Related Experiment Video
Updated: Jul 3, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Enhancing the positive predictive value of early-stage ovarian cancer detection using a two-step machine learning
Mikio Mikami1,2,3, Kazuhiro Tanabe4,5, Saki Nagaki6
1Department of Obstetrics and Gynecology, Tokai University School of Medicine, Isehara, Kanagawa, Japan. mmikami@tokai.ac.jp.
Background:
Early detection of epithelial ovarian cancer (EOC) remains a major clinical challenge. Although serum tumor markers are widely used for detection, their diagnostic performance remains limited. We previously developed a comprehensive serum glycopeptide spectrum analysis (CSGSA) approach that integrates tumor marker measurements and enriched glycopeptides (EGPs) using convolutional neural networks. In this study, we evaluated whether a two-step LightGBM framework incorporating cancer antigen 125 (CA125), human epididymis protein 4 (HE4), cancer antigen 72 - 4 (CA72-4), and EGPs could improve the diagnostic specificity and projected positive predictive value (PPV) for EOC detection compared with conventional biomarker-based approaches.
Methods:
The study included 553 patients with EOC and 1,144 non-EOC controls (healthy individuals or patients with benign conditions). Serum levels of CA125, HE4, and CA72-4 were measured along with 1,712 EGPs. Diagnostic models were developed using machine learning algorithms and evaluated for accuracy, area under the receiver operating characteristic curve (ROC-AUC), PPV, and negative predictive value (NPV).
Results:
The highest diagnostic performance was achieved using a two-step classification framework. First, patients were stratified into high-, intermediate-, and low-risk groups based on tumor markers and age. Second, the intermediate-risk group was reclassified using a model incorporating EGP-derived features. Among the evaluated algorithms, LightGBM achieved the best performance, yielding a prevalence-adjusted (projected) PPV of 18.7% and an NPV of 99.99%. At a predefined specificity of 99.5%, the corresponding sensitivity was 65%.
Conclusions:
The CSGSA method combined with a two-step LightGBM framework demonstrated promising diagnostic performance in an internally validated cohort, with improved specificity and prevalence-adjusted PPV compared with conventional biomarker-based approaches. However, prospective external validation in independent populations is required before clinical implementation or generalizability can be established.
