Machine learning-based prediction of survival prognosis in cervical cancer

Dongyan Ding, Tingyuan Lang, Dongling Zou, Jiawei Tan, Jia Chen, Lei Zhou, Dong Wang, Rong Li, Yunzhe Li, Jingshu Liu, Cui Ma, Qi ZhouBMC Bioinformatics 2021 · DOI

TL;DR

This paper introduces a novel machine learning model (CCSPM) that uses microRNA (miRNA) expression data to predict cervical cancer patient survival with high accuracy. By stratifying patients into three distinct survival groups, this model offers a more nuanced prognostic tool than existing methods, potentially improving clinical management and treatment decisions.

What Problem Does This Paper Address?

Accurately predicting the prognosis of cervical cancer patients is crucial for effective management, yet current clinical features, including TNM staging, often lack sufficient discriminatory power due to the inherent heterogeneity of cancer cells. While molecular features hold promise, developing robust prediction models remains a challenge. This study aims to bridge this gap by leveraging machine learning and miRNA expression profiles to create a more precise survival prediction tool.

Methods

The researchers employed a multi-step machine learning approach:

  • Survival-Related miRNA Identification: They utilized the Cox Proportional-Hazards (Cox-PH) model to identify miRNAs significantly correlated with patient survival.
  • Patient Stratification: K-means clustering was applied to group patients based on the expression of these survival-related miRNAs. This step aimed to partition patients into distinct prognostic categories.
  • Prediction Model Development: A Support Vector Machine (SVM) algorithm was trained on the clustered patient data and miRNA expression profiles to build the final prediction model (CCSPM).
  • Model Evaluation: The model's performance was assessed using Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC) values, and Kaplan-Meier survival analysis.

Datasets

The study utilized the cervical cancer miRNA expression dataset from The Cancer Genome Atlas (TCGA) database. This dataset included miRNA expression data (RPKM) and associated clinical information for cervical cancer patients.

Key Insights and Ideas

  • miRNAs as Powerful Prognostic Markers: The study demonstrates that specific miRNA expression patterns are highly predictive of cervical cancer patient survival, outperforming traditional clinical features.
  • Hierarchical Clustering for Nuanced Prognosis: K-means clustering effectively stratified patients into groups with distinct survival outcomes, revealing that a single TNM stage can encompass patients with vastly different prognoses.
  • SVM for Robust Classification: The SVM algorithm proved adept at learning from the complex miRNA expression data to create a highly accurate predictive model.
  • Cancer Stem Cells (CSCs) as a Mechanistic Link: Pathway analysis revealed that the identified miRNAs target pathways involved in cancer stem cell maintenance, suggesting a biological mechanism underlying their prognostic value and linking them to cancer initiation, progression, and treatment resistance.

Key Results

  • 42 Survival-Related miRNAs Identified: Through Cox-PH analysis, 42 miRNAs were identified as significantly associated with cervical cancer survival.
  • Optimal Patient Stratification: K-means clustering, particularly with the top 10 miRNAs and K=4, successfully stratified patients into four groups with three distinct 5-year survival outcomes: >90% (high survival), ≈65% (moderate survival), and ≤40% (low survival).
  • High Predictive Accuracy: The developed SVM-based CCSPM achieved high performance metrics, including AUC values ranging from 0.914 to 0.996 across different patient groups and datasets (training, test, and combined). The misdiagnosis rate was also remarkably low (0.93% for the training set, 6.52% for the test set, and 2.61% for the whole set).
  • Improved Stratification over Existing Models: The CCSPM's ability to stratify patients into three distinct survival groups offers a significant improvement over existing two-group prognostic models, enhancing its clinical utility.

Limitations and Future Work

  • Interpretation of Clustering: While K=4 provided distinct survival outcomes, the authors note that interpreting the precise meaning of each cluster can be challenging and may benefit from medical expertise. They also observed that while groups 1 and 4 had different miRNA profiles, their 5-year survival outcomes were similar, suggesting potential overlap or complexity in the biological interpretation.
  • Data Heterogeneity: The study acknowledges the challenges posed by biological data, including missing values and batch effects, and describes a two-step imputation strategy to mitigate these issues. However, the inherent heterogeneity of tumor cells remains a factor.
  • Future Directions: The paper suggests that the developed CCSPM, by stratifying patients into more refined survival groups, can significantly improve the forecasting capacity for cervical cancer management. Further validation and clinical implementation of this model are implied next steps.

Summary updated

BibTeX
@article{ding2021machine,
  author = {Dongyan Ding and Tingyuan Lang and Dongling Zou and Jiawei Tan and Jia Chen and Lei Zhou and Dong Wang and Rong Li and Yunzhe Li and Jingshu Liu and Cui Ma and Qi Zhou},
  title = {{Machine learning-based prediction of survival prognosis in cervical cancer}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-04261-x},
}
Figure