Ordinary mobile phones can also measure blood oxygen

Turn on the flashlight of your mobile phone, press it with your finger, and measure blood oxygen saturation easily. The measured concentration range was further extended to 70%. You should know that 70% of the blood oxygen index is an important warning line. If it is lower than this value, it often means that hospitalization is required. At present, the monitoring range of smartwatches and mobile phones on the market is basically more than 80%, which will limit the judgment of the real health of the human body.

And the accuracy of the new method is not bad. In more than 10,000 experiments, the method could tell whether the tester was in low blood oxygen levels 80 percent of the time. At present, the research has been published in the Nature cooperative journal NPJ Digital Medicine, and the research data set has also been open-sourced.

Using Convolutional Neural Networks

The experiment can be roughly divided into two parts. First, a large amount of data is collected in this special way to train a deep-learning model. Second, use the trained model for testing. Let's look at the data collection part first. The researchers recruited 6 subjects to conduct the experiment. By having them inhale different concentrations of oxygen, their blood oxygen concentration levels are changed.

The corresponding author Jason S. Hoffman (Jason S. Hoffman) said that this is very different from the previous method of letting the subjects hold their breath to control the blood oxygen concentration. Data was collected for up to 15 minutes per subject at a time. Then use a smartphone and an ordinary oximeter to monitor the data at the same time.

Among them, ordinary pulse oximeters use transmissive PPG, and mobile phones use reflective PPG. PPG (photoplethysmography) is the most common non-invasive method for measuring blood oxygen concentration. It mainly uses light to irradiate human skin, and the absorption of light by subcutaneous arteries varies due to the different ratios of hemoglobin content. This light change can be further converted into an electrical signal.

For reflective PPG, when the incident light generated by the flash passes through the subcutaneous tissue, venules, and arterioles of the human body, after multiple scattering, a part of the light signal returns to the skin surface, which is to convert this part of the light signal into a current signal. After collecting the data, the researchers used an application to extract clips of more than 30 frames from the video. (In order to ensure that the recorded videos are preferably above 30 frames, the researchers also tied ice packs to the mobile phone to dissipate heat)

Then you can start training the neural network. Using the CNN machine learning model, they designed and trained a neural network consisting of 3 convolutional layers and 2 fully connected layers. After data preprocessing, the PPG signal of each channel can be extracted by calculating the average pixel value of each frame and then averaged. Train and evaluate with Leave-One-Out Cross-Validation (LOOCV), using 1 subject's data as the training set, 1 subject's data as a validation set, and then testing the model on another subject. The input to the model is a 3-second-long video, and the output is blood oxygen saturation.

The test results show that the model works best on subject 4, with a sensitivity of 88% and a specificity of 78%. Hypoxemia was accurately judged 88% of the time.

The dataset is open

Currently, the dataset for this study is freely available as open source. The researchers said that more data support is needed to accurately measure blood oxygen concentration through ordinary smartphones, and the current experimental results cannot be used for medical purposes. For example, in the experiment, it was found that the skin color of the subject and whether there are calluses on the hands may affect the accuracy of the test results.

With only 6 test subjects, the sample size is also very small, which may introduce experimental bias. Therefore, more people are needed to improve and enrich this type of data. The corresponding author and first author of the paper is Jason Hoffman (Jason Hoffman), who is now a Ph.D. student at the University of Washington, and his research direction is the intersection of medical and computer fields. He also had previous work experience in the hardware development department of Microsoft.