Incorporating medical free text into data-driven support for lung cancer multidisciplinary team meetings
Year: 2026
Type: dissertation
Abstract: Lung cancer is the deadliest and most frequently diagnosed cancer worldwide. Mul- tidisciplinary Team (MDT) meetings are fundamental in determining the treatment of lung cancer patients, and Clinical Decision Support Systems (CDSS) can optimize the decision-making process within these meetings by providing data-driven insights. While such systems have shown success in supporting clinical decisions, their application in MDTs is still underexplored, and they often rely solely on structured data. Furthermore, much of the relevant information used in these meetings lies in an unstructured format, such as Electronic Health Records (EHR), complicating analysis and limiting the data’s contribution to the decision-making process. Transformer-based architectures, in particu- lar Bidirectional Encoder Representations from Transformers (BERT), offer a promising solution as they are capable of processing unstructured data and can be fine-tuned for specific problem instances, such as clinical tasks. Nonetheless, the optimal strategies for fine-tuning these models for treatment classification and to gain insights on the model output remain open questions. In this thesis, we demonstrate that fine-tuning BERT-based models significantly outperforms baseline un-fine-tuned approaches in classifying treatment modalities from unstructured EHRs. Additionally, we extract keywords from the EHRs to gain visibility into the model’s learned representations, providing a more interpretable view of the classification process. This project, conducted in collaboration with Instituto Português de Oncologia de Lisboa (IPOL), utilizes institutional clinical data and expert validation to develop a model capable of supporting MDT decision-making, by translating unstructured clinical narratives into structured treatment suggestions.
Institution: Universidade Nova de Lisboa
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213741103
