Recently, the US Patent & Trademark Office has published a patent application from Apple labeled “Interpretable Neural Networks for Cuffless Blood Pressure Estimation.”
Studies show that 1 in 3 deaths in the United States is caused due to heart disease and stroke. It is observed that high blood pressure (HBP) is associated with specific cardiovascular disease (CVD) risk factors which vary in proportion from one environment to another. Yet, high blood pressure generally has no warning signs or symptoms, and therefore many people do not know if they have HBP. Therefore, measuring your blood pressure is essential to keep your health in check. Several methods and instruments, such as blood pressure cuffs, are available to measure blood pressure.
In an interview with CNBC, Apple’s CEO stated that they aim to empower individuals to manage their health. He quotes that in the future, one can claim that Apple’s most significant contribution to mankind has been to health.
Earlier patents include simple wrist-based solutions, including Apple Watch and other inflatable cuffs. Apple’s ongoing project centers around using neural networks for taking blood pressure.
Artificial intelligence is widely adopted in various domains, including medical diagnosis, image processing, speech recognition, and many more. Artificial neural networks include multiple layers and filters. They can be trained by providing large training datasets to get the desired output on the new input data.
Apple’s invention uses an individually-pruned neural network that accepts a Seismocardiogram (SMG) as input to determine systolic and diastolic blood pressures.
Seismocardiogram (SCG) measures the micro-vibrations produced by the heart contraction and blood ejection into the vascular tree. The data collected on the ability of SCG can be used to reflect on cardiac mechanical events, including opening and closure of mitral and aortic valves, atrial filling, and point of maximal aortic blood ejection.
Some examples show that:
- A baseline model can be constructed by training the model with SMG data and blood pressure measurement from multiple subjects. One or more filters can be ranked by separability to prune the model for each unseen user that uses the model after that.
- A user can provide a set of SMG data and blood pressure measurements to the baseline model. The model with an increasing number of filters (ranked by separability) can evaluate the mean absolute error of the predicted blood pressure for multiple runs.
- Including low-separability filters in the model can decrease the accuracy of the model. Therefore, the model can be pruned to include the optimal number of filters ranked by separability for each individual. In some examples, the individual can use the pruned model to calculate blood pressure using SMG data without corresponding blood pressure measurements.