Reliable classification of imbalanced lung cancer data with enhanced split conformal prediction
PeerJ Computer Science · 04.06.2026 · 12 · e3841
DOI: 10.7717/peerj-cs.3841 ↗Kurz gefasst
Fachartikel in der Zeitschrift PeerJ Computer Science, online erschienen am 04.06.2026. Verfasst von Hao Ding und 6 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Natural Science Foundation of China, Precision Radiotherapy Science and Technology Innovation Team of Yunnan Universities.
Abstract (Originalsprache)
Background
Lung cancer is one of the most common and deadliest malignancies worldwide, and early accurate diagnosis is crucial for improving prognosis. Machine learning models based on gene expression profiles offer new approaches for lung cancer classification. However, the inherent high dimensionality of such data and severe class imbalance issues lead to limitations in traditional classification models, such as low recognition rates of minority classes, lack of statistical reliability assessment of predictions, and increased risk of clinical misdiagnosis.
Methods
In this study, we introduce an Enhanced Split Conformal Prediction (ESCP) framework to increase classification accuracy and improve predictive reliability for lung cancer gene expression data. First, applying feature selection using Sure Independence Screening (SIS) to identify a subset of features that are strongly associated with lung cancer status; then, alleviating class imbalance by applying the Synthetic Minority Over-sampling Technique (SMOTE) in the feature space; at last, training a base classifier on the balanced dataset and integrating the Split Conformal Prediction (SCP) method. The SCP module calculates nonconformity scores to generate prediction sets for each sample with statistical guarantees of coverage.
Results
A systematic evaluation across eight public lung cancer gene expression datasets (IM101–IM108) demonstrates that the ESCP framework significantly outperforms baseline models. In terms of classification performance, ESCP achieves higher areas under the receiver operating characteristic curve (AUROC) and precision-recall curve (AUPRC) with most datasets. Regarding predictive reliability, the SCP module ensures that empirical coverage rates closely approximated the pre-specified 95% confidence level ( e.g ., RF-ESCP achieved 96.6% coverage for IM106), with prediction interval lengths (ranging from 0.87 to 1.86) reasonably reflecting model uncertainty. Particularly in ultra-high dimensional datasets, ESCP maintains over 93.6% coverage, effectively overcoming the overfitting issues observed in traditional methods such as single support vector machines, which reportedly have perfect accuracy (accuracy = 1.0) but lack confidence assessment.
Conclusions
This study introduces the split conformal prediction framework into high-dimensional, imbalanced lung cancer gene expression classification tasks. Evaluated on eight public datasets, the ESCP algorithm not only enhances the recognition of minority classes but also provides statistically guaranteed confidence levels for each prediction. This approach reduces model uncertainty and shows potential to offer more interpretable and reliable decision support for clinical applications.
Förderung
- National Natural Science Foundation of China
- Precision Radiotherapy Science and Technology Innovation Team of Yunnan Universities
Zitieren
APA
Ding, H., Fu, G., Fu, Y., Wang, G., Liu, Y., Xiong, W., & Li, R. (2026). Reliable classification of imbalanced lung cancer data with enhanced split conformal prediction. PeerJ Computer Science, 12, e3841. https://doi.org/10.7717/peerj-cs.3841
Vancouver
Ding H, Fu G, Fu Y, Wang G, Liu Y, Xiong W, et al. Reliable classification of imbalanced lung cancer data with enhanced split conformal prediction. PeerJ Computer Science. 2026;12:e3841. doi:10.7717/peerj-cs.3841
BibTeX
@article{Ding2026,
author = {Ding, Hao and Fu, Guanghui and Fu, Yingzi and Wang, Guangxian and Liu, Yongqi and Xiong, Wen and Li, Rongqing},
title = {Reliable classification of imbalanced lung cancer data with enhanced split conformal prediction},
journal = {PeerJ Computer Science},
year = {2026},
volume = {12},
pages = {e3841},
doi = {10.7717/peerj-cs.3841}
}
EUVIMED