Non-parametric modeling apparatus and method for classification, especially of activity state
The activity state classification method of the present invention employs a kernel-based modeling technique, and more specifically a set of similarity-based models, which have been created using example data, to process an input observation or set of input observations, each comprising a set of sensor readings or “features” derived there from or other data, to predict the activity state of a person from whom the sensor data was obtained. A model is created for each class of activity. The input data is processed by each model and the resulting predictions are combined to yield a final prediction of which state of activity is represented by the input data.
1. A method for determining the class of an activity of a person, selected from a plurality of activities of interest, based on multivariate sensor data comprising values or features derived from values of sensors measuring physiological parameters from the person, comprising the steps of:
providing a plurality of kernel-based models, each corresponding to one of said activities of interest, and each model comprising a plurality of reference observations of said multivariate sensor data, at least some said reference observations having been acquired from sensors on a person during a modeled activity to which said model corresponds, and at least some said reference observations having been acquired from sensors on a person during activity different from that to which said model corresponds, and all said reference observations further having a class membership value corresponding to whether said reference observation is of said modeled activity of not;
obtaining a new observation of readings of said multivariate sensor data;
generating in a computer processor an inferential estimate of said class membership value for each of at least some of said kernel-based models using said new observation as input to the at least some of said kernel-based models, where said inferential estimate for a given kernel-based model is generated from a linear combination of at least some of said reference observations of said given kernel-based model; and
determining the class of activity corresponding to said new observation based on a comparison of the class membership estimates.
2. A method according to claim 1 , wherein at least one of said kernel-based models is a similarity-based model.
3. A method according to claim 2 , further comprising the step of determining a class of activity of a window of successive observations in time to be the class most often represented among the successive observations.
4. A method according to claim 3 , wherein said sensor data includes a measure of heart rate.
5. A method according to claim 3 , wherein said sensor data includes a measure of temperature of the person.
6. A method according to claim 3 , wherein said sensor data includes a measure of respiration.
7. A method according to claim 3 , further comprising the step of filtering the window of successive observations as an input to a medical diagnosis based on the determination of the class of the window.
8. A method according to claim 3 , wherein the computer processor used to generate the inferential estimate of said class membership value is embedded in a wearable hardware and further comprising the step of selectively initiating an action by said hardware upon determination of the class of activity corresponding to said window of successive observations.
9. A method according to claim 3 , wherein said new observation is used to localize each said kernel-based model so that each said inferential estimate is generated from a linear combination of just a subset of reference observations of each said kernel-based model.
10. A method according to claim 9 , wherein said subset is determined by a kernel-based similarity comparison of said new observation with said reference observations.
11. A method according to claim 10 , wherein said kernel-based similarity comparison of said new observation with said reference observations marks for inclusion in said subset a predetermined number of those reference observations that have a highest similarity value when compared with said new observation.
12. A method according to claim 10 , wherein said kernel-based similarity comparison of said new observation with said reference observations marks for inclusion in said subset those reference observations that have a similarity value above a predetermined threshold, and that further either have the highest similarity value with the new observation among a set of all reference observations that have a selected sensor data value less than the corresponding sensor data value of said new observation, or have the highest similarity value with the new observation among a set of all reference observations that have said selected sensor data value greater than the corresponding sensor data value of said new observation.
13. A method according to claim 1 , wherein at least one of said kernel-based models is a kernel regression model.
14. A method according to claim 13 , further comprising the step of determining a class of activity of a window of successive observations in time to be the class most often represented among the successive observations.
15. A method according to claim 14 , wherein said sensor data includes a measure of heart rate.
16. A method according to claim 14 , wherein said sensor data includes a measure of temperature of the person.
17. A method according to claim 14 , wherein said sensor data includes a measure of respiration.
18. A method according to claim 1 , further comprising the step of filtering said new observation as an input to a medical diagnosis based on the determination the class of activity.
19. A method according to claim 1 , wherein the computer processor used to generate the inferential estimate of said class membership value is embedded in a wearable hardware and further comprising the step of selectively initiating an action by said hardware upon determination of the class of activity corresponding to said new observation.