«Hello CARU!»

Caru is a voice-first solution that fosters autonomy and safety in the everyday life of elderly people. Users can communicate with relatives and call for help. In addition activity patterns can be derived from different sensors like CO2, temperature, noise, etc. If irregularities in the data are detected, Caru actively notifies caregivers.

Safety for elderly people is an important issue especially now in Covid times when contact and help from the family has to be almost completely avoided. Through Caru, the voice-activated emergency call, contact with family members can be continued with an integrated family chat and safety can be guaranteed through Machine Learning. Caru wishes to use five built-in sensors in the device (voice control, light, CO2, temperature and humidity) and the linked human activity to recognize when a person needs help and triggers the signal to stored contacts that are called one after the other in case of an emergency.
 

As our students were given a clean and structured dataframe, the exploratory data analysis was rapid and straightforward. Several multidimensional representations of data were created to better understand it. Yet, the challenge was to translate Time Series Data into any human activity index, as no available tools existed yet for this purpose. They used the state-the-art of Data Science tools to complementary data analyses based on Machine Learning algorithms:
 

By means of these anonymized Time Series Data that are queried by the device every 20 seconds, correlations with putative activity are established. Thus, automatic emergency calls are handled without the need to press a button  but by Machine Learning.

During the three-month Data Science Bootcamp, a deeper understanding of the basic concepts in Data Science was acquired and skills in R and Python were expanded and put in practice. All data analysis codes were developed in Jupyter notebooks in Python, currently the most popular programming language for Machine Learning. The students also used Pandas, the data analysis and manipulation tool for Python, and scikit-learn, which is considered the most comprehensive library for Machine Learning in Python. Scikit-learn was used for Semi-Supervised and Unsupervised Learning which in this case meant clustering time spans into active and inactive phases and determining the supposed activity from unlabeled data. Furthermore, Prophet, the Facebook library for Time Series Prediction, was used to determine when activities deviated from the normal daily pattern. An attempt was made to determine when activities deviated from the normal daily pattern. In order to visualize the findings, the tool Matplotlib was used and Streamit for prototyping. Furthermore, the students used Visual Studio and Google Colab to perform local and cloud computing of the data. Their project was done entirely remotely and was presented during a webinar.
 

Within the given time frame (4 weeks) our students covered several aspects of a project development: Investigating business goals, exploratory data analysis, and machine learning based data analysis (3 complementary strategies). An extension to the work could be to provide a user-friendly dashboard. Eventually, Caru will be able to use these approaches to contact the person in their flat or to notify relatives if unusual activity for the period has been detected. To achieve this, several steps would need to be realised: assess the robustness of our algorithms regarding the data variability, integrate our algorithms within the existing backend, co-design the UX/UI interface with the Caru’s team and refine our tool based on the caregivers feedback.

Caru is currently in use by many private individuals and is considered a faithful, daily companion.