Multivariate residual-based health index for human health monitoring
Ambulatory or in-hospital monitoring of patients is provided with early warning and prioritization, enabling proactive intervention and amelioration of both costs and risks of health care. Multivariate physiological parameters are estimated by empirical model to remove normal variation. Residuals are tested using a multivariate probability density function to provide a multivariate health index for prioritizing medical effort.
1. A method for monitoring the health of a human, comprising:
obtaining sensor data from a human;
generating with a programmed microprocessor a plurality of features from said sensor data, characteristic of physiological health of said human;
estimating with a programmed microprocessor values for said features characteristic of normal human physiology using a multivariate model, based on the values of said generated plurality of features;
differencing with a programmed microprocessor the estimated values and the generated features to provide a set of residuals for the features, wherein each residual is the difference between the particular feature value expected according to said model, and the corresponding feature value generated from said sensor data; and
determining with a programmed microprocessor a likelihood that said set of residuals is representative of a pattern of normal residuals, by using a Gaussian mixture model based on a set of normal residual reference patterns to approximate a probability distribution for normal residual patterns, and to compute said likelihood that said set of residuals belongs to the distribution, whereby said likelihood consolidates the behaviors of the individual residuals for each of the features into one overall index; and
applying with a programmed microprocessor a test to said likelihood to render a decision whether the generated features are characteristic of normal physiological behavior to provide an early indication of deviation of the physiological health of said human from normal.
2. A method according to claim 1 , wherein said step of applying a test comprises comparing with a programmed microprocessor the logarithm of the inverse of said likelihood to a threshold.
3. A method according to claim 1 , further comprising the step of
testing with a programmed microprocessor a series of said rendered decisions for persistence of like decisions.
4. A method according to claim 1 , wherein said step of estimating values further comprises making a kernel-based comparison of a feature vector, comprising the values of said feature signals, to at least some of a library of exemplary vectors, each comprising values representative of said feature signals in a known health state, in order to generate said estimate as a linear combination of those exemplary vectors, weighted in relation to said comparisons.
5. A method according to claim 4 , wherein said feature vector is compared to said exemplary vectors comprising said library in order to select a subset of said exemplary vectors to use in said kernel-based comparison for generating said estimate.
6. A method according to claim 4 , wherein said estimate is generated as a linear combination of said exemplary vectors, weighted in relation to said comparisons according to:
x
est
=
∑
i
=
1
L
x
i
K
(
x
new
,
x
i
)
∑
i
=
1
L
K
(
x
new
,
x
i
)
where x new is said feature vector, x i are said exemplary vectors, x est is said estimate, and K is said kernel-based comparison.
7. A method according to claim 4 , wherein said estimate is generated as a linear combination of said exemplary vectors, weighted in relation to said comparisons according to:
x
est
=
D
(
D
T
⊗
D
)
-
1
(
D
T
⊗
x
new
)
∑
(
D
T
⊗
D
)
-
1
(
D
T
⊗
x
new
)
where x new is said feature vector, D is a matrix of at least some of said exemplary vectors, x est is said estimate, and {circumflex over (x)} is an operator for performing said kernel-based comparisons between matrices.
8. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
h
(
x
a
,
x
b
)
∝
ⅇ
-
x
a
-
x
b
p
n
h
where x a and x b are vectors being compared, h is a constant, p is the order of the norm, and n is a power to which the norm is raised, and K is the scalar result of the comparison.
9. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
h
(
x
a
,
x
b
)
∝
(
1
+
x
a
-
x
b
p
n
h
)
-
1
where x a and x b are vectors being compared, h is a constant, p is the order of the norm, and n is a power to which the norm is raised, and K is the scalar result of the comparison.
10. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
h
(
x
a
,
x
b
)
∝
1
-
x
a
-
x
b
p
n
h
where x a and x b are vectors being compared, h is a constant, p is the order of the norm, and n is a power to which the norm is raised, and K is the scalar result of the comparison.
11. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
(
x
,
y
)
∝
1
L
∑
m
=
1
L
ⅇ
-
x
m
-
y
m
n
h
m
where x and y are vectors being compared, h m are constants, m is the number of features, and n is a constant power, and K is the scalar result of the comparison.
12. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
(
x
,
y
)
∝
1
L
∑
m
=
1
L
(
1
+
x
m
-
y
m
n
h
m
)
-
1
where x and y are vectors being compared, h m are constants, m is the number of features, and n is a constant power, and K is the scalar result of the comparison.
13. A method according to claim 4 , wherein said kernel-based comparison is of the form:
K
(
x
,
y
)
∝
1
L
∑
m
=
1
L
(
1
-
x
m
-
y
m
n
h
m
)
where x and y are vectors being compared, h m are constants, m is the number of features, and n is a constant power, and K is the scalar result of the comparison.
14. A method according to claim 1 , wherein said step of obtaining sensor data comprises making measurements of sensors embedded inside the monitored human in connection with an implanted cardiac device.
15. A method according to claim 1 , wherein said step of obtaining sensor data comprises receiving wireless transmissions via extremely local radio protocol of measurements of sensors attached to the monitored human.
16. A method according to claim 1 , wherein said step of obtaining sensor data comprises receiving data from a ventilator.
17. A method for monitoring the health of a human, comprising:
obtaining sensor data from a human;
generating with a programmed microprocessor a plurality of features from said sensor data, characteristic of physiological health of said human;
estimating with a programmed microprocessor values for said features characteristic of normal human physiology using a multivariate model, based on the values of said generated plurality of features;
differencing with a programmed microprocessor the estimated values and the generated features to provide a set of residuals for the features, wherein each residual is the difference between the particular feature value expected according to said model, and the corresponding feature value generated from said sensor data;
determining with a programmed microprocessor for each of a plurality of known health states, a likelihood that said set of residuals is representative of a pattern of residuals characteristic of that known health state, by using a Gaussian mixture model based on a set of residual reference patterns for the known health state to approximate a probability distribution for residual patterns of that known health state, and to compute said likelihood that said set of residuals belongs to the distribution; and
applying with a programmed microprocessor a test to the plurality of likelihoods, each corresponding to one of the known health states, to render a ranking of which of said plurality of known health states the generated features are most characteristic of.
18. A method according to claim 17 , wherein said step of estimating values further comprises making a kernel-based comparison of a feature vector, comprising the values of said feature signals, to at least some of a library of exemplary vectors, each comprising values representative of said feature signals in a known health state, in order to generate said estimate as a linear combination of those exemplary vectors, weighted in relation to said comparisons.
19. A method according to claim 18 , wherein said feature vector is compared to said exemplary vectors comprising said library in order to select a subset of said exemplary vectors to use in said kernel-based comparison for generating said estimate.
20. A method according to claim 18 , wherein said estimate is generated as a linear combination of said exemplary vectors, weighted in relation to said comparisons according to:
x
est
=
∑
i
=
1
L
x
i
K
(
x
new
,
x
i
)
∑
i
=
1
L
K
(
x
new
,
x
i
)
where x new is said feature vector, x i are said exemplary vectors, x est is said estimate, and K is said kernel-based comparison.
21. A method according to claim 18 , wherein said estimate is generated as a linear combination of said exemplary vectors, weighted in relation to said comparisons according to:
x
est
=
D
(
D
T
⊗
D
)
-
1
(
D
T
⊗
x
new
)
∑
(
D
T
⊗
D
)
-
1
(
D
T
⊗
x
new
)
where x new is said feature vector, D is a matrix of at least some of said exemplary vectors, x est is said estimate, and {circumflex over (x)} is an operator for performing said kernel-based comparisons between matrices.
22. A method according to claim 1 , wherein the sensor data of said receiving step comprises an electrocardiogram, a bioimpedance, and a photoplethysmogram for at least two wavelengths.
23. A method according to claim 22 , wherein the plurality of features of said generating step comprises a heart rate, a respiration rate, a pulse transit time, and a ratio of absorption of the at least two wavelengths.
24. A method according to claim 23 , wherein the sensor data of said receiving step further comprises at least one accelerometer signal, and further comprising the step of identifying what times the accelerometer signal indicates motion artifact is likely present in the sensor data and ignoring sensor data at those times.
25. A method according to claim 1 , wherein said step of determining a likelihood further comprises scaling said residuals for the generated features, using the means and standard deviations calculated from known normal data residuals, and wherein said set of normal residual reference patterns are likewise scaled.
26. A method according to claim 3 , wherein said step of testing a series of rendered decisions comprises toggling on an alert latch when there have been a first selected minimum count of rendered decisions, within a first selected window of successive observations of the generated features, that the features are not characteristic of normal physiological behavior.
27. A method according to claim 26 , wherein said step of a testing series of rendered decisions further comprises, once the alert latch is on, toggling off the alert latch when there have been a second selected minimum count of rendered decisions, within a second selected window of successive observations of the generated features, that the features are characteristic of normal physiological behavior.
28. A method according to claim 17 , wherein the sensor data of said receiving step comprises an electrocardiogram, a bioimpedance, and a photoplethysmogram for at least two wavelengths.
29. A method according to claim 28 , wherein the plurality of features of said generating step comprises a heart rate, a respiration rate, a pulse transit time, and a ratio of absorption of the at least two wavelengths.
30. A method according to claim 29 , wherein the sensor data of said receiving step further comprises at least one accelerometer signal, and further comprising the step of identifying what times the accelerometer signal indicates motion artifact is likely present in the sensor data and ignoring sensor data at those times.
31. A method according to claim 17 , wherein said step of determining, for each of a plurality of known health states, a likelihood further comprises scaling said residuals for the generated features, using the means and standard deviations calculated from known normal data residuals for each respective known health state, and wherein said set of normal residual reference patterns are likewise scaled.