Diabetes is the body’s inability to efficiently break down sugar or secrete enough insulin required to process glucose, which supports normal bodily functions. Diabetes, as a prevalent chronic disorder, has contributed to numerous underlying health challenges among its carriers and is classified by the WHO as the world’s deadliest disease and silent killer. Its non-communicable nature makes early diagnosis difficult, allowing progression through various stages: type I, type II, pre-diabetes, and gestational. This challenge is further compounded by the imbalanced nature of diabetes datasets, which leads to high misclassification, poor generalization, and reduced accuracy. This study predicts diabetes using a bi-directional long short-term memory (BiLSTM) model applied to two datasets: (a) PIMA Indian Diabetes and (b) Iraqi Society Dataset, to evaluate the impact of six known balancing techniques and assess their effectiveness. Results show that for PID, the SMOTE-Tomek fused BiLSTM outperforms other balancing schemes with F1, Accuracy, Precision, Recall, and Specificity scores of 0.9182, 0.9198, 0.9128, 0.9248, and 0.9208, respectively. For ISD, it also achieves the best performance with values of 0.9367, 0.9369, 0.9386, 0.9388, and 0.9313, respectively. Other balancing approaches yielded F1 scores ranging from [0.6751 to 0.9347], accuracy [0.684 to 0.9358], Precision [0.6851 to 0.9296], Recall [0.6639 to 0.9356], and specificity [0.6658 to 0.9298]. These results imply that BiLSTM is resilient to the vanishing gradient problem and can effectively classify diabetes cases with enhanced performance.

Breast cancer is the most prevalent cancer among women worldwide, requiring early and accurate diagnosis to reduce mortality. This study proposes a hybrid classification pipeline that integrates Hybrid Statistical Feature Selection (HSFS) with unsupervised LSTM-guided feature extraction for breast cancer detection using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. Initially, 20 features were selected using HSFS based on Mutual Information, Chi-square, and Pearson Correlation. To address class imbalance, the training set was balanced using the Synthetic Minority Over-sampling Technique (SMOTE). Subsequently, an LSTM encoder extracted non-linear latent features from the selected features. A fusion strategy was applied by concatenating the statistical and latent features, followed by re-selection of the top 30 features. The final classification was performed using a Support Vector Machine (SVM) with RBF kernel and evaluated using 5-fold cross-validation and a held-out test set. Experimental results showed that the proposed method achieved an average training accuracy of 98.13%, F1-score of 98.13%, and AUC-ROC of 99.55%. On the held-out test set, the model reached an accuracy of 99.30%, precision of 100%, and F1-score of 99.05%, with an AUC-ROC of 0.9973. The proposed pipeline demonstrates improved generalization and interpretability compared to existing methods such as LightGBM-PSO, DHH-GRU, and ensemble deep networks. These results highlight the effectiveness of combining statistical selection and LSTM-based latent feature encoding in a balanced classification framework.