Smart chair for the elderly

in Liketu4 months ago (edited)



405opf2al2g8e71p_IMG_7203.jpg
405opf19l2g8ecah_278981623_745527123485941_2650728297765458670_n.jpg

Hi all,

today I come to you with the progress of my work regarding the checking of non-normative states in elderly people using a specialized chair equipped with a capacitive electrode.

As a reminder for those who have not yet read what all the work is about. My task is to create a neural network model that will be able to take the ECG signal from the e-chair in real-time and then perform a classification of the activity performed. I currently selected three classes, quiet sitting, talking, and coughing. The idea itself came about to help elderly people who rarely want to wear wristbands and spend most of their time in a chair in front of the TV.

I finally got the key to the room and therefore managed to make the first measurements for the database. As it turned out, there are only 20 seconds of measurement time, so the work is a bit tiring. However, I managed to make 90 measurements for each class. Interestingly, the electrodes in the chair fit only men, so I was not able to perform measurements on women, and another thing that could have affected the interference in women could have been the metal clasps in their bras.

Unfortunately today the equipment stopped working and after two hours I gave up on fixing it. I hope it was just a battery failure because there was interference from the mains, but anything could be possible.

In addition to the measurements themselves, the next thing is to update the neural network model. Currently, I have chosen 2D CNN Sequinetial from Keras, which uses CNN, MaxPooling, Flatten, and other layers. In addition, I selected softmax activation functions, something that is usually selected. With these settings, I get about 90% accuracy the only thing that puzzles me is the fact that I get quite large validation loss values, which may be due to the small database, but here still need some time to see if I am right. The next thing is to create a transfer learning that will allow me to retrain the model quite easily once I have a basic database.

I also have a question for everyone who has read this far. What class of activity could I still come up with, I was thinking of holding my breath, screaming, and taking measurements after a big effort. So far I had to skip the yelling due to too much noise that would disturb other staff and students.

That's all for now, thanks a lot for reading!


For the best experience view this post on Liketu