Patients with metastatic cancer and their doctors often struggle with a conundrum: although some people live with their disease for months or years, others have much shorter lifespans. But it can be very challenging for physicians to accurately predict how long any one patient will live. This makes it difficult for doctors to recommend appropriate treatment options and for patients to set goals and priorities about their end-of-life care.

Now radiation oncologist Michael Gensheimer, MD, and his colleagues have designed a computer algorithm to dig through thousands of bits of medical information and pull out nuggets that can predict the lifespan of individual patients. They published their results in the Journal of the National Cancer Institute.

As Gensheimer explained to me in an email:

Patients with metastatic cancer cannot generally be cured of their disease, but there is a lot of variation in how long these patients may have to live. There are studies showing that doctors are not good at accurately predicting their patients' life expectancy and that they tend to over-estimate survival time. Being able to more accurately predict survival could help doctors and patients customize treatments to each patient's situation.

The researchers studied electronic health records, which include doctors' notes, lab results, patients' vital signs, and diagnosis codes, of more than 12,000 patients with metastatic cancer. They then used machine learning techniques to parse through the information and predict how long each patient was likely to live.

Interestingly, the most useful clues were found in the most time-honored method of data storage — the exam notes jotted by the doctors themselves.

As Gensheimer explained:

We used natural language processing methods so that the model could use data that were only present in providers' clinic notes. Some examples of these sorts of data are functional status (is the patient able to walk around, get dressed independently, and so on), and symptoms (shortness of breath, pain, etc.). We found that this note text information was actually the most helpful data source for predicting survival time, more helpful than coded data such as laboratory values.

The researchers are hoping to introduce their model in the clinic soon as part of a pilot study to learn if doctors find it useful and whether it influences how they care for their patients.

To use an example from my field of radiation oncology, a patient who is expected to live for a year or longer may benefit from high-dose radiation or a several-week course of treatment that may control a tumor for longer. But the side effects from this higher dose treatment might not be worthwhile for someone with a life expectancy of only a few months.

Photo by takomabibelot