The aim of this project was to develop a system able to correctly classify and forecast 12 specific sleep positions (visible in the image above), using the data from two accelerometers LIS3DH placed one on the chest and one on the right ankle. The finished project includes:
- development of the wearable device
- data acquisition
- data processing
- GUI for data sampling and visualization
- machine learning architecture able to perform sleep position classification
The PSoC communicates with the LIS3DH sensors using I2C communication, and the accelerometers have a sampling frequency set at 10Hz. The communication between the PSoC device and the pc was created using a HC05 Bluetooth module to make the device wearable. Also, for wearability, the device is powered with a 9V battery and a subsequent 9 to 5 voltage transformer.
- 2 LIS3DH accelerometers
- 1 HC05 Bluetooth module
- 1 9V battery
- 1 9to5 voltage transformer
- 1 PSoC CY8CKIT-059
- 1 breadboard
- 2 3D printed casings were used to encapsulate the accelerometers and place them on the body
- 10 2m long cables to connect the accelerometers to the PSoC
- 2 modular velcro bands to fix the accelerometers in place
The links between the PSoC and the different components can be found in the dedicated folder in the form of the hardware setup documentation and the eagle schematic and board files. The casing's 3D model can also be found in the dedicated folder.
This section of code is responsible for I2C communication between the PSoC and the 2 accelerometers, which are sampling data with an ODR of 10Hz and are set on FIFO mode. The following code averages the 32 samples of the FIFO buffer from both accelerometers, each time they are simultaneously full. Finally, the 6 averaged acceleration values (x, y and z from the chest and the ankle) are combined in a unique string and managed by the UART of the bluetooth module, having a baudrate of 9600 bps.
All code files of this section have been individually commented on for clarity.
This section of code deals with establishing a connection between pc and bluetooth module, visualizing the data sampled in real-time, and producing csv files following a specific sampling protocol, which are further used to train the machine learning module. It follows a description of the most important subsections of code, but the code has been commented throughout for clarity.
This subsection is devoted to scanning all serial ports and creating a list used in the dropdown menu.
The SerialWorker class handles the connection with the bluetooth module by establishing a connection with the desired serial port (selectable in a drop down menu), reading data from this port from start (S) to end (E) tokens as soon as the Start button is pressed, storing iteratively the read lines in the data object and saving them at the end of the protocol in a single csv file.
It's also responsible for sending char data on the serial port and closing the serial port before closing the app.
This sub-section of code deals with creating the window used by all the plot widgets and the various buttons. It's worth noting that the 6 data charts are plotted on the same window by first removing the oldest point, then adding the new one, and finally updating the data.
The main is responsible for setting the window size, creating the thread handler and initializing all the functions previously mentioned.
The Machine Learning portion of this project was created using a Jupyter Notebook. The library used are available in the requirements.txt file inside the machine learning folder and any function needed is clearly listed and commented troughout the code.
In order to run the code, it's necessary to upload the csv files produced by us during the sampling. All the paths are stored in a distinct variable.
The data consists in 11 csv files sampled from 11 different people, it's worth nothing that the file 'fabio_20220601_193018.csv' was discarted due to errors during the sampling procedure, thus we ended up with 10 different people, a sample size big enough to guarantee inter-subject variabilty.
The data available was studied in three different ways, in order to see if the accuracy of the model could be improved:
- Raw sampled data
- Average of data for each sensor
- Average of data for each sensor + minimun and maximum values for each sensor
It's important to note that the first 20 samples of each position recording were removed, due to the patient still not being completely settled down and ready for the sampling. The data was also normalized as good practice dictates.
The data was visualized multiple times, this allowed us to purge it from artifacts and anomalies and improve the learning algorithm, such as the transient data between two positions mentioned in the previous paragraph.
The data was gathered from 2 accelerometers, one on the chest and one on the right ankle. Therefore we have 6 inputs + 1 input for the subject's sex.
- x_chest
- y_chest
- z_chest
- x_ankle
- y_ankle
- z_ankle
- sex
The variable that we need to predict is the sleeping position. It can be one of twelve positions.
To run the code, simply run sequentially every data cell. The code was tested on all three datasets, and the best performing one turned out to be the Random Forest Classifier learned using the average+min_max dataset, which reached 50% of accuracy on the test set.
In order to better assess the performance of the classifier we also tested it only on the four "master" positions: by grouping by three the positions (and therefore eliminating the small variations between each of the main positions) we obtained the main positions: supine, prone, left-sided and right-sided. When we tested our algorithm only with these targets, we easily obtained 100% accuracy.