A Review on Biomarker‐Enhanced Machine Learning for Early Diagnosis and Outcome Prediction in Ovarian Cancer Management

Somayyeh Hormaty, Anwar Nather Seiwan, Bader Rasheed, Hanieh Parvaz, Ali Gharahzadeh, Hamid GhaznaviCancer 2025 · DOI

TL;DR

Machine learning (ML) models, when combined with biomarkers, show significant promise in improving the early detection, diagnosis, and prognosis of ovarian cancer (OC), outperforming traditional methods. Future research needs to focus on multi-center validation, integrating diverse data types (like genomics and imaging), and developing explainable AI to bring these powerful tools into clinical practice and ultimately reduce OC mortality.

What Problem Does This Paper Address?

Ovarian cancer (OC) is a deadly disease, largely because it's often diagnosed at late stages when treatment is less effective. Current diagnostic methods, relying on clinical assessment, imaging, and biomarkers like CA-125, have limitations in terms of sensitivity and specificity, especially for early-stage detection. This leads to poor prognoses, with a 5-year survival rate of only about 30% for advanced stages. The paper addresses the critical need for more accurate and earlier detection and better prediction of patient outcomes to improve management and survival rates.

Methods

This review synthesizes findings from 17 studies that applied machine learning (ML) techniques to biomarker data for OC prediction. The authors analyzed various ML algorithms, including ensemble methods (Random Forest, XGBoost) and deep learning approaches (Neural Networks, RNNs), and examined their performance in diagnostic classification, survival prediction, and treatment response forecasting. The review emphasizes the integration of multi-modal data, such as tumor markers, inflammatory, metabolic, and hematologic parameters, to enhance ML model performance.

Datasets

The paper does not introduce new datasets but reviews studies that utilized existing clinical and biomarker data. The datasets varied in size, with some studies including as few as 44 patients and others involving thousands. The data types included cross-sectional measurements (single-time point) and longitudinal data (serial measurements), which were found to be particularly beneficial for certain ML models like RNNs.

Key Insights and Ideas

  • Biomarker Synergy is Key: Combining traditional biomarkers like CA-125 and HE4 with other markers (inflammatory, metabolic, hematologic) significantly boosts the predictive power of ML models, far beyond what individual markers can achieve.
  • Ensemble and Deep Learning Dominate: Algorithms like Random Forest, XGBoost, and various neural network architectures consistently outperform traditional statistical methods, achieving high accuracy in classification and prediction tasks.
  • Longitudinal Data Offers an Edge: For certain ML models, especially RNNs, using serial biomarker measurements (longitudinal data) provides a richer temporal context that can lead to more accurate early detection compared to single-time point data.
  • Explainable AI is Crucial for Trust: While ML models can achieve high accuracy, understanding why they make certain predictions (using techniques like SHAP) is vital for clinical adoption and building trust among healthcare professionals.
  • Multi-omics Integration is the Future: Moving beyond single data types, integrating genomics, proteomics, and radiomics with clinical and biomarker data holds immense potential for truly personalized OC prediction and management.

Key Results

  • Biomarker-driven ML models achieved AUC values exceeding 0.90 in diagnosing OC and distinguishing malignant from benign tumors.
  • Ensemble methods and deep learning approaches reached classification accuracy up to 99.82% and survival prediction AUCs up to 0.866.
  • Combining CA-125 and HE4 with other markers like CRP and NLR enhanced specificity and sensitivity.
  • Models integrating multiple biomarker categories (tumor, inflammatory, metabolic, hematologic) significantly improved classification accuracy.

Limitations and Future Work

The authors highlight several limitations:

  • Small and Single-Center Datasets: Many studies suffer from small sample sizes and lack of external validation, leading to overfitting and limiting generalizability.
  • Exclusion of Key Data: Many studies exclude valuable data like imaging (radiomics) and genomic information, which could further enhance model accuracy.
  • Challenges in Early/Borderline Cases: Differentiating benign from malignant tumors, especially in early stages or with borderline cases, remains difficult due to biomarker limitations and subjective interpretations.
  • Short Follow-up Periods: Limited follow-up in some studies hinders the reliability of survival and treatment response predictions.

Future work should prioritize:

  • Multi-center validation: To ensure models perform well across diverse populations.
  • Multi-omics integration: Combining genomics, proteomics, and radiomics with biomarker data for a more comprehensive view.
  • Explainable AI (XAI): To increase transparency and trust in ML models.
  • Data augmentation: To mitigate the impact of small sample sizes.
  • Development of robust validation strategies: To ensure models are reliable and clinically applicable.

Summary updated

Abstract

Biomarker-driven ML represents a transformative approach for OC management, improving diagnostic precision and personalized care. Future research should prioritize multi-center validation, multi-omics integration, and explainable AI to overcome current challenges and enable real-world implementation, potentially reducing OC mortality through earlier detection and optimized treatment.
BibTeX
@inproceedings{hormaty2025review,
  author = {Somayyeh Hormaty and Anwar Nather Seiwan and Bader Rasheed and Hanieh Parvaz and Ali Gharahzadeh and Hamid Ghaznavi},
  title = {{A Review on Biomarker‐Enhanced Machine Learning for Early Diagnosis and Outcome Prediction in Ovarian Cancer Management}},
  booktitle = {Cancer},
  year = {2025},
  doi = {10.1002/cam4.71224},
}
Figure