Google knows a lot about you. A lot more than you'd probably be comfortable with.
Their neverending quest for more knowledge has taken a slightly creepy (but medically speaking, quite useful) turn, outlined in a study recently published in npj Digital Medicine. The study involves new Artificial Intelligence (AI) that Google's Medical Brain team have been working on. It has been trained to predict how likely it is that patients entering hospital will make it out alive.
A trial of the machine-learning algorithm has shown that it can predict the likelihood of death with 95 percent accuracy, which is much better than the early warning score system currently used in hospitals.
In one instance reported in the study, a patient with late-stage breast cancer was admitted to hospital. Her lungs filled with fluid, she was seen by several doctors and underwent a scan. According to the hospital's assessment, she had a 9.3 percent chance of dying during her stay, based on her vital signs such as respiratory rate, blood pressure, and pulse.
Google's AI ran its own assessment on the same patient, assessing 175,639 data points on her record, the researchers wrote in their study. These included data points that aren't normally considered during patient evaluations. The AI was able to access previously out of reach data, such as PDFs of notes made by doctors and nurses that indicated evidence of malignant pleural effusions (fluid build-up around the lungs) and potential risk of pressure ulcers.
Looking at this data, the AI put the patient's risk of death during her stay at 19.9 percent. She died 10 days after admission.
Because Google's AI took more into account than the hospital's usual system of assessment, it was able to make a more accurate prediction as a result.
Overall, the study found that the AI was able to predict mortality 24 hours after admission with 95 percent accuracy at one of the hospitals trialed, and 93 percent at the other. This was significantly better than the hospital's traditional predictive model (the augmented Early Warning Score), which predicted mortality with 85 and 86 percent accuracy respectively.
The accuracy of the predictions was put down to the extra data that the AI was able to crunch. Normally when predicting patient outcomes, the time-consuming part is putting all the data together into a readable format, Nigam Shah, an associate professor at Stanford University, told Bloomberg.
"In general, prior work has focused on a subset of features available in the EHR [electronic health record], rather than on all data available in an EHR," the authors wrote in their study. "Which includes clinical free-text notes, as well as large amounts of structured and semi-structured data."
Essentially, Google's AI system copes well with lots of data that hasn't necessarily been put together in a structured way. It creates more accurate predictions with less grunt work from humans.