A hybrid CNN-XLSTM-GRU deep learning model with autoencoder-based feature selection for hypothyroidism diagnosis.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Hypothyroidism diagnosis is challenging due to non-specific symptoms overlapping with other conditions.
• Accurate and early detection of hypothyroidism is crucial for effective patient management.
• Existing diagnostic methods may lack the precision needed for complex cases.

Purpose of the Study:

• To develop and evaluate FusionNet-CXG, a novel hybrid deep-learning model for hypothyroidism prediction.
• To assess the performance of FusionNet-CXG against established deep learning architectures.
• To enhance the interpretability of the model's predictions using feature influence analysis.

Main Methods:

• A hybrid deep-learning model (FusionNet-CXG) integrating CNN, eXtended LSTM, and GRU components was developed.
• Experiments utilized a public dataset of 3,772 records with 30 features, addressing class imbalance with SMOTE-NC.
• Performance was rigorously evaluated using 10 repetitions of 5-fold stratified cross-validation (50 folds total).

Main Results:

• FusionNet-CXG achieved a mean accuracy of 0.9394, F1-score of 0.906, and AUC-ROC of 0.94.
• The hybrid model outperformed baseline models including CNN+LSTM and CNN+BiLSTM.
• SHAP analysis was employed to provide insights into feature importance for model predictions.

Conclusions:

• The combination of local feature extraction (CNN) and recurrent modeling (eXtended LSTM, GRU) is effective for hypothyroidism prediction.
• FusionNet-CXG demonstrates significant potential for improving the accuracy of hypothyroidism diagnosis.
• Future research will focus on external validation and incorporating temporal clinical data for enhanced clinical relevance.