BreastCancerDiagNet - Transformer-Based Clinical Question Generation for Automated History Taking

History-taking is a highly significant procedure in clinical decision-making that remains time-consuming, inconsistent, and can lead to omissions. This study presents BreastCancerDiagNet, a complex transformer-powered system to automate medical history-taking and aid in diagnosis. This model uses structured patient demographics and unstructured clinical symptoms with hybridized ClinicalBERT embeddings, BiLSTM sequence modeling, and self-attention mechanisms. These capabilities are integrated in an encoder-decoder architecture with rotary position embeddings and FlashAttention to enable long-sequence processing. A Reinforcement Learning with Human Feedback (RLHF) strategy is used to refine the question generation strategy to reflect contextual reference to clinical practice. The proposed system was trained and tested using a breast cancer dataset of curated demographic data, symptom data, comorbidities, lifestyle indicators, and physician-curated ground truth questionnaires. The results show that BreastCancerDiagNet achieved a BLEU-4 score of 0.42, a ROUGE-L score of 0.56, and a BERT-F1 score of 0.88, which are higher than the Seq2Seq and Vanilla Transformer baselines. Qualitative analysis confirmed the relevance of the questions generated in clinical practice, covering lump presence, pain, discharge, family history, and drug use. The findings demonstrate the possibility of using BreastCancerDiagNet to save time in consultations, reduce the number of diagnostic errors, and serve as a future-generation Clinical Decision Support System (CDSS) that can be scaled and interpreted.