Share on FacebookTweet about this on TwitterPin on PinterestShare on Google+

Long-Age-Cover


33% of Japan’s population is above 60 years of age. And this is set to increase to 40% by 2050. This altering demographic trend is creating a tremendous pressure on the government to provide better healthcare services and policy support.

A lot of elderly people are already living alone and many deaths go unnoticed for days every year (孤独死 – solitary death). There are young people that take care of their adult children as well as parents (三世代家族 – three generation households). Every year 400 primary schools are converted to care center for elderly. Government burden of healthcare/pension costs has increased and with that an increase in retirement age.

We believe the solutions to this crisis is a shared responsibility of the government and the private sector. We thus intend to introduce technologies with an aim to help elderly people maintain healthy lives by themselves and also actively participate in society.  Technologies that can reduce the burden on young people to keep track of each and every need of their grandparents while still enjoying time with them. Technologies that can make elderly people more active in social life rather than getting secluded. How?

 


 

Enter IoT. A connected health monitoring system + Make it intelligent.

 

A Simple Overview

 

There are essentially 3 elements to it:

  • Sensor Network (Data source)
  • SaaS Platform (Data management and Machine learning)
  • Alter devices (Alerts and predictions)

 

Safe_Aging_Blog

Fig. 1: Healthcare IoT Platform for Smart Aging

 

The platform collects live data from sensors like smart watches, IR sensors & depth sensors installed at home + data from weight readings, blood pressure, ECG etc. which can all be performed at home nowadays.

Other kinds of data from hospital and doctors is also collected. The individuals complete data is monitored by AI/Machine Learning algorithms for anomalous data points and patterns in person’s activity to intelligently determine disease risk and symptoms.

 

What Can be Achieved ?

With powerful machine learning algorithms and availability of  centralised medical data a lot of practical possibilities open up. Some potential applications are –

  • Activity recognition like sitting, standing, coughing, walking, jumping etc. to measure health fitness.
  • Detection of disease symptoms + Research in identification of new symptoms from big data.
  • Disease risk prediction.
  • Alert doctors/family in cases of emergencies.

 


A Case Study: The SPHERE Challenge

We participated in the recently concluded Safe Aging with SPHERE contest which was conducted at datadriven.org in collaboration with ECML-PKDD 2016 conference and AARP foundation.

The data was gathered by a robust set of sensors – cameras, IR sensors, accelerometer sensors. A description of the dataset can be found here. The task was to predict one of 20 different kinds of activities the person may do at every second. This included activities like ascending stairs, descending, jumping, walking, walking with weight, lying, kneeling, standing, sitting, sit-to-lie transition, stand-to-kneel transition etc.

 

Using these sensor readings and tree based machine learning models we concluded at 6th position on the leaderboard. We give a summary of our methodology.

A.  Evaluation Metric

Performance was evaluated using Brier score which gives higher weight to rare activity events like ascending/descending and less weight to common activities like walking/standing/sitting.

B.  Model

We used statistical features like 5-Quantiles, mean, min, max and rolling means/medians of sensor readings in 1 second windows from all sensors. Accelerometer readings of x, y, z axis were also combined to form the sum of axis, sum of absolute values of axis, sum of squares of axis etc.

Models used were Boosted linear models, boosted decision trees (xgboost library), Random forests and Extra trees. Predictions from these models were time shifted by 1, 2, 3 and 4 seconds to form features for further models of the same type. Stacking was done at later stages to achieve the best stacked model. Figure 2 shows schema of the modelling framework.

model_stack

Fig. 2: Our Machine Learning Model Pipeline

 

The main model is a custom xgboost training functions which runs training on 20 models for 20 different targets simultaneously and stops at the best cumulative Brier score. The code is on this gist.

 

Screen Shot 2016-08-02 at 22.32.20

Fig. 3: Final Leader Board

 

We would love to hear from medical sensor manufacturers and other interested in partnerships with this platform.

 

Leave a Reply

メールアドレスが公開されることはありません。 * が付いている欄は必須項目です

You may use these HTML tags and attributes:

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code class="" title="" data-url=""> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong> <pre class="" title="" data-url=""> <span class="" title="" data-url="">