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

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Mean accuracy 0.9394 (93.94%)

• Primary Outcome: Diagnostic accuracy for hypothyroidism detection

• Nnt Or Sensitivity: Sensitivity and specificity not reported; AUC-ROC = 0.94

• Confidence Interval: Not reported

Clinical Application

Limited - requires computational infrastructure and validation in clinical settings Would require TGA approval as diagnostic software, integration with Australian pathology systems, and validation against local thyroid function testing protocols Patients suspected of hypothyroidism, though specific clinical context unclear

Abstract

Hypothyroidism, caused by reduced thyroid-hormone production, is often difficult to diagnose because its symptoms are non-specific and overlap with other disorders. We developed FusionNet-CXG, a hybrid deep-learning model that couples CNN, eXtended LSTM, and GRU components for hypothyroidism prediction. Experiments were performed on a public dataset (3,772 records; 30 features). Class imbalance was mitigated using SMOTE-NC, and performance was estimated using 10 repetitions of 5-fold stratified cross-validation, yielding 50 fold-level evaluations to obtain stable results. Across folds, FusionNet-CXG achieved 0.9394 mean accuracy, F1 = 0.906, and AUC-ROC = 0.94, and it exceeded CNN+LSTM and CNN + BiLSTM baselines under the same protocol. The results indicate that combining local feature extraction with recurrent modelling is effective for this task. To support interpretability, SHAP was used to quantify feature influence on the predictions. Future work will focus on external validation using more diverse real-world cohorts and incorporating temporal clinical data to improve clinical relevance.