arXiv:2611.12940 applies CQL and MOPO to colon-cancer chemo scheduling and reports no endpoint
arXiv:2611.12940 stacks 2020 offline RL methods, CQL and MOPO, on observational colon-cancer chemotherapy trajectories after a simulated Q-learning baseline. The abstract asserts the model supports junior oncologists but reports no N, action space, or effect size. Without an off-policy comparison to the logging policy, efficacy and toxicity claims are not evidenced.
arXiv:2611.12940, submitted 6 November 2026 by Chamani Shiranthika Jayakody Kankanamalage, describes a thesis pipeline for chemotherapy scheduling. The online half is standard Q-learning on a simulated tumor-drug model. The offline half applies Conservative Q-Learning and Model-based Offline Policy Optimization to observational human colon-cancer trajectories. The PDF is listed at 7,729 KB. The abstract names a Chemo RL model and says it supports junior oncologists. It states no dataset accession, cohort size, or primary endpoint.
The abstract’s only evaluation is the phrase "caters well." That is not a result. Kumar, Zhou, Tucker, and Levine, arXiv:2006.04779, define CQL as a penalty on actions outside the data so the Q-function does not exceed the behavior policy. Yu and coauthors, arXiv:2005.13239, define MOPO as short-horizon planning inside a learned dynamics model with an uncertainty penalty. Both papers score D4RL returns, not survival or grade 3 or higher toxicity. Those scores do not transfer. Absent weighted importance sampling or a held-out comparison to the logging policy, the chemotherapy claim has no measured effect size.
Simulated Q-learning on an ordinary differential equation tumor model optimizes a designed reward: tumor burden against drug level. Observational colon-cancer data answer a different question, and they are confounded by indication. Patients with worse disease receive denser schedules, so an uncorrected value function treats intensity as benefit. CQL’s penalty pulls the learned policy back toward that logging policy. Cloning an existing regimen is not a reduction in toxicity and not a gain in efficacy. Calling the 2020 algorithms the novelty mislabels an application study. The unpublished piece is the validation design, not another controller variant.
A tool that proposes a regimen for a named patient is a clinical decision function, not a demo. Deployment then requires a frozen model, a prespecified endpoint on toxicity and overall survival, and a comparison with the clinician policy that generated the data. Cycle-level actions also have to match real schedules: drug, dose bin, and interval, not a continuous control input from a textbook model. Until the PDF reports cohort size, that action space, and an off-policy estimate versus the logging policy, Chemo RL remains an unevaluable claim about scheduling support.
Sources (3)
- [1]Reinforcement Learning based control of Cancer Chemotherapy treatments(https://arxiv.org/abs/2611.12940)
- [2]Conservative Q-Learning for Offline Reinforcement Learning(https://arxiv.org/abs/2006.04779)
- [3]MOPO: Model-based Offline Policy Optimization(https://arxiv.org/abs/2005.13239)