Hypothyroidism is a widespread endocrine disorder characterized by vague and non-specific symptoms, making early diagnosis particularly challenging. The increasing availability of large-scale clinical data presents an opportunity to develop robust, data-driven diagnostic systems. However, traditional machine learning approaches often sacrifice interpretability for accuracy, limiting their deployment in clinical settings. In this study, we present DeepSeqNet, a hybrid and interpretable deep learning framework designed for scalable hypothyroidism diagnosis using sequential patient records. Our architecture synergizes convolutional neural networks (CNN), long short-term memory (LSTM) units, and fully connected (ANN) layers to extract and learn spatial-temporal dependencies in heterogeneous clinical datasets. To ensure transparency, we integrate Polynomial-SHAP, a novel feature attribution method, enabling precise, nonlinear interpretability of model outputs. The framework was evaluated on real-world clinical datasets, achieving 99.34% accuracy, 93.85% precision, 98.39% recall, 96.06% F1-score, and 99.42% specificity on the test set. Performance on the validation set was consistently high, with metrics of 99.59 ± 0.31% accuracy, 99.24 ± 0.42% precision, 99.95 ± 0.04% recall, and 99.59 ± 0.21% F1-score, demonstrating the model’s robustness with minimal performance drop during generalization. Class-wise analysis further shows superior detection of hypothyroidism (F1-score: 96.06%) and normal cases (F1-score: 99.64%). By embedding interpretability into a high-performing deep learning pipeline, our method supports not only accurate prediction but also clinician insight into influential risk factors. While the dataset originates from a single region, the model’s modular design facilitates broad applicability across diverse populations. This study highlights the potential of interpretable AI in improving clinical decision-making for endocrine disorders using big medical data.
Ejiyi et al. (Tue,) studied this question.