Corpus 1,272 assessed · 1,173 scored · 643 reproduced ≥75 · 168 flagged ·∅ 74.1/100
← New search

A robust data scaling algorithm to improve classification accuracies in biomedical data.

BMC Bioinformatics · 2016
87/100 3/4
Why this verdict

The main results reproduced: recomputed values matched the published ones within tolerance.

Reproduced on the brainbox compute brainarbeit.com
How its reproducibility compares
87/100
Reproducibility score
0.7 SD above mean
vs. all fields · 1173 studies
🎯 Scores higher than 72% of all assessed papers rank 301 of 1173 scored

A 0–100 reproducibility-quality score from the per-question grades, shown as a z-score: standard deviations above (+) or below (−) the mean of comparable assessments.

Reproduction agent’s raw note

DESCRIBED WELL ENOUGH — yes. Fresh, independent re-run («infra» work dir had been janitor-reclaimed, so the repo was re-cloned and the whole pipeline re-computed from scratch on «our HPC», single «job» on n167, 12m24s, numpy 2.2.6/scipy 1.15.2/sklearn 1.7.2). The repo allenxhcao/glscale (commit 776a101) ships all 16 datasets as .mat (incl. the named accession GSE29490 = 48 samples x 26916 features) and the GL algorithm is fully specified (generalized-logistic CDF fit to each feature's empirical CDF); the authors' MATLAB needs the Statistics Toolbox, so a faithful Python port of the entire pipeline was run on the shipped data. PRIMARY OUTCOME (1:1): the proposed-method GL SVM AUROC values reproduce within tolerance across ALL 16 datasets — mean |delta| 0.0093, max 0.040, 15/16 within 0.03, 7/16 essentially exact (|d|<=0.003). The named accession GSE29490 reproduces tightly (GL 0.994 reported vs 0.996 reproduced; all four scaling methods within 0.006: None .983/.982, Minmax .984/.983, Zscore .985/.991, GL .994/.996) and GL is the best method there in our run too, so the per-dataset claim holds. These numbers match an earlier independent run («job», different node + library versions) to within rounding, so the GL result is not seed- or library-specific. HONEST DIVERGENCE: the paper's headline 'GL best in 27/29 AUROC, 25/29 ACC' only PARTIALLY reproduces — 14/29 (all methods) / 18/29 (scaled only) for AUROC, 10/29 for ACC. Root cause is the unscaled 'None' baseline: the paper's classifiers collapse on raw features (e.g. Lung SVM None 0.397, Colon 0.670) while sklearn's SVC/LogReg do not (0.925, 1.000), so an inflated None beats GL and GL 'wins' less often. This is a baseline implementation difference, not a data discrepancy and not a GL-value discrepancy; applying our GL-best counting rule to the paper's OWN printed numbers recovers 25/29 ACC (exact) and 26/29 AUROC, confirming our methodology matches the paper. NO FABRICATION indicated — GL numbers independently regenerated and matching in two separate runs. NOT ATTEMPTED: byte-exact 3rd-decimal match (different SVM/optimizer/CV-RNG); Table-3 accuracy operating point (under-specified threshold, advisory); the authors' original MATLAB run (toolbox unavailable). Verdict: the proposed method's reported values reproduce 1:1; its aggregate dominance over the (weak) unscaled baseline is a robustness caveat that does not fully reproduce in an independent pipeline.

💻 Code ↗ 🗄 Data: GSE29490

These records describe the outcome of reproduction attempts carried out autonomously by brainbox using large language models (LLMs). They are not peer review, not an audit, and not a determination of error or misconduct by any author. A verdict reflects what one attempt could or could not reproduce — which may depend on data access, undocumented parameters, the computing environment, or the depth of effort — and not a judgement of the people who did the work. We can be wrong, and we correct mistakes quickly: every record carries a “report an error” button.

Assessment versions

Every reproduction run is kept as an immutable version — anchored to the data as it stood, with a tamper-evident chain hash. A rerun (e.g. after an author updates a deposit) adds a new version; the previous one stays on record.

  1. v1 current initial assessment Score 87
    assessed: 2026-06-16 ⛓ fd3a2aab6ae4
✎ I am an author of this paper

Updated or fixed a deposit, or is there an erratum? Ask us to re-run the metrics. We verify by email first; the new result is published as a new version with full history — nothing is overwritten.

Reason for the rerun

We email you a confirmation link first. The rerun is an objective re-measurement — it cannot change the verdict in your favour, only ask us to look again.

Provenance — full disclosure

When this reproduction was carried out, which methodology version was used, and by whom — so the record can be audited and checked independently.

Reproduced
2026-06-22
Rubric version
not recorded
Assessed by
Last updated
2026-08-05

Provisional, curator- or AI-assessed, and independently checkable. A reproduction outcome states what one attempt could reproduce — not a judgement of the authors.

Deep full-text extraction

Model: sonnet
Founding hypothesis

The paper tests whether scaling biomedical data with a Generalized Logistic (GL) function fit to the empirical cumulative distribution function yields more accurate classification models than the commonly used Min-max and Z-score scaling algorithms, particularly due to robustness to outliers in small-sample biomedical settings.

Core claims
  • Models trained on data scaled by the GL algorithm outperform models trained on data scaled by the Min-max or Z-score algorithms across 16 binary classification tasks, measured by AUROC and percentage of correct classification finding
  • The GL algorithm scales data by learning a generalized logistic function that approximates the empirical cumulative distribution function (ECDF) of each variable method
  • The GL function maps values from the interval (-infinity, infinity) to (0,1), which reduces the distance of outliers from other samples and makes the algorithm intrinsically robust to outliers mechanism
  • The GL algorithm requires no additional denoising or outlier detection step in data preprocessing because it is inherently robust to outliers finding
  • GL function parameters (B, M, Q, ν) are fit by minimizing the sum of squared differences between the GL function and the ECDF via gradient descent, using a novel initialization that assumes the minimum, median, and maximum sample values map to approximately 0.1, 0.5, and 0.9 respectively method
  • The Min-max algorithm linearly maps training data to [-1,1] (or [0,1]) but produces out-of-bound scaled values for unseen/testing samples outside the training range and is very sensitive to outliers finding
  • The Z-score algorithm is sensitive to outliers, and with the small sample sizes typical in biomedical research the estimated mean and standard deviation poorly approximate the true population values, causing poor scaling of future inputs finding
  • The GL data-scaling approach is adapted from the histogram equalization technique used in digital image processing for contrast enhancement, but additionally learns a functional approximation of the CDF so it can scale unseen values mechanism
Experimental setups
Assay System Perturbation Readout Platform
binary classification modeling comparing data-scaling algorithms 16 biomedical datasets spanning different variable types and applications none AUROC and percentage of correct classification
classification with microarray gene expression data (preliminary study, cited ref. 9) gene expression data none classification accuracy
Key results
  • Models learned using data scaled by the GL algorithm outperform models using data scaled by the Min-max and Z-score algorithms
  • GL algorithm was effective in classifying tasks using microarray gene expression data in a preliminary study
Key statistics
  • count 16 (number of binary classification tasks used to evaluate GL vs. Min-max and Z-score scaling)
  • other L(x_min)≈0.1, L(x_med)≈0.5, L(x_max)≈0.9 (heuristic target values used to initialize GL function parameters before gradient descent fitting to the ECDF)

Statistical methods review

Model: opus

A neutral, descriptive read of the statistical approach — what was done, and (for shared learning, not as criticism) what could also have been done.

This is a methodology paper proposing a Generalized Logistic (GL) data scaling algorithm, evaluated empirically by comparing classification models trained on data scaled by GL versus the Min-max and Z-score algorithms. Performance was assessed across 16 binary classification tasks using area under the receiver operating characteristic curve (AUROC) and percentage of correct classification as the reported metrics. The text provided describes the algorithm and evaluation metrics but does not, in the portion available, state formal hypothesis tests, significance levels, or dispersion measures.

Replicationunclear Sample size16 binary classification tasks were used for evaluation; per-dataset sample sizes are not stated in the provided text GroupsGL scaling vs. Min-max vs. Z-score scaling, on classification performance Pairingunclear Randomization/blindingnot stated Dispersionunclear
Approaches that could also have been used
  • Methods were compared using AUROC and percentage of correct classification summarized across 16 tasks.
    Could also: A paired statistical test across datasets (e.g., Wilcoxon signed-rank test) or the Friedman test with a post-hoc Nemenyi comparison could also be applied to compare the three scaling methods across the 16 tasks. — Such tests would add a formal, distribution-free assessment of whether the observed performance differences across multiple datasets are consistent, which is a common convention when benchmarking algorithms over many datasets.
  • Classification performance appears to be reported as point estimates per method.
    Could also: Reporting variability via cross-validation fold standard deviations, 95% confidence intervals, or bootstrap intervals around AUROC could also accompany the point estimates. — Adding a measure of dispersion or interval would convey the stability of the performance estimates, which is especially informative when sample sizes are small as emphasized in this work.
  • The evaluation contrasts scaling methods on overall accuracy and AUROC.
    Could also: Complementary metrics such as the area under the precision-recall curve, F1, or calibration measures could also be reported, possibly with effect-size summaries of the differences. — These additional metrics would describe behavior under class imbalance and the magnitude of differences between methods, giving a fuller picture alongside AUROC.
  • The GL function parameters are fit by minimizing a sum-of-squared-differences objective via gradient descent from a heuristic initialization.
    Could also: Reporting fit quality with goodness-of-fit statistics, or comparing against alternative ECDF-approximation/optimization strategies (e.g., multiple random restarts) could also characterize the fitting step. — Quantifying fit and sensitivity to initialization would describe how reliably the non-convex optimization recovers a good approximation across variables and datasets.

Result convergence & founder nodes

Findings this paper shares with others that ran a comparable experiment. A node’s strength is how many independent papers report it (replication breadth) — not how often it is cited, so a heavily-replicated but under-cited founder still stands out.

Citation network

Where this publication sits in the reproducibility-weighted citation graph — what it is built on, and what is built on it. Citation data from OpenAlex.

Citations
231
Impact: very high
Foundation confidence
None of its references are in our reproducibility record yet — its foundation cannot be assessed.
Topics

No assessed neighbours yet — the network grows as more papers are assessed.

What was reproduced

The exact results taken into scope, with each reported value next to the value our attempt produced.

Scope — pmid-27612635 (GL robust data scaling)

Paper: Cao, Stojkovic, Obradovic. "A robust data scaling algorithm to improve classification accuracies in biomedical data." BMC Bioinformatics 2016. Code: github.com/allenxhcao/glscale (MATLAB + R port). Data: SHIPPED in repo as 16 .mat files under data/ (incl. GSE29490.mat — this RU's named accession).

Pipeline-derived results (IN SCOPE)

The entire paper is a computational pipeline: per dataset, scale features by {None, Minmax, Zscore, GL}, train {LogisticRegression, linear SVM}, evaluate by 5-fold CV (paper Methods), report AUROC (Table 2) and Accuracy (Table 3).

  • Proposed method GL (generalized-logistic CDF fit) is in glscale.m/glscaleR.R.
  • PRIMARY target (this RU): GSE29490 row — SVM only (LR=NA in paper). Table2 AUROC: None .983 Minmax .984 Zscore .985 GL .994 Table3 Acc : None .942 Minmax .954 Zscore .958 GL .979
  • SECONDARY (validate pipeline broadly): self-contained UCI sets Wdbc, Diabetes (Pima), Parkinsons, BreastTissue + the microarray/methylation sets.
  • QUALITATIVE headline claim: "GL best AUROC in 27/29 task-model cases, best accuracy in 25/29."

Out of scope / last-20% NOT attempted (with reason)

  • Byte-exact MATLAB numbers: fitcsvm vs sklearn SVM, fmincon+analytic-Hessian GL fit vs scipy least_squares, and crossvalind RNG fold assignment all differ → exact 3rd-decimal match not expected; grade on within-tol / direction.
  • Accuracy operating-point: script stores only AUROC; Table-3 accuracy threshold (max-F1 vs 0.5) is under-specified → AUROC is the clean grade, accuracy advisory.

Reproduction strategy

Faithful Python re-implementation of the described pipeline (port glscale exactly: logiFunc + SSE fit of generalized-logistic to empirical CDF), run on the SHIPPED .mat data on «our HPC». Authors' own MATLAB needs Statistics Toolbox (fitcsvm, perfcurve, crossvalind, fmincon) unavailable in Octave → re-implementation is the tractable, faithful path (paper describes the algorithm fully).

Figures / tables: Table
GSE29490_SVM_AUROC_GL
Reported
0.994
Reproduced
0.996
exact
GSE29490_SVM_AUROC_None
Reported
0.983
Reproduced
0.982
exact
GSE29490_SVM_AUROC_Minmax
Reported
0.984
Reproduced
0.983
exact
GSE29490_SVM_AUROC_Zscore
Reported
0.985
Reproduced
0.991
within tolerance
BreastTissue_SVM_AUROC_GL
Reported
0.972
Reproduced
0.972
exact
Wdbc_SVM_AUROC_GL
Reported
0.995
Reproduced
0.995
exact
Leukemia_SVM_AUROC_GL
Reported
1.000
Reproduced
0.998
exact
DLBCL_SVM_AUROC_GL
Reported
0.654
Reproduced
0.653
exact
ProstateCancer_SVM_AUROC_GL
Reported
0.822
Reproduced
0.821
exact
GSE25869_SVM_AUROC_GL
Reported
0.943
Reproduced
0.947
exact
Diabetes_SVM_AUROC_GL
Reported
0.834
Reproduced
0.837
exact
ColonCancer_SVM_AUROC_GL
Reported
0.981
Reproduced
0.970
within tolerance
BreastCancer_SVM_AUROC_GL
Reported
0.812
Reproduced
0.824
within tolerance
Parkinsons_SVM_AUROC_GL
Reported
0.891
Reproduced
0.885
within tolerance
Myeloma_SVM_AUROC_GL
Reported
0.750
Reproduced
0.739
within tolerance
IndianLiver_SVM_AUROC_GL
Reported
0.695
Reproduced
0.712
within tolerance
GSE27899IL_SVM_AUROC_GL
Reported
0.824
Reproduced
0.845
within tolerance
LSVT_SVM_AUROC_GL
Reported
0.919
Reproduced
0.901
within tolerance
LungCancer_SVM_AUROC_GL
Reported
0.774
Reproduced
0.814
partial
HEADLINE_GL_best_AUROC
Reported
GL best in 27/29 cases
Reproduced
14/29 among all 4 methods; 18/29 among scaled methods
partial
HEADLINE_GL_best_ACC
Reported
GL best in 25/29 cases
Reproduced
10/29 among all 4 methods
partial

Assessments & scoring basis

Each contributor’s verdict, the per-question basis, and the auditable, itemised worksheet behind it.

No assessment has been recorded yet.
🤝
Reproduced automatically — and fairly

Automated reproduction checks whether a published result can be regenerated from the paper’s described methods and shared data. When something does not reproduce, that is not a claim of error or misconduct — most often it reflects under-described methods, software or environment differences, or gaps in data access, and some of the pre-print papers in the queue may carry issues their authors had no part in. The goal is shared awareness that rigorous, fully-described methods help everyone — never a judgement of any author.

Are you an author? We would genuinely like to hear from you — to clarify the record, add data or code, re-run the pipeline after an accession update, and publish your response right next to the assessment. Everything here is open and auditable.

🚩 Report an error in this record

Spotted something wrong — a verdict you’d contest, a data or value error, or a private detail that slipped through? Tell us, with a short justification. Authors and readers are equally welcome to write in; we review every report.

Prefer email, or the form below not working? Contact us at support@doesitreproduce.com.

Reproduction footprint

claude-opus-4-8

Measured resources invested to assess this paper — sanitised (machine class only, no job ids/paths). Compute = HPC accounting (SLURM); tokens = the AI agent's session.

405.3 k
tokens (I/O) · 36.7 M incl. cache
154 min
runtime · 8.94 CPU-h
0.8 GB
peak RAM
1
HPC jobs
hummel
machine