International Journal of Science, Engineering and TechnologyAuthors: Sonu Kumar, Omvir Singh, Ankush Kumar
Abstract: Accurate disease diagnosis from patient-reported symptoms remains a significant challenge in modern healthcare due to the complexity of symptom interpretation and the growing volume of unstructured medical text. This paper presents a hybrid deep learning model that integrates Bidirectional Long Short-Term Memory (BiLSTM) networks with an Attention Mechanism to classify diseases from free-text symptom descriptions. The BiLSTM component captures sequential and contextual dependencies in symptom narratives, while the attention layer dynamically prioritizes the most diagnostically relevant terms, improving both predictive accuracy and interpretability. The model was trained on the Symptom2Disease dataset following advanced text pre-processing, including lemmatization, stop-word removal, and TF-IDF/tokenized sequence representation, with class-weight balancing to address label imbalance. Evaluated using stratified 5-fold cross-validation, the proposed model achieved a mean accuracy of 90.51% (σ = 1.81%), and a final accuracy of 91.83% after training on the complete dataset. The model outperformed benchmark methods including k-Nearest Neighbours (81.3%), Random Forest (85.5%), and a standard Recurrent Neural Network (88.1%). These results demonstrate that combining sequential modelling with attention-driven feature prioritization yields a robust, interpretable, and clinically promising tool for automated disease prediction, with potential applications in telemedicine and clinical decision support.
DOI: http://doi.org/
