Fig 1.
Workflow of the proposed FusionNet-CXG framework for hypothyroidism prediction.
The pipeline integrates preprocessing, class balancing, feature extraction, and a hybrid CNN–XLSTM–GRU classifier.
Fig 2.
Distribution of thyroid-related hormones (TSH, TT4, T3, and FTI) in the dataset.
Fig 3.
Gender distribution of patients in the dataset.
Fig 4.
Comparison of class balancing techniques.
Fig 5.
Relative contribution of key features.
Table 1.
Optimal Hyperparameter Settings for the Proposed Model.
Table 2.
FusionNet-CXG layer architecture.
Table 3.
Proposed FusionNet-CXG model architecture and parameters.
Table 4.
Experimental Configuration and Pipeline Summary (Hypothyroid Binary Classification).
Table 5.
Software and hardware configuration.
Table 6.
Experimental pipeline and methods (FusionNet-CXG).
Fig 6.
Heatmap of pairwise feature correlations.
Fig 7.
Autoencoder training and validation loss curves for different class balancing approaches: (a) Random Oversampling (b) Random Under sampling (c) ADASYN (d) Borderline-SMOTE (e) SMOTE-NC.
Fig 8.
Correlation between original features and encoded features.
Fig 9.
Illustrates the test Accuracy across hyperparameter tuning results of CNN-LSTM.
Fig 10.
(a) Training and validation loss of CNN-LSTM classifier over epochs. (b) accuracy score between training and validating data over epochs of CNN-LSTM classifier.
Fig 11.
Depicts the test Accuracy across Hyperparameter Configurations for CNN + BiLSTM.
Fig 12.
(a) Training and validation loss of the CNN + BiLSTM classifier. Fig12 (b) training and validation accuracy over epochs of the CNN + BiLSTM classifier.
Fig 13.
Test accuracy across different hyperparameter configurations for the FusionNet-CXG model.
Fig 14.
(a) Training and validation loss of the FusionNet-CXG model. Fig 14 (b) validation and training accuracy of the FusionNet-CXG model.
Fig 15.
Model accuracy among different genders (with 95% Wilson CI).
Table 7.
Overall performance of FusionNet-CXG across 50 runs (mean ± SD; 95% CI).
Table 8.
Overall Performance of FusionNet-CXG (Mean ± SD; 95% CI).
Table 9.
Comparative Performance of CNN+LSTM, CNN + BiLSTM, and Proposed Model with Mean ± SD and 95% Confidence Intervals.
Table 10.
Statistical comparison based on 10 repetition-wise mean accuracies obtained from 10 × 5 repeated stratified cross-validation.
Table 11.
Verification of Assumptions for Paired t-Test.
Fig 16.
Mean performance comparison of the deep learning models on cross-validation test folds under 5-fold stratified cross-validation repeated 10 times.
Fig 17.
SHAP-Based Explainability of the Proposed Hypothyroid Diagnostic Model.
Table 12.
Performance comparison of the Proposed model with state-of-the-art models.