RouseAI

Stacked LSTM Architecture with illustrated gates
EEG Electrode placement on the scalp. 36 electrodes were recorded for the dataset we used.
EEG signal(microvolts) from a single control trial over time
Stages of sleep from N1 to N4

Introduction

Driving alone while tired has few preventable solutions and often causes multi-person crashes and potential death. RouseAI provides warnings to drivers in advance and ensures safety on the roads.

What it does

Drowsy driving can be attributed to around 10% of car crashes, causing over 100,000 crashes and 1550 fatalities annually. In Canada, around 21% of drivers admit to falling asleep at the wheel, putting others on the road at risk and making driving at night, or even in the early hours of the morning exponentially more dangerous. The goal of RouseAI is to use the EEG signals of the brain in order to accurately predict and alert drivers of their next microsleep cycle and take preventative measures.

A specific benefit of the system affects trucking which occupies a large industry in Canada, employs many people working overnight(, and involves icy or dangerous conditions. Hence, being mentally and physically awake is best possible solution, but many truck drivers often don’t have the choice. Driving on rural roads with no places to sleep overnight makes it difficult to ensure that all drivers are at the right attention level to stay on our roads. By predicting the next microsleep cycle, our app will be able to recommend nearby gas stations/truck stops for them to rest so they can safely reach their destination.

How we built it

-Built with an unsupervised learning model using LSTM and preprocessed with Tensorflow.

-The data is specifically electroencephalography data from drivers to measure current microsleep cycles and use this data to predict future cycles. From this data collected, we can extrapolate and determine the next expected microsleep cycle and take preventative measures to ensure that the driver does not end up falling asleep and also recommend places where the driver can rest or take a break.

-Conducts the challenging task of univariate, multistep time series forecasting** which is a part of unsupervised learning.

Scientific Background of EEG

EEG measures post-synaptic potentials from large groups of neurons which trigger during an event. Non-invasive EEGs consist of specially placed electrodes(see Diagram 1) placed on the scalp to measure brain activity and find the difference between signals. During periods of drowsiness, certain regions of the brain such as the hypothalamus will activate and send alpha, beta, gamma, or theta waves. This allows differences in EEG recordings to indicate sleepiness vs wakefulness.

Challenges we ran into

-One of the key assumptions to make predictions into the future is the persistence model of thinking which requires the model to hold a pattern of prediction which led to less optimal predictions.

-Dealing with temporal data was challenging

-Overfitting and regularization with dropout

-Working with large epochs caused longer model-training times under the hackathon time constraint

Accomplishments that we're proud of

-Finishing the stacked LSTM model for unsupervised learning

-Achieving similar or potentially useful pattern fitting and general trends

- Demonstrating valid proof of concept

What we learned

We learned about difficulties with LSTM as well as dealing with aggregate data sources
We learned about frequencies of EEG, spectrograms, and how EEG detects event potentials