In simulated trials of 50 participants, researchers utilized an RL model for the glioblastoma treatments, temozolomide (TMZ) and procarbazine, lomustine and vincristine (PVC), which were administered over weeks or months.

The model constructed treatment cycles that reduced potency to a quarter or half of nearly all doses while maintaining the same tumor-shrinking potential. Doses were many times skipped altogether and scheduled only twice a year rather than monthly.


The model conducted about 20,000 trial-and-error test runs on each patient, learning parameters for optimal regimens to formulate new ones based on various constraints provided by the researchers.

It then was used on 50 other simulated patients, comparing the results to those of a conventional regimen of both TMZ and PVS. Upon receiving no dosage penalty, the model designed nearly identical treatment plans to those of human experts but with substantial cuts in dose frequency and potency.

Shah says such an approach is unorthodox due to its ability to weigh potential negative consequences of actions, in this case, doses, against an outcome like tumor reduction, creating a balance between reducing toxicity and maximizing shrinkage of malignancies.

“AI models' dosing regimens may be used to inform, empower and educate physicians,” Shah said. “We are working on regulatory approvals to move this and related work to clinical settings.”

In addition to single cohorts, the model was designed to treat patients individually, using medical data not often evaluated in traditional clinical trials, such as tumor size, medical histories, genetic profiles and biomarkers. The results were found to be similar.

Patients were randomly selected from a large database of glioblastoma patients who previously underwent traditional treatments. The model was adjusted so as not to put out a maximum number and potency of doses with regimens based on protocols that have been in clinical use for decades and are used by oncologists to predict doses to give patients based on weight.

Reinforced learning has previously been deployed to train driverless cars in maneuvering and was previously used to train the computer program, DeepMind, known for its 2016 victory over one of the best human gamers worldwide in the Japanese game go.

Back to HCB News

Rad Oncology Homepage