Successfully identifying cardiac arrhythmia with Apple Watch and machine learning
Apple WatchTreatment of cancer patientsThere are hospitals that are helping, but before using a doctor 's diagnosis using Apple WatcharrhythmiaHave been discovered and methods to prevent diseases such as heart disease and stroke are being studied.
Can We Detect Atrial Fibrillation using Apple Watch Sensor Data
A kind of arrhythmia "Atrial fibrillation"Has the largest number of patients among arrhythmias, in the United States there are 2.7 million atrial fibrillation patients and it is estimated that about 7000 to 140 million people worldwide suffer from symptoms of atrial fibrillation I will. Atrial fibrillation is also one of causes of stroke, and if atrial fibrillation becomes serious enough to cause a stroke, there is also data that half of patients die within a year. However, since atrial fibrillation is capable of medication treatment, early detection can reduce the risk of stroke by as much as 75%.
Yancheng Liu of data scientist looked at wearable devices such as Apple Watch and Android Wear. Since heart rate data can be obtained in a larger amount than the past data of electrocardiogram, it seems that it was conceived to utilize the measured value of the heart rate collected by the wearable terminal in the medical field.
Mr. Yancheng started research on a method of predicting heart disease by analyzing heart rate measurements. First of all, at the University of California San FranciscoHealth eHeart StudyIn collaboration with the laboratory, we collected 13.5 million data from wearable terminals. One hundred thousand of them were from patients with atrial fibrillation.
The following data is rough data before going through machine learning.
After organizing the time when data was acquired or dividing healthy people and patients with atrial fibrillation, the graph looks like the following, and the patterns of different heart rate in healthy people and patients with atrial fibrillation I see that there is.
From previous studies, healthy people show peaks in electrocardiogram graphs at 0.003 to -0.04 Hz and 0.04 to 0.15 Hz, whereas peaks in atrial fibrillation patients frequently occur like noise I know that. Yancheng thought that he thought he could distinguish between a healthy person and a person with a possibility of atrial fibrillation by analyzing the data and finding the peak of the heart rate.
However, because the proportion of healthy people collected was high, the data of 41 patients with atrial fibrillation, 185 people with health and daily exercise habits, 274 people who are not exercising in particular, It seems that we decided to extract data for a total of 500 people and use it as a sample.
Yancheng analyzed data accuracy, precision and recall using machine learning techniques such as Extremely Randomized Trees (ERT) and Random Forest (RF). The table below compares the analysis results of the four machine learning techniques. According to Yancheng, machine learning called K-nearest neighbors (kNN) was the fastest in analysis speed, but ERT on the left side of the table can identify patients with atrial fibrillation with a probability of 50% Successfully found 86% of people who could develop atrial fibrillation. It seems that ERT is found to be one step excelled at high conformity rate and recall rate.
Patient identified as atrial fibrillation with ERT seems to have used only Apple Watch, Yancheng said, "Data gathered using sensor of wearable terminal is great shock to medical world by using machine learning technique And it will help the doctor to find the patient's atrial fibrillation. "