Can a computer predict cancer outcomes?

Cancer is a serious illness that, unfortunately, is not as simple to treat as most conditions one commonly seeks medical help for. Scientists have yet to discover an effective cure, which means that currently doctors and unfortunate patients have to rely on complicated and often lengthy treatments, in which every day counts. This is why an accurate prognosis at the very start is extremely important in deciding on the course of action both for the doctor and the patient. In some cases, having a good prediction of the outcome may help choosing a more effective treatment; in other cases, where the chances of survival are slim, a less aggressive treatment or otherwise no treatment at all can be opted for, in order to make the last days of a terminal patient as easy as possible. So what if computers could make such prognosis?

At the moment, the most common approach to predicting survival chances for oncology patients rely on a doctor’s educated guess after looking at various test results. The accuracy of such guess would depend on said doctor’s methodology and experience, which also means that two different doctors might disagree on how likely a patient would survive. But what if this decision making could be delegated to a computer? This is the question scientists from Finland and Sweden tried to answer in a recent study looking to find out whether machine learning could be used to accurately predict breast cancer outcomes.

In the study, the researches tried to create a smart model – consider it just a computer program – that would take in only an image of tissue from a tumour and make a prediction for the patient by assigning them a low or high risk of mortality. To train the model, the scientists used historical data available on numerous breast cancer cases throughout Finland.

The first layer of the model was trained to analyse the cancer tissue images and identify their noteworthy features – the kind a human doctor would look for with their eyes to make their own prediction. Then, said features would be converted into a digital form that is easier to work with for a computer algorithm. The second layer was trained to take data from the first and use it to make a prediction of low or high risk of mortality. Since the historical data used for training contained the actual survival outcome for each case, it was possible to give the learning algorithm a set of inputs and a set of expected outputs to create a model that can somehow make sense of connections between the two. After training was completed, the model could take in previously-unseen images of tumour tissue and make its own educated guesses.

The next challenge was to verify exactly how accurate model’s guesses were. To tackle that problem, the researches assembled a panel of oncology experts to compete with the model. A new set of breast cancer cases was pulled from the historical data; the set would not contain cases that the model had seen before. Both the experts and the model would give their survival prediction for each case and then the predictions would be challenged against the actual outcomes. It turned out that the model would give a correct prediction for 60% of the cases, while the experts would hit bull’s eye for 58% of them – a tight win for the model.

While the scientists state that their study has important limitations and doesn’t analyse all variables available to doctors in real life, they think that their approach to outcome prediction has real potential and is worth looking into and developing further. They also discovered that in some cases their model would give accurate predictions based on factors unknown to science – this means that sometimes a machine learning model can see more than a human doctor, which only strengthens the potential of this technology. Who knows, maybe some day all doctors will consult with smart models before giving a prognosis?

Anton Matveev

Turkki R, Byckhov D, Lundin M, et al. Breast cancer outcome prediction with tumour tissue images and machine learning. Breast Cancer Res Treat. 2019;177(1):41-52. doi:10.1007/s10549-019-05281-1 https://pubmed.ncbi.nlm.nih.gov/31119567/

One Reply to “Can a computer predict cancer outcomes?”

  1. Anton – this is so cool and seems like such a smart way to aid in decision-making – I would imagine that anything that can minimize the subjective nature of such decisions would be seen as a plus.
    -Edie

Leave a Reply

Your email address will not be published. Required fields are marked *