Our Process
Our Machine Learning Algorithm
Our Platform is dedicated to detect the vital changes in human behavior as precursors of the clinical deterioration
We program a patient’s smart wearable to collect biometric data at multiple, clinically meaningful intervals. We also constantly update the individual baseline with 24-hour delay period to build personalized, non-risk baseline and alert when any deviation occurs to indicate worsening health condition.
Our Platform integrates the quantitative data with our advanced hybrid AI algorithm dedicated to clinical analysis. The physical activity is measured by the number of occurrences of the multiples of g-force (x, y, z) each second.
We then use a Decision Tree Ensemble based multiclass classification approach to classify risk into 3 levels: high, medium, and no risk. A maximum tree depth of 3 levels is deployed.
The intrinsic graph of the decision tree facilitated the explainability of the model. Due to the nature of the decision tree to bisect the data space and the tendency to overfit the training data when classes are not well separated, we introduced regularization term in order to balance the bias-variance tradeoff (the second term of the training objective equation below)
An initial data generation based on example data to predict clinical outcomes showed that a 60/40 ratio of active vs resting states yielded a healthy outcome. Any derivation that yielded an active state <60% and or a resting state for >40% resulted in a deceased state.
