Developed technology to predict "what next will happen" to Google's going to the patient who visited the hospital


Google reads the data of the patient's electronic medical records and predicts "what will happen next?" To the body of the patient who visited the hospital GoogleNature Partner Journals: Digital MedicineIt is announced in. Deep learning is used for this prediction model, and it is possible to predict "death period of the patient" "long-term hospitalization" "to re-admit" etc with high accuracy.

Scalable and accurate deep learning with electronic health records | npj Digital Medicine

Google AI Blog: Deep Learning for Electronic Health Records

When my health is bad and I go to the hospital, in our head "When will you be able to return home?" "When will you get better?" "You must also come to the hospital Wondering if there are many questions, such as? If these questions can be answered correctly, doctors and nurses can take safe and prompt action. If "what next happens to the patient?" Is predictable, it is also possible for the doctors to get ahead of them before the condition gets worse.

In order to enable such technology, Google collaborated with the University of California San Francisco, Stanford University Medical School and the University of Chicago Hospital.

Researchers focused on machine learning of the same type as predicting traffic conditions during commuting. In order to predict the patient's condition with this machine learning,ScalabilityIt was necessary to clear two points of "accuracy" and "accuracy".

◆ Scalability
"Scalability" means that predictions must be made easily for any hospital system, or for any outcome. In many cases, healthcare data is complex and results are conflicting.

The Electronic Health Record (EHR) is very complicated and even one body temperature has a different meaning depending on whether it was measured by the ear or measured in the mouth or measured under the armpit . In addition, because each hospital customizes EHR in different ways, even two patients receiving similar care, if they are different hospitals will look like completely different data. Therefore, before applying machine learning, researchers can record patient recordsFHIRIt seems that it was necessary to reexpress it in a consistent manner.


Once the formats are complete, the deep learning model reads all the data elements from the initial data to the most recent data and learns which data will help you predict the results. Thousands of data elements are included, but what human beings need to do isRecursive neural networkWhenFeed forward neural networkIt is only to develop some deep learning models based on the.

◆ Accuracy
The presented predictions should inform clinicians of the problem, but they should not be "false alarms". In this regard, it is expected that more accurate prediction models will be created by widening the electronic medical records and using more data.

To measure the accuracy of whether Google's model can distinguish between "a person expected to have a specific result" and "a person who is not so", "Recipient operation characteristics(ROC) "criteria was used. In ROC, 1.00 is "perfect" and 0.5 is "coincidental" whereas the accuracy of the prediction made by Google's model for "whether the patient is hospitalized for a long time" was "0.86". This exceeds the value of "0.76" of the conventional method. The accuracy of the prediction of the death of the patient is "0.95", it exceeds the value of "0.86" of the conventional method, the accuracy of prediction of re-hospitalization is "0.77" and the conventional method "0.7" Since it exceeded the accuracy of accuracy, we can predict "what will happen to the patient next time" with considerably high accuracy.

Also, Google's predictive model predicts what the patient is being treated for, based on what formulation the patient is prescribed from the doctor. Although this does not diagnose the patient's illness, it is meaningful to record what treatment the attending physician is doing.

In addition, Google's forecasting model is considered useful as it indicates important data to doctors to show "important data elements in predicting".

In addition, although Google uses a deep learning model based on electronic charts received from partners to create a prediction model, when Google receives an electronic medical record, it states that personal information was removed and it was non-specified It has been. The data is protected with state-of-the-art security, access restriction and encryption were done.

in Software, Posted by logq_fa