Google is working on deep learning models that can predict, among other things, the risk of a patient dying with great accuracy. The models analyse a patient's entire medical record and can therefore make much more accurate predictions than models that have been available for some time.
This is according to a research paper published in Nature. Google researchers point out that predictive statistical models of patients are expected to play an important role in personalising medication and improving the quality of care. Such models are traditionally based on information from patients' medical records. DThese models are created using a manual selection of the most relevant data from the medical record. This method is not only time-consuming, but also ensures that much information from the medical record is not included in the model.
Analyse unfiltered data
Google is working on deep learning models capable of analysing all unfiltered data from a medical record at lightning speed and making predictions based on this data. Any handwritten notes from doctors are also included in the analyses. As a result, the deep learning model has much more information than the manually created static model and can therefore make more accurate predictions.
The deep learning models can be used to make a variety of medical predictions about a patient. These include the outcome of a specific treatment, how long a patient will need to stay in hospital and at what time they can be discharged from hospital. Deep learning models can also predict the risk of a patient dying.
Tested in two hospitals
The researchers validated their deep learning models using datasets made available by the University of California, San Francisco (UCSF) and the University of Chicago Medicine (UCM). These datasets included medical data from 216,221 adult patients who were hospitalised for a period of at least 24 hours.
During tests, the deep learning models have been used to predict, among other things:
- whether the patient would die during their hospital stay
- whether the patient would be readmitted within 30 days of discharge from hospital
- with which final diagnosis the patient would leave the hospital
In all cases, the deep learning models were found to provide more accurate predictions that traditional real-world prediction models. In addition, the number of false positives was found to halve if the deep learning models were deployed.
Author: Wouter Hoeffnagel
Source: Research paper 'Scalable and accurate deep learning with electronic health records'
Source photo: Pixabay / StockSnap
