Google predicts chance of death with accuracy
AI system could boost patient care
GOOGLE knows everything (or at least it feels that way), and now it can even tell you when you’ll die. The tech giant helped test an artificial intelligence computer system that can predict whether hospital patients will die 24 hours after admission. What’s more staggering is trials put the accuracy of the AI’s predictions as high as 95 per cent. It works by chewing up data about patients – their age, ethnicity and gender. This information is then joined up with hospital information – prior diagnoses, current vital signs and any lab results. What makes the system particularly accurate is that it’s fed data typically out of reach for machines, such as doctors’ notes buried away on charts or in PDFs. Artificial intelligence systems become smarter over time through a process known as machine learning. The AI was developed by researchers from Stanford, the University of Chicago and University of California San Francisco. Google then took the AI system and “taught” it using de-identified data of 216,221 adults from two US medical centres. This meant the AI had more than 46 billion data points to vacuum up. Over time, the AI was able to associate certain words with an outcome (life or death), and understand how likely (or unlikely) someone was to die. What’s particularly exciting about Google’s system is researchers can throw almost any type of data at it. Stanford professor Nigam Shah told Bloomberg about 80 per cent of development time spent on predictive models went toward making the data presentable for the AI. But Google’s system can chew up anything and make predictions based on it through its powerful machine learning abilities. The system can estimate the length of a patient’s hospital stay and chance of being readmitted. So how accurate is the AI? When we talk about probability, a 1.00 score is perfectly accurate. And a 0.50 score is a 50/ 50 chance – an AI that scores 0.50 is no better than a human making random guesses. Here’s how Google’s AI fared on various outcomes: Predicting whether a patient would stay long in a hospital – 0.86 (Google) v 0.76 (traditional methods) Predicting inpatient mortality – 0.95 (Google) v 0.86 (traditional methods) Predicting unexpected readmissions after a patient was discharged – 0.77 (Google) v 0.70 (traditional methods). “These models outperformed traditional, clinically-used predictive models in all cases,” Google’s Alvin Rajkomar said. Professor Rajkomar said hospitals adopting the AI could use it to “improve care” for patients.
Photo:iStock HAVE YOU GOOGLED IT? An artificial intelligence system, tested by Google, can predict patients’ health outcomes.