[Medical prior-guided TabMap deep learning model for ovarian cancer prediction and interpretability analysis].

objectivesTo develop a TabMap image mapping and deep learning prediction framework that integrates medical prior knowledge to address the challenges of complex feature associations in tabular medical data and insufficient model interpretability in early ovarian cancer diagnosis.

methodsBased on clinical semantics, 36 medical diagnostic features were partitioned into 4 spatially continuous regions, namely the routine blood test partition, biochemical indicators partition, tumor markers partition, and other (coagulation, inflammation, and other clinical variables) partition. The Gromov-Wasserstein optimal transport algorithm was then employed to solve the optimal coupling between feature space and pixel space, thus generating TabMap images that preserve the topological structures. A lightweight convolutional neural network incorporating SE attention mechanism and global average pooling was designed to handle sample imbalance using class-weighted loss and weighted sampling strategies. Finally, class activation mapping (CAM) technique was utilized to visualize the model's decision-making process for interpretability analysis.

resultsExperiments on a real ovarian cancer dataset demonstrated that the proposed method achieved a test accuracy of 91.82% with a precision of 89.96%, recall of 89.45%, F1-score of 0.8970 and balanced accuracy of 88.51%, representing accuracy improvements of 8.26%-14.93% over traditional machine learning methods and 2.28%-13.60% over standard deep learning models. Interpretability analysis showed that the model exhibits pronounced activation patterns in the tumor-marker region as well as in coagulation-, inflammation-, and age-related zones, consistent with established clinical understanding. The feature-importance ranking further underscored the pivotal roles of CA125, HE4, and other key biomarkers in guiding the model's decisions.

conclusionsThe proposed medical prior-guided TabMap method effectively integrates domain knowledge with data-driven learning, which enhances both its prediction performance and clinical interpretability. This strategy provides a novel approach for deep learning modeling of tabular medical data with good clinical potentials.