“Doctor, how much time do I have left?” is one of the most difficult but inevitable questions that the doctors of the seriously ill face. Yet, giving an accurate answer is critical to help the person prepare for possible death. Recently, a team of researchers from Stanford University has designed an AI tool to help doctors do that, and it turns out it is surprisingly accurate. A preprint of their paper can be found here.
Palliative, or end-of-life care, is offered to seriously ill patients with a view of improving the quality of their remaining life. It focuses on comfort and managing symptoms rather than finding a cure. The trick, though, is to identify those patients who could benefit from this type of care, on time.
Palliative Care Challenges
In an interview with Gizmodo, Anand Avati (the lead author of the study) said: “The problem we address is that only a small fraction of patients who can benefit from palliative care actually receive it—partly due to being identified too late, and partly due to shortage of [human resources] in palliative care services to proactively identify them early on.”
The best time to offer palliative care is when the patient is within three to 12 months of possible death. Less than that is simply not enough for the proper preparations, while more than 12 months could place an unnecessary burden on the limited resources which palliative care teams typically have.
The problem is, doctors are often wrong in their prognoses. In fact, a study published in the British Medical Journal in 2003 found that physicians regularly overestimated the survival time of their terminally ill patients. Thus, many patients who could benefit from palliative care are not receiving it.
The new tool developed by the researchers is an algorithm capable of predicting how likely a patient is to die within this target time span. It uses information like the severity of the disease, the medicines prescribed, and the number of days spent in the hospital to make its prediction. The patient’s electronic health record is fed into the algorithm, which then assigns and adjusts the weights given to the different factors, and calculates the probability that the patient will die within 3-12 months.
How It Works
The new algorithm is a deep neural network, a method for machine learning called that way because it is inspired in the neuron networks of the brain. In this type of machine learning, the computer learns by analyzing examples. For instance, to learn to identify cars, an algorithm would analyze “training” images of cars, labeled beforehand into “car” or “no car.” The machine would then recognize features in the images that correspond to the concept “car.” By analyzing many images, the algorithm can be refined to astonishing accuracy.
To develop this particular algorithm, the researchers used the electronic health records of past patients. They sifted through the records of 2 million patients treated either at the Stanford Hospital or the Lucile Packard Children’s Hospital from 1995 to 2014. This included patients of different ages, and affected by different types of diseases. They identified 200,000 suitable records, and used 160,000 of them to train the algorithm.
The remaining 40,000 records were used to test the accuracy of the new tool, and the results were surprising. In 9 out of 10 cases it was correct when predicting a patient would die within 3 to 12 months. Additionally, 95% of patients for whom the algorithm gave a low probability of death lived longer than 12 months.
Problem Areas
As promising as this new technology sounds, there are some shortcomings to be aware of. Exactly how neuron networks work is a bit of a mystery. They are often referred to as a black box, because it is difficult to explain their inner workings. In other words, this new tool is able to give a fairly accurate prediction that a specific patient will die within 3-12 months, but it does not tell doctors why.
However, it is important to understand how these algorithms work, especially in applications concerning human health, to be able to recognize how and why they might fail. In the end, the program is not meant to replace a trained, reasoning physician but rather to be used as a tool to improve performance.
Nigam Shah, one of the co-authors of the study told Gizmodo: “We believe that a black-box model can lead physicians to good decisions but only if they keep human intelligence in the loop.”
Fortunately, there might be some things doctors can learn from it, after all. Siddhartha Mukherjee points out in an article for The New York Times that looking into the calculations for individual patients can generate some interesting insights, pointing to factors doctors might not have considered as predictors for demise.
The researchers indicate that in the future, they would like to broaden the information fed to the algorithm. Particularly, they discovered that some of the information used to train the program would not be available to be used for present-day patients. Additionally, they also plan on expanding the patient database used to train it, since at present the information came from only two hospitals.