PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.17100v1cs.AI

TL;DR

Limited and small HCC datasets constrain work on genomic-data classification. This paper constructs a multi-stage dataset by combining three genomic datasets through semi-supervised learning; it contains 770 samples and the XGBoost model achieved 96.5% accuracy.

  • Problem

    The study addresses the limited availability of HCC datasets and small sample sizes for multi-stage HCC classification.

    Source: §I · p. 2

  • Method

    The authors use semi-supervised learning with three genomic datasets, mapping labels and retaining shared gene-expression features to construct an HCC dataset.

    Source: §I · p. 2, B. Dataset Pre-Processing · p. 3, C. Semi-Supervised Learning · p. 3

  • Results

    770 samples with 11,150 gene-expression features each form the five-class dataset, and the XGBoost model achieved 96.5% accuracy.

    Source: B. HCC Dataset · p. 4, A. XGBoost Model · p. 4

  • Takeaways & Limitations

    The constructed dataset can support medical research, early-diagnosis work, and AI-driven classification and predictive models for HCC.

    Source: §IV · p. 5

    The authors identify class imbalance as a limitation of the dataset.

    Source: §IV · p. 5

Abstract

from arXiv · show

Liver cancer is a complex disease responsible for a high number of deaths across the globe each year, making automated solutions for liver cancer classification urgent. The most common form of liver cancer is hepatocellular carcinoma (HCC), accounting for over 90% of liver cancer cases. There is a distinct lack of publicly available HCC datasets utilizing genomic data, which is necessary for training artificial intelligence (AI) models for automated HCC classification. This study proposes constructing a multi-stage HCC dataset using XGBoost and Semi-Supervised learning on three separate datasets of genomic biomarkers, utilizing their existing labels in the Semi-Supervised learning process to label the proposed dataset. The proposed dataset consists of 770 patient samples in total, categorized into five classes that represent normal tissue alongside different stages of HCC. Each sample in the dataset consists of 11,150 different gene expression levels. The XGBoost model demonstrated a final classification accuracy of 96.5% during the Semi-Supervised learning process.

I. INTRODUCTION AND RELATED WORK

The section reviews HCC biomarker and gene-classification research alongside limitations in available data, then introduces a five-class, 770-sample dataset with 11,150 genomic expressions per sample.

I. INTRODUCTION AND RELATED WORK

  • Prior biomarker research identified 28 metabolites and 169 genes correlated with aggressive HCC progression and outcomes.The section also describes body-fluid monitoring, assay technologies, and multi-omics approaches.Source: §I · p. 1
  • Early HCC is typically smaller and less symptomatic, while advanced HCC can involve larger tumors and vascular or organ invasion.LGDN and HGDN are precancerous lesions; HGDN has more pronounced atypia and a higher risk of progression to HCC.Source: §I · p. 1
  • Das et al.'s gene-sequence classifier used only four healthy and four HCC genes, and the authors noted insufficient data for training a new CNN.Reported accuracies were 80.36% for the one-dimensional CNN, 98.86% for SVM, and 100% for fine-tuned VGG16.Source: §I · pp. 1–2
  • 770 samples with 11,150 genomic expressions each form a five-class dataset intended to address limited, small HCC datasets.The classes represent tissue samples from different HCC stages.Source: §I · p. 2

II. METHODOLOGY

The methodology combines three genomic datasets through collection, preprocessing, and semi-supervised learning to construct the HCC dataset.

II. METHODOLOGY

  • The workflow obtains a base dataset from medical domain experts and uses three different datasets to build the final HCC dataset.Figure 1 outlines the collection, preprocessing and formatting, and semi-supervised learning phases.Source: §II · p. 2
  • The collected data are preprocessed and properly formatted before the semi-supervised learning phase constructs the HCC dataset.Source: §II · p. 2
  • Semi-supervised learning is the final stated phase in the dataset-construction workflow.Source: §II · p. 2

A. Lubeck Dataset and Target Classes

The Lubeck dataset supplies five ground-truth tissue classes and serves as the starting point for semi-supervised learning.

A. Lubeck Dataset and Target Classes

  • The paper's methodology figure provides a visual summary of the steps used to construct the HCC dataset.Source: Figure 1 · p. 2
  • The five classes are Surrounding Liver (sl), early HCC (ehcc), progressed HCC (phcc), low-grade dysplastic nodules (lgdn), and high-grade dysplastic nodules (hgdn).LGDN is associated with lower HCC risk, while HGDN is associated with higher risk.Source: A. Lubeck Dataset and Target Classes · p. 2, Table I · p. 2
  • Lubeck's five labels are described as the most detailed HCC-stage labels among the three datasets and provide the semi-supervised learning starting point.Source: A. Lubeck Dataset and Target Classes · p. 2

B. Dataset Pre-Processing

Three sources provide labeled or stage-specific genomic data; preprocessing aligns shared gene features to enable their combination.

B. Dataset Pre-Processing

  • The Lubeck private dataset contains 29 samples, each with 16,381 gene-expression features and one of five class labels.Its labels are sl, ehcc, phcc, lgdn, and hgdn.Source: B. Dataset Pre-Processing · p. 3
  • The Cancer Genome Atlas (TCGA) source contains 782 HCC patient samples with 16,382 gene-expression features and Normal, HCC, and Transition labels.Source: B. Dataset Pre-Processing · p. 3
  • GSE89377 contains 107 liver-tissue samples across nine HCC-development stages, with around 48,000 gene-expression features per sample.Source: B. Dataset Pre-Processing · p. 3
  • For GSE89377, gene keywords were matched to symbols on the Illumina HumanHT-12 V3.0 Expression BeadChip before combining data.Source: B. Dataset Pre-Processing · p. 3
  • Only gene-expression features common to all three datasets were retained so samples from each source could be combined.Source: B. Dataset Pre-Processing · p. 3

C. Semi-Supervised Learning

The semi-supervised procedure uses Lubeck’s five labels to expand the dataset with samples from TCGA and GSE89377, whose labels differ from the target classes. XGBoost predictions are added under confidence and label-mapping rules, with retraining repeated until performance stops improving.

C. Semi-Supervised Learning

  • GSE89377 samples are mapped into the existing datasets’ labels, since GSE89377 and TCGA use labels different from Lubeck’s five target classes.Cirrhosis samples are classified as Normal; chronic hepatitis samples are discarded as unrelated to HCC.Source: C. Semi-Supervised Learning · p. 3
  • The process trains XGBoost on Lubeck, predicts TCGA labels, adds high-confidence predictions matching TCGA’s real labels, then retrains on the expanded Lubeck dataset.The train/test split is stratified 70%/30%.Source: C. Semi-Supervised Learning · p. 3
  • A 40% prediction-confidence threshold and medical-expert label rules govern which TCGA samples are eligible for addition.HCC predictions must be ehcc or phcc, Transition predictions lgdn or hgdn, and Normal predictions sl.Source: C. Semi-Supervised Learning · p. 3
  • The expansion and retraining steps repeat until performance degrades or no longer changes.These conditions define convergence.Source: C. Semi-Supervised Learning · p. 3

III. RESULTS AND DISCUSSION

The results section covers the final XGBoost model and its performance, as well as the constructed dataset’s class distribution and genomic features.

III. RESULTS AND DISCUSSION

  • The results section presents the final XGBoost model and discusses its performance, the dataset’s class distribution, and its genomic features.Source: §III · p. 3
  • Figure 3 summarizes the semi-supervised training process that uses XGBoost to create the HCC dataset.Source: Figure 3 · p. 4

A. XGBoost Model

The final XGBoost model, trained using all gene features, reports 96.5% accuracy on the HCC dataset, alongside precision, recall, F1, and crossentropy metrics.

A. XGBoost Model

  • 96.5% accuracy, 90.6% precision, 88.9% recall, and 89.3% F1 score were reported for XGBoost on the HCC dataset.The accompanying Table II records accuracy, precision, recall, and F1 score.Source: A. XGBoost Model · p. 4, Table II · p. 4
  • 0.1015 crossentropy loss accompanies the reported metrics, while the authors note a precision-recall trade-off that leans toward reducing false positives.The authors caution that class imbalance should be considered when comparing model performance; XGBoost was mainly used to construct the dataset during semi-supervised learning.Source: A. XGBoost Model · p. 4
  • Figures 4 and 5 show performance across the five classes and the model’s misclassifications, respectively.Source: Figure 4 · p. 4, Figure 5 · p. 4, A. XGBoost Model · p. 4

B. HCC Dataset

The constructed HCC dataset contains 770 patient samples with 11,150 gene-expression features per patient, across five classes whose counts are imbalanced.

B. HCC Dataset

  • 770 patient samples, each with 11,150 gene-expression features, make up the constructed dataset.The features represent genomic expression levels.Source: B. HCC Dataset · p. 4
  • The 770-sample HCC dataset is significantly imbalanced: phcc and ehcc have the highest sample counts, while lgdn, hgdn, and sl are underrepresented; Figure 6 displays class frequencies.Source: B. HCC Dataset · p. 4, Figure 6 · p. 5
  • Figure 7’s PCA plot visualizes how the five classes are distributed in feature space.Source: B. HCC Dataset · p. 5, Figure 7 · p. 5

IV. CONCLUSION

The paper presents a relatively large, high-dimensional HCC dataset built from one private and two public datasets, mapping the public datasets’ labels to the Lubeck dataset. It identifies class imbalance and further validation, calibration, and sample-diversity work as priorities.

  • The conclusion presents a relatively large, high-dimensional HCC dataset built from one private and two public datasets, mapping the public datasets’ labels to the Lubeck dataset.Source: §IV · p. 5
  • Dataset imbalance is identified as a limitation, with class weighting or synthetic minority oversampling proposed to mitigate it and improve generalization.Source: §IV · p. 5
  • Further external-dataset validation and probability-calibration checks are proposed to assess robustness and generalizability.Source: §IV · p. 5
  • Interdisciplinary medical and machine-learning collaboration is proposed to improve dataset quality and sample diversity and identify genes for clinical validation.Source: §IV · p. 5
Loading 2609.17100v1…