Learning chemotherapy schedules from patient records: a small study with major limits

From simulated patients to hospital records

The supplied manuscript is a National Taipei University master’s thesis dated 2021, accompanying a preprint posting listed in November 2026. It investigates reinforcement learning, a machine-learning approach that searches for sequences of actions that maximize a numerical score, or reward.

The research begins with conventional Q-learning in mathematical simulations. Equations describe changes in tumor cells, normal cells, immune cells and drug concentration. The software explores dosing schedules for three hypothetical patient categories and additional simulated patients.

These experiments illustrate how a learning algorithm behaves inside a specified model. They do not establish how people would respond. The thesis itself identifies a central limitation: equations informed by animal studies cannot adequately capture the complexity of human cancer and treatment.

The main project, called Chemo RL, therefore turns to offline reinforcement learning. Rather than learning by experimentally treating patients, offline learning uses previously collected records. This avoids exposing people to exploratory treatment decisions during training, but does not by itself make the resulting recommendations safe.

What the system learned from

The clinical dataset contained records from 40 people with stage IV colon cancer receiving first-line treatment at a hospital in Taiwan between 2017 and 2020. Patients were aged 39–84. The thesis reports approximately 575 treatment records, although counts differ slightly across sections.

Inputs included tumor measurements, body measurements, side effects, treatment response, other illnesses and performance status—a measure of everyday functioning. Missing tumor measurements were filled using interpolation, meaning intermediate values were estimated rather than directly observed.

A first component used supervised learning to reproduce an oncologist-assigned patient-state score. The researcher compared a convolutional neural network with a random forest, which combines predictions from multiple decision trees. These scores then fed into the treatment-learning framework.

For the principal dosing model, patients with fewer than six or more than 15 treatment cycles were excluded. That left 27 patients for training and seven for testing. The system considered six drugs, although the only patient receiving oxaliplatin was excluded, leaving no retained clinical example for that drug.

The framework combined Conservative Q-Learning, intended to discourage overly optimistic estimates for poorly supported actions, with Model-Based Offline Policy Optimization, which generates simulated transitions and penalizes model uncertainty. Neither mechanism guarantees safety in this dataset.

Recommendations also blended the learned policy with historical treatment decisions: 60% weight went to the target policy and 40% to the recorded behavior policy. Similarity to clinicians’ prescriptions was therefore partly built into the output.

What the results show—and do not show

The random forest generally performed better than the neural network at reproducing the assigned patient-state labels. However, the reported random-forest test accuracy is inconsistent: one passage gives 92.37%, while its confusion matrix gives about 89.2%. The neural-network confusion matrix reports about 82.1%.

Those numbers concern agreement with human-assigned labels, not whether a proposed treatment benefits a patient. The labels were also highly imbalanced, with no examples for scores zero through two. The thesis describes upsampling for the neural network but does not clearly establish that duplicated examples were kept separate across training and evaluation.

For dosing, recommendations became closer to recorded prescriptions as training cohorts increased from five to 27 patients. Reported mean cetuximab dose differences, for example, fell from 247 to roughly 107 milligrams. This measures agreement with past prescribing, not identification of a better treatment.

The system also assigned higher predicted rewards to its suggested schedules than to recorded schedules in illustrated comparisons. Its reward function combined tumor change, weight change, side effects, treatment response and death, with substantially different numerical weights. Higher predicted scores therefore reflect both the model’s assumptions and the researcher’s chosen priorities; they are not observed improvements in health.

Why clinical conclusions remain premature

The study reports no prospective trial in which patients received the suggested schedules, and no demonstrated survival or toxicity benefit from those schedules. Its small, single-hospital dataset provides limited evidence about performance elsewhere. No confidence intervals are supplied for the headline comparisons.

Historical records also do not reveal what would have happened had the same patient received a different dose. Unrecorded factors influencing both prescribing and outcomes could mislead the model. Compressing clinical information into one patient-state score could further discard important distinctions.

The thesis presents the system as potential support for oncologists, but its dedicated limitations section is empty. The supported conclusion is narrower: Chemo RL generated treatment suggestions from a small retrospective dataset. Whether those suggestions are useful, effective or safe remains unestablished.