Explainable Multitask Transformers for Early Detection of Smoking Behaviors and Lung Cancer Symptoms from Danish Electronic Health Records.

Lung cancer is the leading cause of cancer-related mortality, and early detection is challenging due to asymptomatic onset. Accurate detection of smoking behaviors and symptoms is essential for risk evaluation and timely intervention. This study developed multitask transformer-based models to simultaneously detect smoking status, pack-year, and lung cancer symptom categories from Danish electronic health records. Texts were annotated for smoking behaviors, and 16 lung cancer-related symptoms were grouped into six categories. Preprocessing and tokenization were performed, and three transformer models, including BioBERT, XLM-RoBERTa, and Danish BERT, were trained in a multitask framework. Model performance was evaluated, and LIME was applied for explainability. BioBERT performed best for smoking status (Accuracy: 97.93%, F1-score: 86.30%), while XLM-RoBERTa excelled in pack-year (Accuracy: 96.23%, F1-score: 92.02%) and symptom detection (Accuracy: 99.55%, F1-score: 96.67%). LIME highlighted key predictive features, enhancing transparency. These results demonstrate that multitask transformer models can support early risk assessment and improved clinical decision-making, with future work applying large language models.