Co-author: Khin Radanar Pyae Phyo 

  Nearly 150,000 people die each day around the world, according to 2017 data. It is important to understand what is meant by the cause of death and the risk factor associated with a premature death. The following survey is the measure of the cause of death among the age between 15 to 49.

Figure 1: Causes of Death for 15 to 49 year olds

   Let’s look at the causes of death in the above survey and we can see that  globally the leading cause of death in this age group is cardiovascular disease which concerns heart diseases, followed by cancers which both account for more than one million deaths. Therefore, Mitsufuji Corporation announced Hamon smartwear that saves human lives from death and dangerous environments. This is a simple approach: employees who perform their tasks in challenging environments wear IOT devices to monitor and alert them in high-risk circumstances.

Figure 2: Left Side: CEO of smartwear maker , Right side : hamon sensor

Figure 3: Visualizing the workflow of Hamon Smartwear

   Sensors-based projects are a little complicated and need more knowledge about biometric information, physics phenomenon, machine learning models and data structure. As technologies become more arising, the trend of today we have already known is Artificial Intelligence and A.I technologies are applied with sensor based projects to predict future outcomes. 

   Our research is aimed to support better products for Mitsufuji Corporation with AI Technologies. Our research demonstrates that we can predict the future heart rate which uses the previous heart rate training data whether the person is in safe or dangerous mode. If the person who is wearing a hamon device’s heart rate is abnormal, the transmitter will be alerted, the information will be sent to the cloud server and notify the administration department which can save the labourer’s safety in time.

Dataset

   In this research, we used electrocardiogram (ECG) heart rate data from Mitsufuji Corporation to train the AI model. Let’s take a look at the 8 main kinds of sensors being tested in this research. 

1) Drownsine Sensor

The level of drowsiness is measured approximately every 5min and tests for whether a person is sleepy or not. 

2) Falldown Sensor

This sensor examines the movement of the human body, recognizes a fall from normal daily activities by an effective algorithm and automatically sends requests for help to the caregivers with the patients’ location. 

3) Acceleration Trend Sensor

There are three axis models to convert the mechanical motion into an electrical signal.

4) Position Sensor

Position Sensor permits position measurement. It can be linear,angular or multi-axis. 

5) RRI Trend Sensor

The R-R interval fluctuation in electrocardiogram(ECG) is called heart rate variability, which reflects activities of the autonomic nervous system(ANS) and has been used in monitoring systems. 

6) Gyro Sensor

Gyro uses 3-axis positions, calculates degree of freedom, movement of things.

7) Stress Sensor

A contact sensor that responds to the forces produced by mechanical contact. 

8) Hypothermia Sensor

The hypothermia sensor will detect temperature and send an alarm  to the 24 hour response center if the temperature drops below a safe level.

   According to the previous survey result of 2017 death’s cause,  heart diseases are the highest death cause. In that case, among various data from different sensor types, we choose rri_trend sensor data, which represents heart rate variability and reflects activities of the autonomic nervous system(ANS) data, to make analysis for this research.

Training

   In order to predict future heart rate, we used an AI model called LSTM (Long Short Term Memory). LSTM is perhaps the most successful type of recurrent neural network that is capable of directly supporting multivariate sequence prediction problems.

Figure 4: Sliding Window Approach to Modeling Time Series Taken from “Time-series Extreme Event Forecasting with Neural Networks at Uber”.

   Training dataset was prepared by splitting the original data into input sequence (X) and output sequence (Y) using sliding windows mechanism. We tested by changing different sizes of the look-back and forecast horizon in the experiments. The following is one of our trained research that shows how precise the model is. By seeing the variance (difference) between the train loss and validation loss, we can conclude that the model is good because the variance is only about 0.001.

Figure 5: Comparison between actual heart rate data and predicted heart rate result

Summary

   Nowadays, Smartwear technology is becoming the most popular trend and we hope to penetrate the smart IOT device solutions which combine with AI technologies in the business market. Our Nexidea AI research team was leading to get accurate predictions in case this smart wear prediction is based on each person’s safety and people’s lives. That’s why we chose the most advanced AI model to predict accurately. 

   Since this research is based on time series data, the data preprocessing part is important to be in a regular interval time sequence. The research outcome is supposed to be a good prediction because the training loss and prediction loss between them is actually small. There might also be challenges as the training time will be longer when the input sequence is longer and how to handle peaks as the heart rate reaches to the high rate.

   Our Nexidea AI Research team are looking forward to better real-life research to aid the society to be more comfortable and to be safe as much as we can. According to this forecasting project , we got a lot of knowledge about the time sequence based AI models, how time series models are applied in business sectors and the spreading of smart technologies.