Explainable AI-Driven Improved Disease Prediction through Symptom Analysis with Custom BERT

This study proposes a custom BERT-based hybrid model aimed at enhancing the accuracy and explainability of disease prediction based on patient symptom descriptions. Utilizing the Symptom2Disease dataset, which includes 1,200 samples across 24 disease categories, the model processes textual symptom data and maps it to corresponding diseases. Key contributions of this work include the integration of Explainable AI (XAI) techniques to provide transparency in the model’s decision-making process and the employment of dropout to mitigate overfitting, ensuring that the model generalizes well on unseen data. The proposed model builds upon BERT’s pre-trained architecture by incorporating additional layers that refine the embeddings and enhance classification accuracy. Through rigorous experimentation, the model achieved an impressive 99% accuracy, outperforming existing models such as DistilBERT and MCN-BERT in both accuracy and training efficiency. The inclusion of XAI allows for visualizations of word-level contributions, offering insights into how specific symptoms influence disease predictions. Furthermore, a comparative analysis with previous studies demonstrates that the proposed model not only performs with higher accuracy but also achieves faster training times, making it both computationally efficient and effective for practical medical applications. The paper highlights the potential for applying hybrid architectures in healthcare, improving disease prediction through advanced natural language processing techniques.