Falls are the leading cause of injury death in adults aged 65 and over, causing about 38,000 deaths and almost 3 million hospital visits in 2021 (for Disease Control & Prevention, 2024). Through the synthesis of current deep learning techniques and gait (walking patterns) based data, our team proposes a new and viable solution to predicting and preventing elderly falls. Gait data surrounding elderly people can be recorded using gyroscopes and accelerometers located in cell phones or other portable technologies. While mobile devices make this data easily accessible, it is often complex, containing time series data and multiple sensors. Neural networks can easily extract and recognize patterns, making it a good tool for analyzing the complex sensor data and returning an accurate reading. Additionally, a neural network’s ability to learn over time allows it to scale well with large amounts of data that other models may struggle with. To accomplish our goal, the team will combine the efficiency of a CNN with the accuracy of an LSTM to create our own hybrid model. The model will receive input from sensor readings (off of the user’s phone) and output information that can be used as data for future research as well as issue a notification to nearby emergency responders. The use of a neural network in this way should significantly alleviate the pressure on both the patient and the emergency responders and help fall victims to receive immediate and effective care.
Our model has been trained on the SisFall Dataset. It contains accelerometer and gyroscope sensor data of 19 different types of ADL and 15 different types of Fall, performed by young and old adults. The data has been recorded using the ADXL345 and MMA8451Q accelerometers along with ITG3200 gyroscope. It has been sampled at 200 Hz and is stored in the form of text files containing the x, y and z directional components of all sensors. The procedure given below has been utilized to clean, process and extract all training and testing data. For more information, kindly access: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5298771/
To access the dataset, we would like to request the reader to kindly reach out to the original authors directly.
Team collected our own data using the Sensor Logger on the iPhone which was used for model selection.
Testing was done using the MobiFall and MobiAct Datasets, containing different types of FALLs and ADL samples recorded using a Samsung S3 phone. For more information, kindly access: https://bmi.hmu.gr/the-mobifall-and-mobiact-datasets-2/
Kindly Check Final Report in the reports folder for an extensive description of the Data Processing Pipeline utilized.
To establish a baseline to compare our model, the team utilized a Support Vector Machine (SVM) with a linear kernel (implemented in SVM.ipynb) to classify our data. The SVM finds a hyper plane that maximizes the margin between the two nearest points of each label (in binary classification), fits the data, and then creates a decision boundary to classify based on the chosen hyperplane. The model was able to receive an overall testing accuracy of 99.72- percent (0.28-percent error) on the remaining data by correctly classifying 718 out of 720 ADLs and 361 out of 362 falls on the SisFall dataset's test split.
The fall detection system deploys a hybrid CNN-LSTM model that leverages the strength of both convolutional neural networks and long short-term memory networks, and is designed to efficiently process and analyze complex temporal sensor data. After thorough data preprocessing (segmenting raw inertial sensor data into fixed-length sequences and applying normalization and noise reduction techniques), the model accepts a 3D tensor with dimensions (sample number, time step, and feature type). This data first gets passed through three convolutional layers that each increase in complexity to detect increasingly intricate features: first layer has 64 filters, the second layer has 128 filters, and the final layer has 256 filters. Each layer is followed by a ReLU activation function and max pooling (pool size of 2). The output from the CNN then gets passed into two LSTM layers - each with 100 units. After this, the output is sent to a dense layer with 100 units and ReLU activation, a dropout layer with rate of 0.5, and the output layer is 1 unit with a sigmoid activation function to produce the accurate output.
The model achieved a 99.2% training, 99.1% validation and 99.2% test accuracy, highlighting the model’s effectiveness in classification between Fall and ADLs on SisFall Dataset. The model also achieves a Sensitivity of 98.3% and Specificity of 99.6% across the entire SisFall Dataset.
The proposed model performed most effectively, achieving an accuracy of 80.5%. Sensitivity of Phone Data stood at 77% whereas Specificity was 84.4% across the entire collected data.
The overall metrics from the MobiFall data set are as follows: accuracy 92.7%, specificity 91.8%, sensitivity of 97.3%.
We would like to thank the authors of the MobiFall and SisFall Dataset for allowing us access to do this project.