Semi-Supervised Learning-Based Genetic Biomarkers Dataset for Multiple-Stage Hepatocellular Carcinoma Prediction
Ahmed Ammar Kubba, Manar Abu Talib, Jibran Sualeh Muhammad, Ali Bou Nassif, Abdalla Sayed Mohamed, Darko Castven, Jens U. Marquardt
cs.AI
Quick verdict
Proposes a 770-sample, 5-stage HCC gene-expression dataset built by label mapping and XGBoost self-training; 96.5% accuracy is inflated by circular evaluation.
Summary
This paper tackles the scarcity of public genomic datasets for multi-stage hepatocellular carcinoma (HCC) classification by building a new dataset from three sources: a private 29-sample Lubeck cohort carrying five pathology-defined labels (surrounding liver, early HCC, progressed HCC, low- and high-grade dysplastic nodules), TCGA (782 samples with coarse labels), and GEO series GSE89377 (107 samples, nine stage labels). Features are restricted to the 11,150 genes common to all three platforms, and GSE89377 labels are mapped onto the five target classes 'based on medical domain expert analysis'. The authors iteratively train XGBoost on the Lubeck set to predict TCGA labels, admitting TCGA samples only when predictions exceed a 40% confidence threshold and agree with the original coarse label; the process converged at 770 samples with severe class imbalance (test support: 99 progressed HCC vs. 8 surrounding-liver samples). On a stratified 70/30 split the final model reports 96.5% accuracy, 90.6% macro precision, 88.9% macro recall, 89.3% macro F1 and 0.1015 cross-entropy loss, with per-class F1 spanning 100% (progressed HCC) to 69.6% (high-grade dysplastic nodules, 11 samples). Key limitations: test samples were admitted only when model predictions matched the mapped labels, structurally inflating evaluation; no batch-effect correction is reported for merging RNA-seq and microarray platforms; no public release of data or code is stated; and there is no baseline comparison, cross-validation, or external validation.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A 770-sample, five-class, 11,150-feature HCC gene-expression dataset can be constructed by combining a private 29-sample Lubeck cohort with public TCGA and GSE89377 data via label mapping and iterative XGBoost self-training.
- An acceptance rule of >40% prediction confidence plus agreement with the original coarse label suffices to transfer the five-class label scheme onto public samples until performance converges.
- The final XGBoost model achieves 96.5% accuracy (macro precision 90.6%, macro recall 88.9%, macro F1 89.3%, cross-entropy loss 0.1015) on a stratified 70/30 split of the constructed dataset.
- The dataset is more detailed than other publicly available HCC genomic datasets because it covers five developmental stages including premalignant dysplastic nodules.
Strengths
Genuinely multi-stage label granularity
The five-class scheme (Table I) spans premalignant lesions (LGDN/HGDN), early and progressed HCC, and surrounding liver — finer than the binary tumor/normal framing of most public HCC resources, and directly useful for early-detection research.
Explicit, reproducible assembly logic on paper
Section II-C spells out the four acceptance conditions (40% confidence plus label-agreement rules per TCGA class) and a stated convergence criterion, and Section II-B documents the GSE89377 preprocessing (mapping to Illumina HumanHT-12 V3.0/GPL6947, intersecting to 11,150 common genes).
Honest per-class reporting
Cross-disciplinary authorship
Weaknesses
Circular, selection-biased evaluation
No batch-effect correction for a cross-platform merge
No baselines, cross-validation, or uncertainty
Deliverable not demonstrably available
Arbitrary threshold and contestable mappings
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a dataset-construction paper in computational biology. The authors claim three things: (1) that a five-class, multi-stage HCC gene-expression dataset — 770 samples, 11,150 features, spanning surrounding liver (sl), low/high-grade dysplastic nodules (lgdn/hgdn), early (ehcc) and progressed (phcc) HCC — can be assembled from a private 29-sample Lübeck cohort plus public TCGA (782 samples) and GEO GSE89377 (107 samples); (2) that an iterative XGBoost self-training loop with a 40% confidence threshold and label-agreement rules can transfer the five-class scheme onto the public samples; and (3) that the resulting model reaches 96.5% accuracy (Table II) on the constructed dataset. The abstract motivates the work by 'a distinct lack of publicly available HCC datasets utilizing genomic data', which makes public release of the deliverable central to the claim.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the HCC dataset presented in this paper?
It is a gene-expression dataset of 770 patient liver-tissue samples in five classes — surrounding liver, low-grade and high-grade dysplastic nodules, early hepatocellular carcinoma, and progressed hepatocellular carcinoma — with 11,150 gene expression levels per sample. It was built by combining a private 29-sample clinical cohort with the public TCGA (782 samples) and GEO GSE89377 (107 samples) datasets, using an iterative XGBoost loop to assign the five classes to public samples whose predictions passed a 40% confidence threshold and matched their original coarse labels.
How accurate is the XGBoost model in this study?
The paper reports 96.5% overall accuracy on a stratified 70/30 split of the constructed 770-sample dataset, with macro precision 90.6%, macro recall 88.9%, macro F1 89.3%, and cross-entropy loss 0.1015. However, the test pool partly consists of samples admitted only because the model already agreed with their mapped labels, so this figure likely overstates real-world performance; the rarest classes performed worst (high-grade dysplastic nodules F1 69.6% on 11 test samples).
Is the dataset publicly available?
The paper states no repository, accession number, or DOI for the 770-sample dataset or for the code, and its five-class ground truth derives from a private 29-sample clinical cohort from the Lübeck group. Public availability therefore cannot be confirmed from the paper itself.
Why do reviewers consider the reported accuracy optimistic?
During dataset construction, public samples were only added when XGBoost's predictions exceeded a 40% confidence threshold and matched the mapped coarse label. Because the final test split is drawn from this pre-filtered pool, the evaluation partly measures the acceptance filter itself rather than generalization to unseen patients — a structural selection/circularity bias that inflates accuracy. Additional concerns are the absence of batch-effect correction across merged sequencing and microarray platforms, and of any baseline, cross-validation, or external validation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-7 (pipeline, SSL illustration, per-class comparison, confusion matrix, class distribution, PCA) are referenced but not analyzable from the extracted text
- The per-source composition of the final 770 samples (how many accepted from TCGA vs. GSE89377 vs. Lubeck, per iteration) is not reported in the text
- TCGA's three-class scheme including a 'Transition' label is unusual and cannot be verified against the actual TCGA-LIHC composition from the text alone
- XGBoost hyperparameters and the number of self-training iterations are not stated
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.