SCOPE-MOVE

1. Start with the Overview to confirm the purpose of the app, the evidence base size, and the main navigation routes.
2. Use the Primary Tool to compare validation-level estimates. Select the metric, filter by domain, model group, age group, device, validation type, environment, and protocol, then inspect the summary tiles and forest view.
3. Read the pooled line as an orientation aid, not as a universal benchmark. Rows above the pooled line and rows marked as external validations should be checked against the study context, protocol, population, device, and source paper.
4. Use Model Ranking for an exploratory shortlist of externally validated model identities. Expand a model card to review devices, age groups, validation types, and studies that externally validated the model.
5. Use Studies to understand the broader evidence landscape, including publication timing, countries, device brands, participant characteristics, and the compact prediction model type chart.
6. Use Quality to inspect study-level risk-of-bias summaries and the concise signalling-question wording behind SQ1-SQ11.
7. Use Data to inspect the bundled source and derived tables when you need extraction-level detail or want to audit a displayed result.
8. Use DOI links and the data dictionary when selecting or reusing a candidate method. The app supports evidence-guided selection; it does not declare a universally best model.

This list gives the minimal wording behind SQ1-SQ11 for interpreting the app summaries. It is not an instruction manual for re-performing the assessment.

The data dictionary describes the raw input files used to generate the analyses and app data bundle. App-facing definitions, interpretation notes, and hover-field explanations are presented in the Definitions and field guide panel above.

Open raw-input data dictionary

• Each plotted point is the extracted validation performance metric for a specific validation. The classifier metrics are F1, sensitivity, and specificity.
• Metrics were extracted as the unweighted mean and standard deviation across behaviour classes when reported.
• When mean and SD were not reported but class-specific performance was reported, metrics were macro-averaged across classes to obtain an unweighted mean and SD.
• When a confusion matrix included validation counts for each behaviour class, metrics were computed and macro-averaged across classes to obtain an unweighted mean and SD.
• Extracted SDs were converted to sampling variances and divided by cross-validation folds when SDs reflected CV, or by number of behaviour classes when SDs reflected class variation. When SDs were unavailable, a lower-bounded behaviour-domain-specific median constant variance was used to stabilise weights.
• The constructed variance was used to build the displayed 95% confidence interval for each validation estimate.
• For meta-analysis, classifier metrics are analysed on the logit scale and back-transformed to proportions for display.
• The covariate-adjusted orientation line represents a browser-side orientation estimate for the selected subset, using the review covariate set where available. The simple pooled line is a random-effects line with no covariates.
• The covariate-adjusted orientation line uses the review covariate set: centred number of behaviour classes, centred validation folds, missingness indicators, test environment, test protocol, and behaviour domain.
• Covariates that are constant or effectively unavailable in the filtered subset are dropped from that pooled-line calculation and listed under the plot.
• The simple random-effects line uses the validation table variances and includes no covariates.
• The dashed pooled line can mask heterogeneity across behaviours, protocols, devices, and populations.
• Missing metrics reflect reporting and extraction availability, not automatic exclusion by this app.