Novelty detection with missing parameters

• US 8,949,072 B2
• Filed: 10/09/2009
• Issued: 02/03/2015
• Est. Priority Date: 10/09/2008
• Status: Expired due to Fees

• Current Assignee: Oxford Biosignals (Us) Incorporated
• Original Assignee: Oxford Biosignals (Us) Incorporated
• Inventors: Tarassenko, Lionel, Hann, Alistair
• Primary Examiner(s): Tsai, Carol S
• Application Number:
  US13/122,801
• Publication Number:
  US 20110265026A1
• Time in Patent Office: 1,943 Days
• Field of Search: 702/19, 702/67, 702/181, 702/189, 600/301, 600/484, 600/512, 600/544
• US Class Current: 702/181
• CPC Class Codes: A61B 2560/0276 Determining malfunction
• CPC Class Codes: A61B 5/02055 Simultaneously evaluating both cardiovascular condition and temperature
• CPC Class Codes: A61B 5/021 Measuring pressure in heart or blood vessels A61B5/0205 takes precedence
• CPC Class Codes: A61B 5/0245 by using sensing means generating electric signals, i.e. ECG signals
• CPC Class Codes: A61B 5/0816 Measuring devices for examining respiratory frequency measuring frequency of electric signals G01R23/00
• CPC Class Codes: A61B 5/145 Measuring characteristics of blood in vivo, e.g. gas concentration, pH value; Measuring characteristics of body fluids or tissues, e.g. interstitial fluid, cerebral tissuemeasuring of blood pressure or blood flow A61B5/02; non-radiation detecting or locating of foreign bodies in blood A61B5/06
• CPC Class Codes: A61B 5/6843 Monitoring or controlling sensor contact pressure
• CPC Class Codes: G06F 18/2433 Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G16H 10/00 ICT specially adapted for the handling or processing of patient-related medical or healthcare data for medical reports G16H15/00; for therapies or health-improving plans G16H20/00; for the handling or processing of medical images G16H30/00
• CPC Class Codes: G16H 40/67 for remote operation
• CPC Class Codes: G16H 50/70 for mining of medical data, e.g. analysing previous cases of other patients

Abstract

A method of obtaining a consistent evaluation of the state of the system which has been monitored by measurement of multiple parameters of that system. The multiple parameters are used to calculate a single dimensional value based on the distance between the current state and normal states of the system using a Parzen Windows probability function. Consistent single dimensional values regardless of the dimensionality of the original data set can be obtained by finding a relationship between the single dimensional value and the probability of status of the system. Different relationships are obtained for different dimensionalities of data sets. Sensor malfunction can also be detected by testing the probability of the state implied by measuring all of the available parameters against the probability of the state implied by ignoring different individual ones of the parameters. A significant disparity in the two probabilities indicate possible sensor malfunction.

Claims

1. A method of determining and displaying an index of novelty representing the state of a system based on measurements of a plurality of different parameters of the system, the system being a physical system or a biological system, the plurality of parameters defining respective dimensions of a multi-dimensional measurement space, the method being performed by a processor, and the method comprising the steps of:

receiving sets of values each value in a set being a measurement of a different one of said plurality of parameters, each set defining a data point in said multi-dimensional measurement space;

calculating the index of novelty for each data point by comparing the position in said multi-dimensional measurement space of each data point to a set of prototype data points comprising measurements of said parameters representative of the system being in a normal state,

wherein in the event of receipt of a deficient data point missing one or more of said parameter values of said set, the method further comprises the steps of;

calculating a marginal index of novelty of the deficient data point based on the position of the deficient data point in a reduced dimensionality space omitting the dimensions corresponding to the missing parameter values,

using a predefined relationship, calculated for said reduced dimensionality space, between the marginal index of novelty and the marginal probability of different states of the system to find a marginal probability value representing the probability of the system being at least as close to said normal state as the state represented by the deficient data point,

using a further predefined relationship, calculated for said multi-dimensional measurement space, between the index of novelty and the probability of different states of the system to find the index of novelty in said multi-dimensional measurement space corresponding to a probability equal to said marginal probability value; and

displaying the index of novelty so found as the index of novelty for the deficient data point.

2. A method according to claim 1 wherein the marginal index of novelty of the deficient data point is based on comparing the distance of the deficient data point in the reduced dimensionality space from a set of prototype data points comprising measurements of said parameters representative of the system being in a normal state.

3. A method according to claim 2 wherein the set of prototype data points in the reduced dimensionality space are the prototype points in the multi-dimensional measurement space omitting the dimensions corresponding to the missing parameter values.

4. A method according to claim 1 wherein in calculating the index, or marginal index, of novelty for each data point its position is compared to the set of prototype data points using a probability density function based on the Euclidian distance between that data point and each of the prototype data points.

5. A method according to claim 4 wherein the probability density function is the Parzen Windows function.

6. A method according to claim 1 wherein the prototype points are cluster centres in a training set of multiple samples of data points.

7. A method according to claim 6 wherein the same cluster centres are used in calculating the marginal index of novelty in the reduced dimensionality space regardless of which parameter values are missing.

8. A method according to claim 6 wherein the cluster centres are recalculated in each reduced dimensionality space.

9. A method according to claim 1 wherein the predefined relationships in said reduced dimensionality space and in said multi-dimensional measurement space are based on a training set of multiple samples of data points.

10. A method according to claim 9 wherein the predefined relationships are based on taking as thresholds values of the index, or marginal index, of novelty progressively ranging from a minimum to a maximum value, finding for each of said thresholds the fraction of data points in the training set which have an index, or respectively marginal index, of novelty below the threshold, and taking that fraction as the probability value, or respectively marginal probability value, for that index, or respectively marginal index, of novelty.

11. A method according to claim 1 wherein a plurality of the predefined relationships in different reduced dimensionality spaces are used, each being based on the same single training set of data points and omitting a different parameter or combination of parameters from the sets of values forming each data point of the training set.

12. An apparatus for determining and displaying an index of novelty representing the state of a system based on measurements of a plurality of different parameters of the system, the apparatus comprising the processor performing the method of claim 1, an input for receiving said measurements of a plurality of different parameters, and a display for displaying the index of novelty.

28. A method according to claim 1 wherein the system is human or animal.

29. A method according to claim 1 wherein the parameters are medical measurements.

30. A method according to claim 29 wherein the parameters comprise any of heart rate, blood pressure, temperature, blood oxygen saturation, breathing rate, Glasgow Coma Score.

13. A method of determining and displaying an index of novelty representing the state of a system based on measurement of a plurality of different parameters of the system, the system being a physical system or a biological system, the method being performed by a processor, and the method comprising the steps of:

receiving a data set comprising a plurality of sets of values each value being a measurement of a different one of said plurality of parameters, each set of values defining a data point in measurement space of dimension D, where D is a positive integer greater than one;

calculating a provisional index of novelty for each data point by comparing the position in said measurement space of dimension D of each data point to a set of prototype data points comprising measurements of said parameters representative of the system being in a normal state;

calculating a relationship, in said measurement space of dimension D, between the provisional index of novelty and the probability of different states of the system to find a probability value representing the probability of the system being at least as close to said normal state as the state represented by a data point having that provisional index of novelty,

using said calculated relationship to find for each data point of the data set the probability value corresponding to its provisional index of novelty;

using a further predefined relationship between the probability of different states of the system and an index of novelty defined in a space of dimension L, where L is a positive integer less than D, and based on comparing the distance between a data point and said normal state of the system, to find for each data point of the data set the index of novelty corresponding in the space of dimension L to said probability value; and

displaying the index of novelty so found as the index of novelty for that data point.

14. A method according to claim 13 wherein in calculating the index, or provisional index, of novelty for each data point its position is compared to the set of prototype data points using a probability density function based on the Euclidian distance between that data point and each of the prototype data points.

15. A method according to claim 14 wherein the probability density function is the Parzen Windows function.

16. A method according to claim 13 wherein the predefined relationships are based on taking as thresholds values of the index, or provisional index, of novelty progressively ranging from a minimum to a maximum value, finding for each of said thresholds the fraction of data points which have an index, or respectively provisional index, of novelty below the threshold, and taking that fraction as the probability value for that index, or respectively provisional index, of novelty.

17. An apparatus for determining and displaying an index of novelty representing the state of a system based on measurements of a plurality of different parameters of the system, the apparatus comprising the processor performing the method of claim 13, an input for receiving said measurements of a plurality of different parameters, and a display for displaying the index of novelty.

18. A method of detecting sensor malfunction amongst a plurality of different sensors each measuring a parameter of the system, the plurality of parameters defining respective dimensions of a D-dimensional measurement space, where D is equal to the number of sensors, the method comprising the steps of:

receiving sets of values, each value in a set being a measurement from a different one of said sensors, each set defining a data point in said D-dimensional measurement space;

calculating an index of novelty for each data point by comparing its position in said D-dimensional measurement space to a set of prototype data points comprising measurements of said parameters representative of the system being in a normal state;

using a predefined relationship, calculated for said D-dimensional measurement space, between the index of novelty and the probability of different states of the system to find a probability value representing the probability of the system being at least as close to said normal state as the state represented by the data point;

calculating for each data point at least one marginal novelty index by ignoring one parameter value from said set of values and comparing the position of the data point in a D-1 dimensional space to the set of prototype data points also in the D-1 dimensional space, the D-1 dimensional space omitting the dimension corresponding to the ignored parameter;

using a further predefined relationship, calculated for said D-1 dimensional space, between the marginal index of novelty and the marginal probability of different states of the system to find a marginal probability value representing the probability of the system being at least as close to said normal state as the state represented by the marginal novelty index;

comparing said probability value and said marginal probability value; and

in the event of the probability value and marginal probability value differing by more than a predefined threshold, outputting an alert for malfunction of the sensor whose parameter value was ignored.

19. A method according to claim 18 further comprising repeating the steps of calculating a marginal novelty index and finding a marginal probability value for each data point, each time ignoring a different parameter value, and comparing each resulting marginal probability value to the probability value, to check for malfunction of each of said sensors.

20. A method according to claim 18 further comprising outputting the alert if only one marginal probability value differs by more than a predefined threshold from said probability value.

21. A method according to claim 18 wherein in calculating the index, or marginal index, of novelty for each data point its position is compared to the set of prototype data points using a probability density function based on the Euclidian distance between that data point and each of the prototype data points.

22. A method according to claim 21 wherein the probability density function is the Parzen Windows function.

23. A method according to claim 18 wherein the prototype points are cluster centres in a training set of multiple samples of data points.

24. A method according to claim 23 wherein the same cluster centres are used in calculating the marginal index of novelty in the D-1 dimensionality space regardless of which parameter value is missing.

25. A method according to claim 18 wherein the predefined relationships in said D-1 dimensionality space and in said D-dimensional measurement space are based on a training set of multiple samples of data points.

26. A method according to claim 25 wherein the predefined relationships are based on taking as thresholds values of the index, or marginal index, of novelty progressively ranging from a minimum to a maximum value, finding for each of said thresholds the fraction of data points in the training set which have an index, or respectively marginal index, of novelty below the threshold, and taking that fraction as the probability value, or respectively marginal probability value, for that index, or respectively marginal index, of novelty.

27. An apparatus for detecting sensor malfunction amongst a plurality of different sensors each measuring a parameter of the system, the apparatus comprising an input for receiving inputs from said sensors, a processor adapted to execute the method steps of claim 18, and a device for outputting said alert.