Ovarian cancer screening moves past a single marker: PSO-tuned fusion of four rankings lifts full-pipeline ablation accuracy to 85.51%
Synopsis
Early ovarian cancer diagnosis leans on serum markers and routine laboratory tests, yet redundant, noisy and correlated variables in high-dimensional clinical data degrade both classification and interpretability. A hybrid biomarker selection pipeline fuses correlation analysis, mutual information, XGBoost importance and SHAP relevance with particle-swarm-optimized adaptive weights, feeding an attention-enhanced multi-scale 1D convolutional network; on a public Kaggle ovarian cancer clinical dataset under stratified 10-fold cross-validation it reports 93.35% accuracy and 90.66% ROC-AUC. In ablation, the full pipeline reaches 85.51% accuracy, 19.33 percentage points above a baseline 1D-CNN without selection or attention.
Interpretation
Ovarian cancer classification no longer takes every clinical laboratory variable as input; it uses a biomarker subset chosen by fusing four complementary rankings, then passed to an attention-enhanced multi-scale 1D convolutional network. Earlier approaches mostly relied on a single feature-selection strategy or gave fixed weights to several rankings, so redundant and correlated markers tended to remain in the model. The ablation shows a baseline 1D-CNN without biomarker selection or attention at 66.18% accuracy, rising to 81.88% with hybrid selection and 85.51% for the full pipeline, a 19.33 percentage-point gap; the intermediate multi-scale model's accuracy is written as 76.13% in the ablation and 69.63% in the comparison, so the two figures disagree.
The hybrid score places statistical ranking, ensemble importance and explainability-driven relevance inside one weighting scheme whose weights are learned from the data. Particle swarm optimization here searches the weight of each of the four rankings so that the fused biomarker ordering performs best for the downstream classifier. Evaluation uses a public Kaggle ovarian cancer clinical dataset with stratified 10-fold cross-validation, and Wilcoxon signed-rank and Friedman tests for significance.
The classifier runs three parallel 1D convolution branches with different kernel lengths to extract biomarker patterns, then an attention module assigns weights to feature channels. A single-scale 1D convolution sees the biomarker sequence through one receptive field, which makes it hard to cover local fluctuations and longer-range relationships at once. With kernel lengths 3, 5 and 7 plus batch normalization, residual connections and global average pooling, stratified 10-fold cross-validation reports 93.35% accuracy, 93.18% precision, 94.25% recall and 93.71% F1.
Perspective
The results rest on a secondary analysis of a public Kaggle ovarian cancer clinical dataset, where inputs are structured numeric tables of serum markers such as CA125, HE4, CA19-9 and CEA plus blood count and biochemistry variables, and the task is binary classification of ovarian cancer versus non-cancer. Evaluation uses stratified 10-fold cross-validation and reports accuracy, precision, recall, F1 and ROC-AUC. SHAP and LIME provide global and single-case explanations, with CA125 ranked first by SHAP, followed by HE4, age, CA19-9, BMI and menopausal status. The intended readers are researchers and engineers working on clinical marker modeling and explainable AI.
The training and validation accuracy curves show a clear gap, which the source attributes to a relatively small and unbalanced clinical dataset and reads as a risk of overfitting. Some metrics are not fully consistent across figures: the intermediate multi-scale model's accuracy is written as 76.13% in the ablation and 69.63% in the comparison. All results come from a secondary analysis of one public dataset and have not been validated on independent or multicenter cohorts.