The information in diagnostic tests

Methodological evidence map and secondary synthesis of diagnostic accuracy reviews.

Analysis completed; manuscript version 1.11.0 released on September 8, 2026.

The complete analysis and manuscript release are documented by September 8, 2026. The exact review completion date is not separately recorded.

Corresponding author: Shuhan He, she8@mgb.org.

Shuhan He conceived the study, assembled the structured review data, specified and performed the analyses, interpreted the findings, created the figures and drafted the article. The other named authors contributed to interpretation and critically revised the manuscript. Shuhan He is the guarantor. OpenAI Codex assisted retrieval, code and language editing under author review.

Across reproducibly accessible diagnostic accuracy reviews, how much uncertainty about a binary target condition is a test or threshold expected to resolve at a stated starting probability, and how does this information relate to conventional test-performance measures?

Diagnostic testing changes what is known about a target condition. Sensitivity and specificity describe test performance; likelihood ratios and post-test probabilities describe an observed result. Mutual information expresses the expected reduction in uncertainty before either result is known. Earlier studies established diagnostic applications of this measure. This study assembles a broad empirical reference scale from accessible diagnostic-review data so readers can interpret a diagnostic profile at a stated starting probability and compare its information with conventional accuracy measures. Information describes learning about a binary target. Clinical value also depends on consequences, available actions, and patient preferences.

• Build an empirical reference distribution of expected diagnostic information from source-defined tests and thresholds.
• Compare uncertainty resolved at 5%, 20%, and 50% starting disease probability and across a 1%-50% grid.
• Examine agreement and differences between diagnostic information and Youden's J using source-comparable examples.
• Distinguish expected information before testing from the probability update and uncertainty change after each result.

• Distribution of 100 x I(D;T)/H(D) at 5%, 20%, and 50% starting probabilities, with corresponding information in bits; each pooled estimate has equal weight.
• Information profiles from 1% to 50% in one-percentage-point steps.
• Spearman correlations with Youden's J and descriptive near-tie comparisons.
• CEA information crossings and Ottawa ankle rule result branches.
• Sensitivity to zero-cell correction and the regularized direct-binomial estimator.

Include

• Diagnostic-review data with explicit study identifiers, source links, and non-negative integer TP, FP, FN, TN counts.
• Groups with at least five unique study identifiers, one result per identifier, and at least three studies with both disease states represented.
• Aggregate sensitivity plus specificity at least one in the source-defined orientation, with one canonical representation after exact row-signature deduplication.
• Historical Cochrane groups linked to review-author-reported pooled findings; modern Cochrane main analyses; at most one canonical main table per OSF, Zenodo, or PMC review under the recorded priority rules.

Exclude

• Too few studies, repeated identifiers within a group, too few studies representing both disease states, or below-chance performance in the original orientation.
• Counts reconstructed from rounded accuracy measures, pooled-only or consequence tables, and unsupported narrative transcription.
• Duplicate representations and alternatives not selected by source-specific primary designation; preserve their status in the audit record.

Secondary methodological synthesis of diagnostic accuracy reviews with accessible study-level cross-classification data; original designs are inherited from each source review.

No common date window: historical Cochrane is a fixed 2018 release; PMC uses 2020-2026 publication years; other sources use recorded inventories or queries. Source-specific limits govern inclusion.

Search terms were English. The recorded API queries contain no language filter; eligibility depended on accessible structured study-level tables. No translation workflow was used in the recorded computational extraction.

Public OSF, Zenodo, Figshare, Dryad, and SRDR deposits were examined for reusable diagnostic-review data. Dryad and SRDR contributed no primary pooled estimates. This accessible evidence sample is not an exhaustive search of unpublished studies or all diagnostic medicine.

Searches are reproducible from the recorded queries, fixed inventories, dates and checksums. The Cochrane benchmark and two modern review packages were used for source-format validation. No separate search-validation exercise against an independent seed set is documented.

The recorded acquisition used public repository files, published review packages and machine-readable PMC tables. No author-supplied private dataset or author-contact campaign is documented in this evidence base.

This is a fixed evidence map. The registered analysis uses the dated inventories below; no rolling search update is part of this version. Any later update will be recorded as an amendment.

The seven frozen inventories contain 3,979 discovery records. The primary analysis includes 210 reviews. Public discovery CSVs retain identifiers, titles, source links and recorded source terms; abstracts and raw documents are omitted. Their hashes are distinct from the original inventory hashes.

Retain source identifiers and compare exact normalized count-row signatures. Keep one result per identifier within a group; repeated identifiers make that group ineligible. The same study may contribute to different diagnostic profiles, so study-result appearances are not unique studies.

Shuhan He directed the computational selection and author audit. Fixed source priorities resolved representations; ambiguous groups were excluded or retained outside the primary sample. Source identities and metadata remained visible. The recorded workflow used one screener; no inter-reviewer agreement statistic was reported.

Inspect structured review deposits and machine-readable PMC tables for explicit study-level counts. Preserve article, table, file and review identifiers and exclusion reasons. Selection used the single author-directed workflow described above.

Codex assisted retrieval, parsing and code drafting. Selection used deterministic rules with author oversight.

Parse explicit study identifiers and integer TP, FP, FN, TN cells while retaining source location, orientation and checksums. Do not reconstruct counts from rounded accuracy metrics. Apply canonical workbook/table rules. Shuhan He directed extraction and rechecked 50 table decisions and 1,640 extracted cells; the recorded audit found no acceptance or transcription errors. This was an author audit within the same workflow, with source identities retained.

Groups require explicit study identifiers and integer 2x2 counts. Unavailable counts are not imputed or reconstructed from rounded accuracy measures. Missing threshold, setting, population and reference-standard descriptors remain missing and limit contextual comparisons.

Interpret the estimates in the context of source-reported risk of bias and clinical setting. The atlas did not undertake a uniform study-level QUADAS-2 assessment; source reporting varies.

Pool each eligible source-defined test or threshold separately with bivariate random-effects REML on sensitivity/specificity logits, adding 0.5 to all four cells in every study. Back-transform the fitted mean logits. Compute entropy, mutual information and percentage uncertainty resolved at the specified starting probabilities. Give each pooled estimate equal reference-distribution weight. Report pooled-mean intervals and plug-in predictive ranges conditional on fitted heterogeneity. Bootstrap reference medians over pooled estimates with 2,000 percentile replicates and a fixed seed; this does not propagate within-estimate uncertainty. Repeat analyses with correction only in zero-containing studies and with the regularized direct-binomial estimator. Treat within-review contrasts and CEA crossings as descriptive; no paired superiority inference or clinical utility ranking is supported. Distinguish signed entropy change, posterior-to-prior KL divergence, log2 likelihood ratio, and expected mutual information.

The completed analysis contains 444 converged eligible bivariate fits. The primary reference scale retains 273 pooled estimates from 210 reviews and 4,104 study-result appearances. At a 20% starting probability, the median estimate removes 29.2% of starting uncertainty (0.211 of 0.722 bits).

Shuhan He specified and performed the analyses. The workflow retained source identities and used deterministic scripts, numerical tests, source-summary comparisons and estimator sensitivity checks. No independent blinded second synthesis is documented.

No formal publication-bias analysis is part of this methodological reference scale. The distribution describes reproducibly accessible reviews across different tests and conditions; it is not a pooled estimate of one clinical effect.

Report source fidelity, numerical diagnostics, estimator sensitivity, model-based uncertainty and sample limitations. This methodological reference scale does not assign GRADE ratings to clinical effects or recommendations.

The archived release environment is CPython 3.13.12 on Windows x86-64, NumPy 2.4.4, pandas 3.0.2, SciPy 1.17.1 and Matplotlib 3.10.8. The downloadable code package includes the complete pinned requirements and original script versions.

Open the complete ConductScience Data release

Download the frozen data and analysis archive

Dataset DOI: https://doi.org/10.55157/csd.di.2026.001. Crossref registered this DOI on September 14, 2026 local time (September 15, 2026 at 1:59 AM UTC), after the data release was published on September 14, 2026.

View the ConductScience Data record

Cite the dataset

Download release 1.13.3 (ZIP)

Inspect the file manifest

The historical Cochrane DTA Reference Dataset is available at https://zenodo.org/records/1303259 under CC BY-NC 4.0. Other source data retain their original terms. The protocol's CC BY 4.0 license applies to protocol text; it does not relicense third-party data or code.

National Academy of Medicine, Agreement 2026A008797. The manuscript states that the funder had no role in design, acquisition, analysis, interpretation, writing, or submission.

The authors declare support from the National Academy of Medicine for the submitted work (Agreement No. 2026A008797); no other competing interests are declared.

The study was registered internally with ConductScience on 16 August 2026. The public protocol record and its DOI were established on 13 September 2026. The protocol has a versioned amendment history: source-priority rules developed during staged audits were fixed before the final reference analysis. This version documents the completed methods and distinguishes the original internal registration from subsequent amendments and public posting.

• August 16, 2026: Internal study registration with ConductScience; subsequent method amendments retain their own history.
• September 8, 2026: Manuscript version 1.11.0 records the completed analysis.
• September 13, 2026: Public protocol record posted and DOI created by Crossref.
• September 13, 2026: Protocol version 1.2.0 clarifies date meanings and adds the documented methods and support files.
• September 13, 2026: Protocol version 1.2.1 improves navigation, formatting and file access, and labels the registration and publication dates. Scientific methods and supporting files are unchanged.
• September 14, 2026: Protocol presentation 1.3.0 links the unchanged registered protocol to the complete ConductScience Data release version 1.13.3. The dataset DOI is recorded only after Crossref verification.
• September 14, 2026: Crossref created dataset DOI 10.55157/csd.di.2026.001 on September 14 in the publisher's local time zone (01:59:05 UTC on September 15) for the ConductScience Data release published on September 14. The protocol remains registered August 16 and published September 13.

Previous protocol version 1.3.0 (JSON)

The DOI identifies this protocol record.

• Pilot search started: August 15, 2026
• Formal search started: August 16, 2026
• Screening started: August 16, 2026
• Extraction started: August 16, 2026
• Synthesis started: August 16, 2026

He S, Joseph JW, Liebovitz DM, et al. The information in diagnostic tests: registered protocol. ConductScience; 2026. Version 1.3.1. CSR-DI-2026-001. https://doi.org/10.55157/csr.di.2026.001

Registered with ConductScience: August 16, 2026. Published online: September 13, 2026.

Registration, date and source metadata (JSON)