• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Thursday, November 19, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Cancer

Stacking Machine Learning Ensemble Predicts Ovarian Cancer Risk in Screening Cohort

by
November 10, 2026
in Cancer, Health
Reading Time: 6 mins read
0
Stacking Machine Learning Ensemble Predicts Ovarian Cancer Risk in Screening Cohort

Stacking Machine Learning Ensemble Predicts Ovarian Cancer Risk in Screening Cohort

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Researchers have developed an interpretable, leakage-aware machine learning framework to evaluate ovarian cancer risk using data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial, showing that an ensemble combining K-Nearest Neighbors and AdaBoost achieved an accuracy of 0.7644 and a test-set receiver operating characteristic area under the curve of 0.8399. The study, published in Frontiers in Oncology by Sheela Lavanya J. M. and Subbulakshmi P. from the Vellore Institute of Technology, examined risk stratification across demographic, clinical, questionnaire, and biomarker variables while eliminating retrospective outcome variables that artificially inflate predictive performance. Across fifteen stacking configurations and nine standalone algorithms, combining distance-based pattern recognition and gradient-weighted boosting delivered the highest observed predictive balance on an analytically balanced testing cohort of 208 participants. However, hypothesis testing revealed that the ensemble’s accuracy was not statistically significantly better than standalone AdaBoost, establishing the stacking framework as a comparable alternative rather than a superior architecture.

The clinical rationale underpinning the framework stems from the difficulty of detecting ovarian malignancies before metastasis occurs, as early-stage cases remain largely asymptomatic and standard screening parameters lack adequate sensitivity. In gynecological oncology, ovarian cancer accounts for an estimated 300,000 new diagnoses and over 200,000 deaths annually, ranking as the eighth leading cause of cancer mortality among women worldwide. Survival trajectories depend acutely on clinical stage at diagnosis: patients diagnosed at localized stages achieve five-year survival rates exceeding 90 percent, whereas individuals diagnosed with advanced disease face survival below 30 percent. Routine screening modalities relying on serum cancer antigen 125 blood tests and transvaginal ultrasound examinations lack sufficient sensitivity and specificity, leaving nearly 70 percent of cases diagnosed at advanced stages. Randomized screening trials, including the United Kingdom Collaborative Trial of Ovarian Cancer Screening and the trial analyzed here, showed that universal screening with these tools produced no statistically significant mortality reduction, highlighting the necessity of multidimensional algorithmic models.

To construct and validate predictive models within an extensive population, the authors examined data from the randomized screening trial conducted between 1993 and 2001 by the National Cancer Institute. The full study enrolled approximately 155,000 individuals, including an ovarian screening arm of 78,209 women aged 55 to 74 receiving either annual multimodal screening or usual medical care. Across this cohort, 518 participants were diagnosed with ovarian cancer alongside 77,691 unaffected controls, creating severe class imbalance where malignancies constituted only 0.66 percent of records. Because machine learning classifiers trained on imbalanced data tend to overlook minority classes, the investigators applied random undersampling without replacement to the control population, matching 518 randomly chosen controls to the 518 cancer cases. This created a balanced analytical sample of 1,036 participants, split using stratified sampling into an 80 percent training set of 828 participants and an independent 20 percent test set of 208 participants, maintaining equal proportions of cases and controls.

A central contribution of the study involved identifying and removing data leakage variables prior to feature selection, model training, and explainability assessments. Data leakage occurs when information unavailable at the prediction time point enters model training, causing algorithms to learn from downstream outcomes rather than pre-diagnostic signals and producing overoptimistic performance estimates. Preliminary evaluations identified eight variables that demonstrated unusually high predictive power and violated temporal causality. Two variables—binary mortality status and death recorded with cause-of-death documentation—were classified as direct leakage because they explicitly encoded study outcomes. Six additional variables—days to study exit or mortality, exit mortality status, age at exit or censoring, age at first cancer diagnosis, age at ovarian cancer exit, and ovarian cancer survival indicators—represented indirect leakage capturing post-diagnostic endpoints. The authors excluded all eight variables, restricting all subsequent preprocessing, feature selection, and model training strictly to pre-diagnostic demographic measures, baseline questionnaires, and screening records obtained before cancer ascertainment.

To reduce high-dimensional clinical data into stable, non-redundant predictors, the researchers constructed a multi-method feature selection pipeline combining statistical filters and tree-based embedded algorithms applied exclusively to the training cohort. Clinical screening datasets feature numerous interrelated variables, requiring consensus strategies to prevent overfitting and improve interpretability. The pipeline utilized analysis of variance F-tests to assess numerical distinctions across outcome groups and Chi-square tests for categorical associations, paired with embedded importance metrics from Random Forest Gini impurity reductions and XGBoost gain calculations. Rather than depending on a single technique, the authors calculated an aggregate sum of ranks across all four methods, prioritizing features with the lowest cumulative rank sums. This consensus approach selected twenty top-ranked pre-diagnostic predictors, capturing participant-reported personal histories of ovarian and other cancers, questionnaire entry ages, screening timeline intervals, and serial CA-125 and ultrasound measurements, yielding compact training and testing matrices of 828 by 20 and 208 by 20.

Algorithmic benchmarking began with baseline evaluations across the full feature repository, revealing distinct classification patterns across linear models, margin-based classifiers, distance metrics, and boosting ensembles. Prior to feature selection, boosting algorithms achieved the highest test accuracies, led by AdaBoost at 0.7644, CatBoost at 0.7500, and LightGBM at 0.7356, while Random Forest attained 0.7163. Linear Support Vector Machines and Logistic Regression achieved moderate accuracies of 0.6875 and 0.6635, whereas K-Nearest Neighbors recorded 0.5385. Retraining on the selected twenty-feature subset showed that boosting architectures remained the most robust classifiers. AdaBoost delivered the highest individual accuracy at 0.7452, accompanied by a precision of 0.7463 and an F1-score of 0.7449. XGBoost achieved an accuracy of 0.7404, CatBoost reached 0.7356, and LightGBM achieved 0.7308. Among remaining models, Random Forest registered 0.7260, Linear Support Vector Machines and Decision Trees both reached 0.7115, Logistic Regression scored 0.6923, and K-Nearest Neighbors improved to 0.6058.

Threshold-independent receiver operating characteristic evaluations further demonstrated the generalization capacity of boosting architectures on the twenty-feature subset. While training-set areas under the curve reached 0.9721 for CatBoost, 0.9630 for LightGBM, and 0.9548 for XGBoost, test-set discrimination remained closely grouped among ensemble models. On independent test data, AdaBoost achieved the highest area under the curve at 0.8401, followed by XGBoost at 0.8364, Random Forest at 0.8259, CatBoost at 0.8246, Decision Trees at 0.8207, and LightGBM at 0.8149, while Logistic Regression recorded 0.7224 and K-Nearest Neighbors scored 0.6124. To enhance performance, the researchers evaluated fifteen stacking ensemble configurations using Logistic Regression as a common meta-learner trained on out-of-fold probability predictions via five-fold cross-validation. The stacking ensemble combining K-Nearest Neighbors and AdaBoost achieved the highest observed test accuracy of 0.7644, with an error rate of 0.2356, a precision of 0.7700, and an F1-score of 0.7632, restoring overall accuracy over the reduced feature set.

Diagnostic metrics evaluated on the test cohort provided clinical context beyond overall accuracy, assessing the top-performing stacking model across confusion matrix derivatives and hypothesis testing. The K-Nearest Neighbors and AdaBoost ensemble demonstrated a sensitivity of 0.6923, detecting 72 of 104 cancer cases, and a specificity of 0.8365, identifying 87 of 104 controls. The model yielded a positive predictive value of 0.8089, a negative predictive value of 0.7310, a balanced accuracy of 0.7644, a Matthews correlation coefficient of 0.5344, and a precision-recall area under the curve of 0.8767. Class-wise breakdowns showed an F1-score of 0.7802 for controls and 0.7461 for cancer cases. Ninety-five percent bootstrap confidence intervals for accuracy ranged from 0.7019 to 0.8174. Crucially, a paired McNemar test comparing the stacking model against standalone AdaBoost produced a statistic of 0.45 and a p-value of 0.5023, indicating no statistically significant difference and confirming the stacking model functions as a comparable rather than superior predictive tool.

To ensure clinical interpretability and foster transparency, the authors implemented post hoc explainable artificial intelligence techniques using SHapley Additive exPlanations and permutation feature importance across the complete ensemble. Evaluating the stacking model across test observations revealed that pre-diagnostic questionnaire responses, age indicators, and serial screening biomarkers exerted the strongest influence on predicted risk. Key contributors identified by SHAP included participant-reported personal history of ovarian cancer from the supplemental questionnaire, questionnaire entry age, fifth-round CA-125 results, screening timeline intervals, chronological age, and baseline ultrasound findings. Permutation importance corroborated this hierarchy: supplemental questionnaire personal history produced the highest importance score at 0.1207, followed by entry age at 0.0231, screening timeline days at 0.0197, dietary survey timeline days at 0.0192, and dietary questionnaire personal history at 0.0139. Serial CA-125 measurements, chronological age, and ultrasound results also maintained positive contributions. Built-in feature importance similarly scored the leading personal history variable at 0.3202, confirming alignment across independent explainability methods.

In evaluating probabilistic calibration and clinical utility, the authors examined model reliability while detailing explicit limitations that necessitate external validation prior to clinical translation. The K-Nearest Neighbors and AdaBoost ensemble achieved a Brier score of 0.1686 and a calibration intercept of 0.0021, reflecting good probabilistic accuracy and minimal systematic bias. However, an observed calibration slope of 2.4486 departed from the ideal value of 1.0, showing that predicted risks were lower than observed event rates and establishing that formal recalibration is needed before probabilities can serve as absolute risk estimates. Decision curve analysis indicated potential net clinical benefit across threshold probabilities, though the authors noted these findings are exploratory because the model utilized an undersampled cohort. Additionally, evaluation was restricted to a single retrospective cohort without external, temporal, or prospective validation. The researchers concluded that external validation in population-representative cohorts, probability recalibration, and prospective assessments remain essential prerequisites before such machine learning frameworks can support clinical ovarian cancer screening.

Subject of Research: Machine learning risk prediction for ovarian cancer using screening trial data

Article Title: A leakage-aware and interpretable machine learning framework for ovarian cancer prediction using the PLCO screening dataset

Article References: M., S. L. J., & P., S. (2026). A leakage-aware and interpretable machine learning framework for ovarian cancer prediction using the PLCO screening dataset. Frontiers in Oncology, 16, Article 1926361. https://doi.org/10.3389/fonc.2026.1926361

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1926361

Keywords: ovarian cancer, machine learning, cancer screening, stacking ensemble, PLCO trial, data leakage, explainable AI, risk stratification, leakage-aware, interpretable, machine, learning

News Source: Nathaniel Bowman. (November 10, 2026). Stacking Machine Learning Ensemble Predicts Ovarian Cancer Risk in Screening Cohort. Scienmag.

Tags: cancer screeningdata leakageExplainable AIinterpretableleakage-awarelearningmachineMachine LearningOvarian cancerPLCO trialrisk stratificationstacking ensemble
Share12Tweet7Share2ShareShareShare1

Related Posts

Researchers Design Quality Indicator System for Integrated Chinese and Western Medicine Hospitals

Researchers Design Quality Indicator System for Integrated Chinese and Western Medicine Hospitals

November 19, 2026
Automated Echocardiography Artificial Intelligence Cuts Analysis Time While Preserving Ejection Fraction Accuracy

Automated Echocardiography Artificial Intelligence Cuts Analysis Time While Preserving Ejection Fraction Accuracy

November 19, 2026

Brain Structural Alterations Show Marked Regional Heterogeneity Across Insomnia and Depression

November 19, 2026

Stretchable Nanocomposites Enable Adaptive Closed-Loop Bioelectronic Systems Across Dynamic Tissues

November 19, 2026

POPULAR NEWS

  • Study Finds Autism Diagnosed in Twenty-Eight Percent of Children With Neurofibromatosis Type 1

    Study Finds Autism Diagnosed in Twenty-Eight Percent of Children With Neurofibromatosis Type 1

    29 shares
    Share 12 Tweet 7
  • Esketamine with Pulsed Radiofrequency Alleviates Neuropathic Pain in Herpes Zoster Patients

    29 shares
    Share 12 Tweet 7
  • Fences Restrict Movement and Alter Space Use in Endangered Masai Giraffes

    29 shares
    Share 12 Tweet 7
  • Stem Cell Delivery Repairs Post-Stroke Endoplasmic Reticulum Dynamics in Rat Model

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Researchers Design Quality Indicator System for Integrated Chinese and Western Medicine Hospitals

Automated Echocardiography Artificial Intelligence Cuts Analysis Time While Preserving Ejection Fraction Accuracy

Brain Structural Alterations Show Marked Regional Heterogeneity Across Insomnia and Depression

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 89 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.