Integrating Temporal and Spatial Learning for Prostate Cancer Prediction using a Pruned 1D-CNN BiLSTM Hybrid Model

Prostate cancer remains one of the most common and serious cancers affecting men worldwide. Early and precise diagnosis is crucial for improving patient outcomes and directing timely medical actions. In this work, we present a new deep learning-based diagnostic model that combines a 1D convolutional neural network (1D-CNN) with bidirectional long short-term memory (BiLSTM), optimized through a feature selection strategy called adaptive medical feature pruning (AMFP). Our approach uses the prostate cancer dataset from Kaggle, which includes key biological indicators related to prostate health. A strong preprocessing pipeline that includes label encoding, min-max normalization, and filling in missing values ensures high-quality input data. AMFP is vital; it applies entropy-based analysis, filtering by the coefficient of variation, and removing multicollinearity to refine the feature set. This significantly improves model generalization and learning efficiency. We conducted extensive evaluations using stratified 5-fold cross-validation. Our model consistently outperformed traditional classifiers such as random forest, SVM, XGBoost, MLP, and Logistic Regression. It achieved impressive performance metrics, including an average accuracy of 93.60%, precision of 92.89%, recall of 94.40%, F1-score of 93.63%, and ROC-AUC of 94.96%. Ablation studies further show the important contribution of AMFP in enhancing model robustness. Additionally, the model has strong computational efficiency, with a training time of just 8.4 seconds and an inference time of 0.0021 seconds per sample. This makes it wellsuited for real-time clinical diagnostic support. Overall, our proposed hybrid architecture offers a lightweight, interpretable, and high-performing solution for early prostate cancer detection, with significant potential for use in clinical settings.