Neuro-symbolisk AI för släktträdsanalys : Ett neuro-symboliskt frågesvarssystem för släktträdsanalys vid bedömning av ärftlig cancerrisk

The use of Machine Learning (ML) & Deep Learning (DL) in healthcare has shown promise in improving diagnostics, treatment planning, and predictive tasks. However, their adoption in healthcare is also limited by the lack of transparency in how these models derive their outputs. This problem is also more commonly known as the ’black-box’ problem. This thesis investigates the use of Neuro-Symbolic Artificial Intelligence (NeSy AI) as a potential solution, focusing on pedigree analysis for hereditary cancer risk assessment. NeSy AI combines the pattern recognition capabilities of neural networks with the structured, rule-based reasoning of symbolic AI. This hybrid approach addresses the transparency limitations of traditional ML models. Despite the recent interest in the field, its application in healthcare is limited and relatively unexplored. To analyze this, a modular Question Answering (QA) system was developed. This QA model integrates a fine-tuned Large Language Model (LLM) for parsing natural language queries into symbolic programs, and a symbolic executor that reasons over structured, synthetically generated pedigree data, aligned with established clinical guidelines. The system was evaluated on 15 synthetic pedigrees, and 26 questions were asked about each pedigree’s unique features and cancer risk. Results showed that the fine-tuned NeSy QA model achieved high end-to-end accuracy (91.5%) and lower query latency (0.95 seconds), compared to the baseline’s 90.3% and 11.97 seconds latency, and consistently outperformed the baseline in its reasoning capabilities. Additionally, the model showcased robustness to linguistic noise and could generalize to Swedish-language queries. All in all, this thesis provides evidence that NeSy AI can be an interesting option for automating pedigree analysis in hereditary cancer risk assessment. By combining statistical learning with symbolic reasoning, the system indicates enhanced interpretability in the decision-making process while still not compromising performance.