Machine learning experts working at Google Health have published a new study in tandem with the University of California San Francisco (UCSF)’s computational health sciences department that describes a machine learning model the researchers built that can anticipate normal physician drug prescribing patterns, using a patient’s electronic health records (EHR) as input. That’s useful because around 2 percent of patients who end up hospitalized are affected by preventable mistakes in medication prescriptions, some instances of which can even lead to death.
The researchers describe the system as working in a similar manner to automated, machine learning-based fraud detection tools that are commonly used by credit card companies to alert customers of possible fraudulent transactions: They essentially build a baseline of what’s normal consumer behavior based on past transactions, and then alert your bank’s fraud department or freeze access when they detect a behavior that is not in line with and individual’s baseline behavior.
Similarly, the model trained by Google and UCSF worked by identifying any prescriptions that “looked abnormal for the patient and their current situation.” That’s a much more challenging proposition in the case of prescription drugs, vs. consumer activity – because courses of medication, their interactions with one another, and the specific needs, sensitivities and conditions of any given patient all present an incredibly complex web to untangle.
To make it possible, the researchers used electronic health records from de-identified patient that include vital signs, lab results, prior medications and medical procedures, as well as diagnoses and changes over time. They paired this historical data with current state information, and came up with various models to attempt to output an accurate prediction of a course of prescription for a given patient.
Their best-performing model was accurate “three quarters of the time,” Google says, which means that it matched up with what a physician actually decided to prescribe in a large majority of cases. It was also even more accurate (93%) in terms of predicting at least one medication that would fall within a top ten list of a physician’s most likely medicine choices for a patient – even if its top choice didn’t match the doctor’s.
The researchers are quick to note that though the model thus far has been fairly accurate in predicting a normal course of prescription, that doesn’t mean it’s able to successfully detect deviations from that yet with any high degree of accuracy. Still, it’s a good first step upon which to build that kind of flagging system.