Cardiopulmonary Resuscitation Assistant

A smartphone app for detecting the chest compression rate during CPR (dissertation research). Evaluates the chest compression rate from the bottom-up view by harnessing and thresholding the acceleration of pixels. Tested with 88% accuracy over ideal and non-ideal environments. Report PDF available in repository. Visualisations available at cpr-assistant.herokuapp.com.

Python

The solution was implemented in Python to prove the feasibility of the concept.

Installation

Requires Python >=3.5, pip3 and virtualenv to be installed on your machine.

1. Install virtualenv globally using pip3 with pip3 install virtualenv.
2. Create a virtualenv named "venv" with virtualenv venv.
3. Activate the virtualenv with . venv/bin/activate.
4. Install dependencies with pip3 install -r requirements.txt.

Datasets (Test Cases)

Code	Test Case
L	Subject has long, loose hair.
S	Subject has short hair.
B	Background disturbances wholly intended to disrupt performance.
C	Crop at neck-level (the subject’s neck-level and below is cropped out).

Graphs

Provided that the CSV files are in the correct directories, interactive graphs may be viewed using the following command: python3 solution.py -d=DATASET where DATASET can be one of L1, L2, L3, S1, S2, S3, S4, LB1, LB2, LB3, SB1, SC1, SCB1.

Ground Truth

• Ground truth data was extracted from front view videos, designated with the suffix GT (e.g. L1_GT.mp4) using the following command: python3 ground_truth_extractor.py -i=L1_GT.mp4 -v -o (warning: this overwrites the existing data used for graphing).
• This command calculates ground truth for dataset L1 from the video at ./videos/GT/L1_GT.mp4 and dumps the data in a CSV at directory ./csv_gt/L1_GT.mp4.csv.
• CSV files outputted by the script contain the following data per column: (X, Y, S) where X is the X-coordinate of the ball in the frame, Y the Y-coordinate of the ball in the frame, and S the state observed in the frame: S = { None, Compression, Breathing }. Each row denotes a frame.
• For more details on command-line arguments, run python3 ground_truth_extractor.py --help.

Solution

• Solution data was extracted from the bottom-up view videos, designated with the suffix BUV (e.g. L1_BUV.mp4) using the following command: python3 solution.py -d=L1 -r -o -w=./weights/S1.npy (warning: this overwrites the existing CSV used for graphs).
• This command extracts solution data for dataset L1 from the video at ./videos/BUV/L1_BUV.mp4 using preloaded weights at ./weights/S1.npy.
• The script solution.py contains parameters that may be configured accordingly: BREATHING_MODE, LEARNING_RATE, LOOKBACK_TIME, MINIMUM_ACCELERATION, MOVING_AVG_PERIOD, MAX_TIME_FOR_UPWARD_ACCELERATION, MIN_MOVEMENT_PCG, MIN_FLOW_THRESHOLD, SCALE, MIN_BREATHING_MOVEMENT.
• CSV files outputted by the script contain the following data per column: (vertical displacement, upward displacement sum, downward displacement sum, S, total percentage of pixels moved, vertical acceleration) where S is the state observed in the frame: S = { None, Compression, Breathing }. Each row denotes a frame.
• The resulting weights after each execution may be saved by providing the -s flag.
• For more details on command-line arguments, run python3 solution.py --help.

Android

An Android application was implemented with similar logic to the Python solution. This requires further device-specific optimisations.