Comparison-level dataset and analysis code for a systematic review of deep learning versus embedding-based classification

# Comparison-level dataset and analysis code Supporting data and code for the systematic review "When Does Deep LearningOutperform Embedding-Based Classification? A Systematic Review and QuantitativeSynthesis of Data Requirements, Performance Trade-offs, and Paradigm Selection". The review compares end-to-end deep learning against embedding-basedclassification, in which a frozen pre-trained encoder classifies by cosinesimilarity to class prototypes. 27 primary studies published betweenJanuary 2018 and June 2026 met the inclusion criteria, contributing 113 pairedcross-paradigm comparisons. No new empirical data were generated. Every performance value was extracted fromthe published studies cited in the manuscript's reference list. The dataset isoriginal in that the comparison-level compilation, the paradigm classification ofeach arm, and the quality appraisal were produced by the authors. ## Files **Supplementary_Table_S1.xlsx** - Comparison-level dataset. One row per pairedcomparison (113 rows). Records the study identifier, dataset, task type, numberof classes, training set size and size band, the deep learning arm and theembedding arm, primary metric, both reported performance values, the reportedvariance and the basis on which it was computed where available, codeavailability, and the raw performance difference. Blank variance fields indicatethe source study reported a single value; those comparisons contribute to theraw-difference synthesis only. **Supplementary_Table_S2.xlsx** - Quality appraisal. One row per included study(27 rows), scored 0, 1 or 2 against five criteria: reproducibility, datasetstandardisation, evaluation rigor, baseline representativeness, and reportingcompleteness. Totals band as 8 to 10 High, 5 to 7 Moderate, 0 to 4 Low. **Supplementary_File_S3_analysis_script.py** - Analysis script. Reads theextraction workbook and computes the pooled standardized mean difference,subgroup analyses by domain, training-set-size band and model generation,publication bias tests, sensitivity analyses, and the figures. Requires Python3.9 or later with openpyxl, numpy, scipy and matplotlib. **Supplementary_File_S3b_computed_outputs.xlsx** - Computed output tables. Everyvalue quoted in Section 4.4, Table 2 and Section 4.7 of the manuscript,reproduced from the script above, including the pooled estimate(d = -0.194, I-squared = 99.4 percent, k = 48). **Supplementary_File_S4_PRISMA_checklist.xlsx** - Completed PRISMA 2020checklist, 42 items, each mapped to the section, table or figure of themanuscript where it is reported. ## Reproducing the analysis Place the extraction data and the script in the same directory and run thescript. It writes the results file, the fill map and the figures. Values in theoutput should match Supplementary_File_S3b exactly. ## Licence CC BY 4.0. Reuse is permitted with attribution.