Trustworthy RF/Microwave Biomedical AI: Information Fusion and Evaluation Design for Clinical Translation

This deposit is the companion dataset and reproducibility package for the manuscript "Trustworthy RF/Microwave Biomedical AI: Information Fusion and Evaluation Design for Clinical Translation" by Aldelemy, Al-Dulaimi, Ibrahim, Twigg, Qahwaji and Abd-Alhameed, submitted to Artificial Intelligence in Medicine (Elsevier) in 2026.

The manuscript is a PRISMA-ScR-informed critical scoping review and methodological audit of RF, microwave and ultra-wideband biomedical AI studies published between 1 January 2010 and 31 January 2026. It charts and deeply appraises 20 empirical studies (19 AI-inference studies plus 1 contextual translational anchor) against an 11-dimension RF-specific risk-of-bias framework, yielding 220 transparent ratings, and derives a tiered evidence framework (T0–T5) linking simulation feasibility, phantom realism, experimental robustness, single-centre clinical accuracy, multi-site external validation and prospective workflow utility to required evidence.

What this deposit contains

Protocol and reporting documentation (02_protocol/): PRISMA-ScR checklist, SANRA self-assessment, search strategy with the full Boolean string, inclusion/exclusion criteria (E1–E6), deviations log, and the protocol used to govern screening, charting and appraisal.

Machine-readable review data (03_review_data/): the search log with per-database counts (357 records identified), the deduplication log (231 after deduplication), title/abstract screening counts (63 full-text reports assessed), full-text exclusion reasons (E1–E6, 32 exclusions), the 22-field evidence map for all 20 charted studies, the full 220-cell risk-of-bias matrix (20 studies × 11 RF-specific signalling dimensions), per-study claim-tier adjudication, the denominator audit ledger, the author-linked-study sensitivity table, the worked source-specific audit example, the reusable pre-consensus checklist template, and Wilson 95% confidence-interval calculations for every headline proportion reported in the manuscript. A complete data dictionary is included.

Reusable templates (06_templates_for_reuse/): blank CSV and Excel templates that other research teams can use to run the same RF-specific risk-of-bias appraisal on their own corpora.

Reproducibility scripts (07_scripts/): four Python scripts (pure standard library, no third-party dependencies) that verify the PRISMA arithmetic, recompute the headline statistics (17/19, 11/19, 10/19, 12/20, 9/20), regenerate the Wilson 95% CIs, and produce the claim-tier divergence summary directly from the CSV files. All scripts exit cleanly with code 0 and reproduce every number cited in the manuscript.

Submission-support materials (08_upload_instructions/): the journal-upload checklist and the Data Availability statement to paste into the manuscript.

Supplementary workbook: Supplementary_Tables_S1_to_S5_PUBLICATION_READY.xlsx — a 12-sheet Excel file containing Tables S1 (PRISMA-ScR checklist), S2 (search log, PRISMA flow, exclusion reasons), S3 (SANRA self-assessment, deviations log), S4 (evidence map, 220-cell risk-of-bias matrix, claim-tier adjudication, denominator ledger), S5 (worked audit) and a blank checklist template.

Repository governance: MANIFEST.csv enumerating all 42 files with sizes, VALIDATION_REPORT.txt showing structural checks and live script outputs, AUDIT_FIX_REPORT_v2_2.md documenting the v2.1 → v2.2 corrections, LICENSE.txt (CC BY 4.0), and README.md.

Headline findings reproducible from this deposit

Across the 19-study AI-inference corpus: calibration or uncertainty reporting absent in 17/19 (89%; 95% Wilson CI 69–97%); preprocessing/tuning operators not confirmed to be fitted inside the resampling loop in 17/19 (89%); subgroup or worst-group performance with interval estimates absent in 11/19 (58%; CI 36–77%); multi-site or device-level external validation absent in 10/19 (53%; CI 32–73%); auditable code or split-file release absent in 17/19 (89%). At corpus level (20 studies): claim-tier divergence in 12/20 (60%; CI 39–78%) and high overall concern in 9/20 (45%; CI 26–66%). Sensitivity analyses excluding the author-linked study do not change the qualitative ordering of these findings.

Author-linked study disclosure

One eligible peer-reviewed paper in the corpus is co-authored by the first reviewer: Aldelemy et al. 2026, "Development of a Traceable Dielectric-Property Dataset for Femoral-Fracture Monitoring in the 1–4 GHz Band with Cross-Database Validation and Machine-Learning Proof of Concept" (IEEE Access). For this study, full-text eligibility, evidence extraction and risk-of-bias ratings were performed by the second author (Ali Al-Dulaimi, British University Vietnam) with the first reviewer recused; methodological consistency of the appraisal was reviewed by the fifth author (Rami Qahwaji). The dielectric-property dataset itself is independently archived at Zenodo DOI 10.5281/zenodo.18027914 and is referenced from this deposit but is not redistributed here.

What this deposit does not contain

No new RF measurement datasets were generated for this review. The deposit does not redistribute raw RF measurements from included studies, source PDFs of those studies, or any private clinical data.

Reproducibility

After unzipping, run the following from the deposit root:

python 07_scripts/verify_prisma_flow.py
python 07_scripts/compute_headline_statistics.py
python 07_scripts/compute_claim_tier_divergence.py
python 07_scripts/compute_wilson_intervals.py

All scripts use only the Python standard library (Python ≥ 3.9