‘Unorthodox’ neural network generates optimal brain cancer treatment plans
Image credit: Dreamstime
Researchers based at Massachusetts Institute of Technology have created an artificial neural network capable of determining the minimum doses of toxic chemotherapy and radiotherapy treatments that will still prove effective.
The team focused their efforts on the most aggressive type of brain cancer, glioblastoma, which can also be found in the spinal cord, and often kills its host within a year of diagnosis. Standard treatment involves surgery followed by months of chemotherapy (often using the oral drug temozolomide) and radiotherapy. Temozolomide comes with a range of unpleasant and common side effects, including bone marrow suppression, nausea and vomiting, as well as having damaging effects if taken during pregnancy or breastfeeding.
Despite unpleasant side effects patients must suffer through, clinicians tend to apply the maximum safe doses in order to shrink the tumours as much as possible.
Now, researchers based at the MIT Media Lab have suggested a machine learning model which could continue to offer effective dosing while minimising toxic side effects. This artificial neural network is presented with treatment plans currently in use – which apply temozolomide, procarbazine, lomustine and vincristine – and adjusts the monthly doses iteration after iteration until it settles upon an optimal treatment plan balancing efficacy with tolerability.
This makes use of a performance-focused machine learning technique called reinforced learning, in which a network learns that certain actions lead to improved outcomes, and iteratively adjusts its actions to achieve better outcomes. Reinforcement learning has a wide range of practical applications, and was notably used to train the artificially intelligent DeepMind program to play Go; in 2016 the program beat the world Go champion.
In order to ensure that the neural network did not simply propose maximum doses – which would offer the best outcome in terms of shrinking the tumour, but cause many agonising side effects – the researchers penalised the network for suggesting treatments involving maximum doses. This resulted in an “unorthodox” reinforcement learning model which balances potential negative consequences of its actions against the desired outcome (shrinking the tumour).
“If all we want to do is reduce the mean tumour diameter, and let it take whatever actions it wants, it will administer drugs irresponsibly,” said Dr Pratik Shah, who supervised the study. “Instead, we said: we need to reduce the harmful actions it takes to get to that outcome.”
In a simulated trial of 50 patients, the model’s suggested treatments reduced the potency of nearly all the doses to a quarter or half, while maintaining the same tumour-shrinking potential. In some cases, it skipped doses entirely, at times scheduling doses every six months rather than every month. When the network was not penalised for suggesting maximum doses, however, it offered treatment suggestions almost identical to those of human experts.
“We kept the goal, where we have to help patients by reducing tumour sizes but, at the same time, we want to make sure quality of life- the dosing toxicity – doesn’t lead to overwhelming sickness and harmful side effects,” said Shah.
Considerable excitement – as well as trepidation – surrounds the possibility of rolling out artificial intelligence in healthcare, both in research (such as to identify causes of and treatments for diseases) and in managing treatment. It is thought by many experts that machine learning could become a useful tool in cancer diagnosis and treatment, and earlier this year, Prime Minster Theresa May called for machine learning to be integrated with cancer treatment in order to prevent thousands of deaths every year.