Night Science Workshop: Causal Inference and Reinforcement Learning

Mathematical modeling in radiation biology and oncology has a rich history, leveraging animal, in vitro, and clinical data. These models, while insightful, are limited in capturing complex features like patient demographics, treatment specifics, or omics and imaging data. Causal machine learning (CML) excels at incorporating such multimodal features into causal inference analyses aimed at estimating treatment effects. Integrating mechanistic concepts, like Biologically Effective Dose (BED), as engineered features into CML can enhance both accuracy and interpretability, guiding clinically actionable insights. Our study applied this approach to the RADCURE dataset on head and neck cancers, combining dose-independent (DI) and dose-dependent (DD) tumor repopulation models with Causal Survival Forests (CSF). BED was calculated using tumor killing and repopulation terms, with the DD variant assuming repopulation onset and rate depend on cell-killing intensity. This CML approach integrated mechanistic modeling concepts (BED variants) with clinical and radiomics data. The analysis suggested that high BED DD and BED DI, alongside chemotherapy, causally increased restricted mean survival time (RMST) and survival probability. Chemotherapy effects exceeded those reported in previous meta-analyses, possibly due to younger/healthier patients receiving treatment and/or evolving treatment protocols. Refutation tests and sensitivity analyses were used to test the robustness of causal modeling. Such integration of mechanistic and CML models on non-randomized clinical data can support personalized medicine, optimize radiotherapy protocols, and generate hypotheses to complement randomized clinical trials.

Add to Calendarcalendarcalendarcalendarcalendar

1. Symbolic Reasoning for Enhancing LLMs and RL: Reinforcement Learning via Symbolic Feedback (RLSF) leverages symbolic tools like provers and simulators to provide fine-grained feedback for fine-tuning LLMs. This approach achieves significant advancements in chemistry, programming, and arithmetic tasks, achieving substantial model size reductions while outperforming GPT-4.
    
2. Reinforcement Learning for Improving Symbolic Reasoning Tools: AlphaMapleSAT (AMS) demonstrates the power of RL techniques, such as Monte Carlo Tree Search (MCTS) with symbolic feedback, in enhancing parallel SAT solving. This method achieves a significant speed-up on challenging combinatorial problems and finds applications in areas like quantum foundations.

Add to Calendarcalendarcalendarcalendarcalendar