log₁₀(v₂), then:

Users enter the actual numeric values represented by labelled ticks—for example, 1 for 10⁰, 10 for 10¹, and 100 for 10². The logarithmic transformation is applied internally.

Central values and uncertainty

The extraction workflow separates the central estimate from its uncertainty representation. Supported structures include a central value alone, one-sided symmetric error, two-sided symmetric error, asymmetric bounds and XY points.

Two-sided symmetric error

The user selects the central value and the visible upper and lower error endpoints. The exported symmetric error magnitude is calculated as:

The central marker is retained independently. If the selected central value is materially away from the midpoint of the two error endpoints, the interface can warn the user rather than silently redefining the central estimate.

One-sided symmetric error

When only one visible endpoint is available, the error magnitude is:

Asymmetric bounds

For genuinely asymmetric intervals, the lower and upper bounds are retained separately. The analysis CSV does not force them into one symmetric error value.

Controlled validation

A controlled PDF was generated from exactly known values to test common and deliberately challenging plotting conditions. The validation set included bar charts and XY plots, linear and logarithmic axes, SD and SE error bars, asymmetric 95% confidence intervals, and a deliberately low-resolution/dense-label figure.

The completed dataset contained 33 plotted records and two extraction attempts per record, yielding 66 central-value extractions, 50 symmetric-error extractions, 32 asymmetric-bound extractions and 148 extracted numerical quantities in total.

All applicable pre-specified controlled-validation accuracy and repeatability thresholds were met.

Published-paper external validation

A separate real-world validation used five figure panels from independent open-access eLife articles. Publisher-supplied source-data spreadsheets were used to derive ground-truth values only after the plot extractions had been completed.

The external validation contained 26 unique plotted records, each extracted twice. For the accuracy analysis, the two attempts were averaged within each unique record so that repeated measurements were not incorrectly treated as independent observations. Repeatability was analysed separately from the paired attempts.

The logarithmic-axis panel was the most challenging published condition. Excluding that panel, the other 24 published records had a 95th-percentile central-value error of 0.67% and a maximum error of 0.72%.

Plot Data Extractor Pro was evaluated using both controlled figures with known numerical values and an independent set of figures from published scientific papers. In the published-paper validation, 26 plotted values from five figure panels were extracted twice. The median absolute error for central estimates was 0.31% relative to the published source values, with a 95th-percentile error of 1.46%. Repeated extractions differed by a median of 0.13% of the source value. These findings complement controlled validation in which the median central-value error was approximately 0.13% of the plotted axis span.

Published panels used in the external validation

• eLife 100490, Figure 2A — cell-volume data with publisher-supplied source data.
• eLife 100129, Figure 2B — quantitative metabolic data with publisher-supplied source data.
• eLife 75233, Figure 2B — MDGA localisation data with raw/source data supplied for the figure.
• eLife 102222, Figure 4D — boxplot around the median with quantitative source data.
• eLife 71557, Figure 3A — logarithmic y-axis with values reported as averages and SD.

Limitations

• The validation was performed by a single operator; repeatability therefore describes within-operator repeatability rather than inter-rater reproducibility.
• The external validation included five published panels and 26 unique records. It complements, rather than duplicates, the broader controlled validation dataset.
• Plot digitisation remains dependent on figure resolution, line thickness, compression, overlapping symbols and marker-placement precision.
• Very short error bars can have relatively large percentage errors even when the absolute graphical placement error is small.
• For the eLife 102222 boxplot, the median was independently validated; quartile edges were not used as primary accuracy evidence because the source material did not specify the exact quartile algorithm used to draw the box.
• These results describe the tested datasets and should not be interpreted as a guarantee of identical accuracy for every possible figure.