Researchers at MIT are now exploring “machine-learning strategies” as a means of regulating dosages of chemotherapy and radiotherapy administered to cancer patients, writes Rob Matheson in a recently published article for MIT News.
The MIT Media Lab research team is focusing on glioblastoma, an aggressive form of cancer. The glioblastoma tumor appears in the central nervous system (either in the brain or spinal cord), and the prognosis for diagnosed adults is “no more than five years,” Matheson writes.
To treat glioblastoma, medical professionals administer the “maximum safe drug doses,” but the side-effects can be severe.
The research team, lead by Gregory Yauney, adopted a technique called reinforced learning (RL), loosely inspired by behavioral psychology, to test treatments that administer the drugs temozolomide (TMZ), procarbazine, lomustine, and vincristine (PVC) to patients with glioblastoma.
The artificially intelligent “agent,” Matheson writes, is tasked with completing an “action.” In this case, that action is whether to withhold or administer a drug dose. The “agent” scans available data, which is comprised of traditional regimens by which medical professionals administer dosages set by practice and trials as the clinical standard. For each point at which an “action” is required, the “agent” determines whether to administer or withhold a drug dose. If the “agent” determines to administer a dose of the drug, it then decides whether to administer the full dose or partial dose.
If the “action” the “agent” chooses is expected to shrink the diameter of the tumor (based on a clinical model), then the “agent” receives a “reward,” presented as a positive number, like +1.
But this is an “unorthodox” RL model, states Pratik Shah, who supervised the research, because the agent also weighs potential negative consequences of its “actions.” If the agent administers doses of maximum potency at all intervals an “action” is required, it receives a “penalty,” presented as a negative number, like -1.
According to clinical design expert Nicholas J. Schork, this new method of administering drugs is a “major improvement,” states Matheson. As Schork tells Matheson: “[Humans don’t] have the in-depth perception that a machine looking at tons of data has, so the human process is slow, tedious, and inexact.” Machines, on the other hand, are able to sift through data and identify patterns both quickly and effectively.
The MIT Media Lab team will be presenting their research at the 2018 Machine Learning for Healthcare Conference at Stanford University this week.
Read the full article at MIT News.