Speaker
Description
Non-small cell lung cancer (NSCLC) treated with tyrosine-kinase inhibitors (TKIs) initially responds well but rapidly develops resistance. This happens because in heterogeneous tumour microenvironments, continuous standard-of-care treatment rapidly eliminates drug sensitive cells, enabling resistant populations to dominate and leading to treatment failure. Evolutionary therapies (ET), guided by mathematical models, aim to delay treatment failure and prolong survival by maintaining a sufficient sensitive population to suppress resistance, and have shown promise in recent prostate cancer trials. Clinical implementation of ET in NSCLC is challenging due to its fast growth and limited biomarkers. Recent work suggests that deep reinforcement learning (DRL) can generate more robust, patient specific treatment schedules. We present a DRL framework tailored to NSCLC treated with TKIs. Using virtual patients based on clinical data, our learned treatment schedules predict improved survival across all patients compared with both standard-of-care protocols and conventional ET strategies. We then extend the DRL reward structures to incorporate patient specific quality of life preferences, enabling evaluation of treatment policies through quality-adjusted survival and supporting more clinically meaningful decision making.
