Research at Hypnos
Machine Learning at Hypnos
Our dataset resources are the followings:
Data structure for training
Our data is a time series and we needed our algorithms to recognize previous and next data for a specific event. To do so, we prepared windows with various sizes containing heart rates, heart rate mean, heart rate consecutive differences etc. if we are going to be training with heart rates, for example.
We used following algorithms for machine learning,
- Neural Networks
- Support Vector Machine
- Linear Discriminant Analysis
- Random Forest Classifier
- K-Nearest Neighbors Algorithm
These are currently the most succesful models trained with machine learning.
- Apnea model - 65% accuracy
- Sleep state model - 65% accuracy
- Heart attack model - 70% accuracy
The purpose of unsupervised learning is to support the supervised model in its classifications. As shown in the diagram, the outlier detection model is trained real-time with users' own data. This data is continuously fed to the model. As more and more a user sleeps with Hypnos, the more Hypnos will "get used to" the user, as a result, the more accurate its personal feedback will be.