The present invention relates to a method and apparatus for monitoring the status of a patient as measured by a plurality of vital signs.
It is common in a clinical setting, particularly a high dependency care unit or intensive care unit, to monitor continuously the condition or status of patients. Typically vital signs such as one or more of blood pressure, heart rate, skin or body temperature, blood oxygen saturation (for instance measured by pulse oximetry with a finger probe), breathing rate (for instance measured by electrical impedance pneumography) may be monitored. The heart rate, and sometimes breathing rate, may be obtained from one or more channels of an electrocardiogram (ECG), which may also be monitored continuously. It is also possible to derive from these primary signals some secondary parameters such as heart rate variability, S-T segment elevation/depression and so on. Normally the individual measurements are displayed to a clinician in a form of a number, and possibly a graphical trace, as illustrated, for example, in
However, normal values for vital signs can vary considerably from one patient to another due to factors such as age, type of disease, medication, type of surgery, etc. Replacing the population-based statistical model with a personal model for an individual patient is problematic because no valid set of measurements for developing such a model exists for an individual, especially when they have just changed status, for example because of a surgical operation or developing a disease or other medical condition. In the applicant's own U.S. Pat. No. 7,031,857 B2 it was proposed that after an initial period of judging a patient's condition against a population-based statistical model, enough measurements on the individual patient would have been obtained to replace the population based model with a patient-specific one. However, in practice this is not possible as a patient's condition is changing and obtaining regulatory clearance for monitoring on the basis of a not-yet validated set of data is impossible. Therefore it is difficult to provide a scoring system which can provide a reliable and valid indication of variations from normality for an individual patient. Further it would require a long period of monitoring the patient once they have reached a stable state to provide a statistically reliable set of measurements from which a valid model can be generated. At the start of monitoring therefore, when the patient may be most at risk of sudden deterioration, a personalised model is not available. So it is not clear how to obtain a personalised model which is useful in a reasonably quick time.
The present invention provides an enhancement to a continuous patient monitoring system such as that disclosed in U.S. Pat. No. 7,031,857 B2 by comparing the status of a patient, as measured by a plurality of vital signs, simultaneously to two different statistical models, one being population-based and one being personal to the patient. The statistical model which is personal to the patient is obtained by applying machine learning techniques to an initial population-based model and updating it based on each current vital signs measurement from the patient as it is received. By simultaneously evaluating the patient's status against two statistical models an alert notification can be generated relative to the population-based model if a computed score representing the patient's condition breaches a predefined threshold (this model can be a regulatory approved model), thus providing safety for the patient if their condition deteriorates, but also a notification of a change in the patient's status compared to the patient-specific statistical model can also be generated. Such a change in status could be an improvement or deterioration for that particular patient, but which is still within an overall region of normality for the population, and it can thus provide an earlier indication of change.
In more detail, therefore, the present invention provides a method of monitoring the status of a patient comprising the steps of continuously measuring a plurality of patient vital signs to provide a succession of multi-parameter vital signs observations, comparing each successive multi-parameter vital signs observation to two statistical models, a first of the statistical models being a population-based reference model mapping the probability distribution of multi-parameter vital signs observations for a population of different individuals, and a second of the statistical models being a patient-specific model which is a probability distribution based at least partly on previous multi-parameter vital signs observations for the patient being monitored, calculating first and second numerical indices based respectively on the probability densities of the current multi-parameter vital signs observation obtained by comparison with the first and second statistical models, wherein an increased value of said indices represent an increase in abnormality with respect to the statistical model, outputting an alert notification if the first numerical index exceeds a first threshold, and outputting a status change notification if the second numerical index exceeds a second threshold.
The first threshold on the first numerical index in the population-based reference model effectively defines a global region of normality for the population. The region of parameter space occupied by the patient's current state would normally be expected to be smaller than the global region of normality and may be embedded within it. The second threshold effectively defines a region of normality personal to the individual patient and changes from this could reflect either an improvement or deterioration in the individual patient's condition depending on whether the change is accompanied by an increase or decrease in the first numerical index obtained by reference to the population-based model.
The vital signs measured may typically comprise heart rate, breathing rate, blood pressure, blood oxygen saturation and body temperature and at each time point the measured values form a multi-parameter vital signs observation for evaluation against the statistical models. Typically the various parameters are collected at different sampling rates, for example the heart rate may be calculated every few seconds whereas the blood pressure may be measured only once every 2 to 4 hours, and so to provide a continuous set of samples the lower sampling rate measurements are repeated in successive multi-parameter observations until they are re-measured.
Preferably the patient-specific statistical model is updated at each time step by using the most recent multi-parameter vital signs observation. However the updating step can be selectively enabled when the patient status is within the global normality region for the population, i.e. when the first numerical index does not exceed the first threshold. This can avoid updating the model based on noisy or spurious measurements. The inhibition on updating can be automatic based on the first numerical index, or can be under the control of a clinician. The threshold defining the region of global normality and the threshold for determining whether or not to update the patient specific model with an individual measurement can be different from each other. Thus it is possible for the clinician to selectively allow the patient-specific model to be updated even if the population-based model shows some degree of abnormality. In the case of one or more of the parameters in an observation being spurious or missing it is possible for this to be treated as a missing variable (for example replaced by its mean value) in determining whether to update the patient-specific model, but if it is decided to update then that the complete original observation is used to update the model. The decision on whether to update may be automatic, e.g. based on the degree of abnormality, or under the control of the clinician or operator.
The step of updating the patient-specific model can be stopped when the model becomes stable (for example when the parameters do not change significantly for a predetermined period of time, e.g. 20 or 30 minutes), and restarted when a comparison of the most recent multi-parameter vital signs observation (or some predefined number of them) with the patient-specific model indicates a change in patient status.
Preferably the patient-specific model is initialised to be identical to the population model, or to have the same statistical properties (e.g. mean and variance) as the population model such that it would give approximately the same values for the numerical indices. As patient-specific observations are received, this initialized model evolves to cover the area of normality for the particular patient being monitored.
The patient-specific model and the population-based model may be Kernel Density Estimators, such as a plurality of spherical Gaussian density functions with all kernels having equal weights. The patient's specific model may be updated by a Bayesian process which comprises calculating from the patient-specific model the likelihood of the most recent multi-parameter vital signs observation, multiplying the model's prior probability distribution by the calculated likelihood and then renormalizing it to form the Bayesian posterior probability distribution for the model, given the current sample, which then serves as the prior probability distribution for the next sample.
This update process for the patient-specific model may utilise particle filtering. This is a standard technique in which “particles” (being samples initially generated from the population-based model by random sampling of it) have a noise term added to each of them and then the likelihood of each particle plus noise is computed. A plurality of the particles is then resampled with a probability of selection being set proportional to its computed likelihood. The resampled set of particles constitutes the new patient-specific model. As an alternative to the standard particle filtering approach, the new samples can be drawn by sampling from the nearest kernel (to the selected particle) in the population-based model. This has the effect that the patient-specific model is always bounded by the population-based model. If it is desired for certain patients that a patient specific model may lie outside the population-based region of normality, then the standard resampling approach should be used.
As mentioned above the status change notification generated with reference to the patient-specific model can correspond to an improvement or deterioration in the patient's condition and it thus provides extra information to clinical staff compared to the results from the population-based reference model. Preferably an alert based on the patient-specific model is only generated if the current observation or observations are associated with an increase in the first numerical index generated with reference to the population-based reference model. This means that an alert is only generated if it reflects deterioration in the patient's condition.
Preferably such an increase of the first numerical index is judged by comparing the value of the first numerical index obtained for the current multi-parameter vital signs observation to the value of the first numerical index that was obtained at some previous time, for example at the time when the patient-specific model became stable or when the monitoring process was started.
To avoid generating alerts or status change notifications when spurious or erroneous measurements are received (for example because of noisy signals), the notifications may only be generated if a threshold is exceeded by a filtered version of the index (e.g. for more than a predefined number of observations in succession or more than a predefined number of observations within a predetermined time period, though other filtered versions may be used). This effectively smoothes the output of the system.
The invention may be embodied in an apparatus for executing the method and such an apparatus may therefore comprise an input for receiving the patient's vital signs measurements, a memory for storing the statistical models, a processor programmed to execute the steps of comparing the measurement to the statistical models, calculating the numerical indices and comparing them to the thresholds, and to output the alert and status change notifications. The output may comprise a display for displaying the notifications in a graphical form, though audible notifications may also be generated. Typically the display can display the individual measured vital signs in a conventional manner together with the numerical indices and alert and status notifications.
The apparatus may be incorporated into a standard vital signs monitor or can be embodied in a programmed computer system which receives the vital signs measurements.
The invention will be further described by way of examples with reference the accompanying drawings in which:—
The small ellipse within the larger ellipse in
As illustrated in
In this embodiment the population model (large ellipse in
In one embodiment of the invention, the approach for updating the patient-specific model of vital signs is to employ particle filtering (sequential Monte Carlo technique), though by imposing some assumptions other machine-learning techniques such as Kalman Filtering or Unscented Kalman Filtering can be used (see: Beyond the Kalman Filter: Particle Filters for Tracking Applications, Branko Ristic, Artech House Publishers, 2004, ISBN 158053631X, or The unscented Kalman filter for nonlinear estimation, Wan E. A, and Van der Merwe, R, Adaptive Systems for Signal Processing, Communications, and Control Symposium 2000. AS-SPCC. The IEEE 2000).
A particle filter is a non-Gaussian density estimator, which is formulated in the Bayesian framework. Bayesian learning estimates the distribution of interest (posterior distribution), using the prior distribution and the likelihood of the current observation (more details on dynamic Bayesian updating can be found in, for example, Arulampalam, M. S.; Maskell, S.; Gordon, N.; Clapp, T., “A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking, ”IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 174, 188, February 2002).
As mentioned above, the population distribution is used as the prior distribution for estimation of the posterior distribution for a specific patient. This allows measurement artefacts to be handled better and reduces the updating rate when the vital signs are well-away from the normal values. It also assists a better decision about the improvement or deterioration of the patient, for example, when the vital signs are moving away from the centre of normality of the patient model, but they are moving towards the centre of the normality of the population model as illustrated by arrow B in
Thus starting from the population distribution, to obtain the patient-specific distribution, suppose that the population distribution, i.e., the prior, is represented using a kernel density estimator such as Parzen window [see, e.g., U.S. Pat. No. 7,031,857B2]. Let N be the number of kernels, e.g. 400 in this embodiment, each of which is a Gaussian function
with mean μj, j=1, . . . , N and covariance C
, respectively. Here d is the dimensionality of data. It is also possible to have different covariance matrices for each kernel, as in the Gaussian mixture models (GMM), but here for the sake of simplicity of notation, we assume that the covariance matrices are the same for all kernels. The population distribution that the process starts from is therefore given by:
So the probability density p for a particular set of values xt is the sum of the value of each of the Gaussian kernel. Furthermore, suppose that the vital signs and their noisy observations at time t are denoted by xt and yt, respectively, we have:
y
t
=x
t
+v
t (2)
where vt is the additive and zero-mean noise. Based on the Bayesian framework, we are interested in the estimation of the posterior distribution of the vital signs denoted by p(xt|yt;) i.e. the probability of state xt given measurement yt for patient
. Suppose also that xt and yt obey the following Markovian state space:
x
t
˜p(xt|xt-1;) (3)
y
t
˜p(yt|xt;) (4)
i.e. that the values of xt are distributed according to p(xt|xt-1;) and the observations yt are distributed according to p(yt|xt;
). The posterior distribution, i.e. the patient-specific distribution, is assumed to be non-Gaussian, and thus N
(e.g. 1000 in this embodiment) Gaussian kernels, each with potentially a different weight wti are used to approximate this distribution:
p(xt|yt;)=Σt=1N
wti
(xt;xti,C
) (5)
where wti, i=1, . . . , N are kernel weights and
(xt; xti, C
) is the Gaussian function centered at xti with covariance matrix C
. Thus, instead of the Dirac delta function as in standard particle filter which is known as selective importance resampling (SIR), we use Gaussian kernels for approximation of posterior distribution, and therefore the regularised particle filter is employed. The regularised particle filter is the same as the SIR filter except it has a modified resampling technique to take into account the kernel used in the approximation of the posterior density [for more details see: Arulampalam, M. S.; Maskell, S.; Gordon, N.; Clapp, T., “A tutorial on particle filters for online nonlinear/non-Gaussian Bayesian tracking, “IEEE Transactions on Signal Processing, vol. 50, no. 2, pp. 174, 188, February 2002].
Based on the above definitions and assumptions, an algorithm using a generic particle filter may be given by Algorithm 1.
:
In this algorithm, π(xt|nt-1i,yk) is the importance density and the measurement transition probability is also given by:
p(yt|xti;)≈
(yt|0,Cv) (6)
where (yt|0, Cv) is the distribution of the additive noise vt, which has been assumed to be distributed independently and identically. In this embodiment, the importance density is chosen to be equal to the prior density π(xti|xt-1i,yt)≈p(xti|xt-1i), and therefore the update procedure for weights is simplified to (as in SIR filter):
w
i
i
∝w
t-1
i
p(yt|xti) (7)
Optionally, the state transition probabilities can be approximated by the population model:
p(xti|xt-1i;)≈p(xt|xt-1i;
) (8)
Here, we assume that the transition density is the same as the population density model given in equation (1). We may further approximate the state transition probability by:
p(xt|xt-1i;)≈
(xt;
,
) (6)
where =argmaxμ
(xt-1i;μj,
), for j=1, . . . ,
. That is to say, to sample xti from the population model, we sample from the kernel in the population model that is nearest to the particle xt-1i, or we consider the nearest kernel as the transition distribution. Equations (8) and (9) have the desirable effect that the patient-specific model's region of normality is always bounded by the population-based model. If a clinician decides, for a particular patient, that their region of normality lies outside the population-based region of normality, then this further approximation can be omitted.
Note that the measurement equation (2) is linear, but the transition of the state is non-Gaussian (nonlinear), since it is based on the population distribution, which is a non-Gaussian density. However, for a faster implementation, one may approximate the state transition by a Gaussian distribution and effectively employ the Kalman filter, which assumes Gaussian distribution for both initial and transition distributions. Based on the above assumptions the implementation the generic algorithm given in Algorithm 1 is converted to Algorithm 2.
particles (e.g.
= 1000) each
As in the standard particle filter, the algorithm has some free parameters including the covariance of the process noise, the covariance of the measurement noise (if we assume they are Gaussian) and the number of effective samples. These parameters can be set empirically, according to the available data set in hand. A simple approach is to set the process noise covariance equal to the covariance of the kernel in the patient-specific model, and to set the measurement noise covariance equal to the covariance of the kernel of the population model.
A numerical index such as novelty score, e.g. the Visensia index (VSI) used by the applicant, is a measure for the wellness of the patient, and in U.S. Pat. No. 7,031,857 B2 it was defined as a VSI score using equation (1) by:
VS(xt)=−log p(xt;
)+log p(
;
) (7)
where is the mean of p(xt;
). The term log p(
;
) is added to make VS
(xt) zero at the centre of the distribution. This same index is used in this embodiment as the first numerical index which is based on the population-based model. A threshold on this index defines the boundaries of the region regarded as normal and illustrated as the large ellipse in
in step 126. In step 128 an alert can be generated if VS
or a filtered version of it is outside a threshold, i.e. outside the large ellipse of
To obtain a second numerical index, based on the patient specific model, two approaches, as explained below, can be used to calculate the novelty score and therefore generate relevant alerts.
Similar to equation (10), a novelty (VSI) score for the patient-specific model can be defined (which sets the boundaries of the small ellipse in
VS(xt)=−log p(xt|yt;
)+log p(
;
) (8)
where is the mean of p(xt|yt;
). This definition sets the boundaries of the small ellipse in
Thus to obtain the second numerical index, as illustrated in ) inserting this into equation (11) in step 136. Again the second term based on the mean of the model is fixed (once the patient-specific model is frozen its mean would be recalculated when it is updated) and can be pre-calculated.
If the value of personalised VSI exceeds a threshold this is used to operate a status change notification in step 140 which can be displayed or transmitted to a clinician, who then determines whether the patient has deteriorated or improved. In the case of deterioration, as shown in arrow A in
It is envisaged that either type of notification would summon a clinician to make an assessment of what has changed. In the event of a type A alert, this could warrant escalation of the intervention, and will have therefore provided an earlier warning than the original system. In the event of a B type alert, the option exists for the clinician to authorise the system to re-engage the learning process, in order to track the patient's improvement.
In another scenario the patient state might change along a trajectory that leads to only small changes (+ or −) in VSI. In this case, the system can also be allowed to learn the patient's new state.
As a more sophisticated alternative, the VSI for personalised monitoring can consider both population and patient models, which leads to the second approach. By assuming that these distributions are independent, we may write:
p(xt;,
)∝p(xt|yt;
)p(xt;
) (9)
This leads to the definition of the personalised VSI for a specific patient :
VS(xt)∝−log p(xt;
,
)+log p(
;
,
) (10)
where is the mean of p(xt;
,
). Same as above, the term log p(
;
,
) has been added to make VS
(xt) zero at the centre (mean) of the distribution. It is clear that if the values of both population and patient probability density values decrease, which is a sign of deterioration, the VSI increases, whereas if the probability density values increase, which is a sign of improvement, the VSI decreases.
Based on equation (12), and according to the equations (1) and (5), after resampling we have:
Since the product of two Gaussian probability density functions (pdf) is Gaussian itself, this pdf is a Gaussian mixture model. By algebraic manipulation of equation (14), we obtain:
p(xt;,
)∝
(xt;
,
) (12)
where
=(
+
(
xti+
μj),
=(
+
)−1 (13)
If it is assumed that the covariance is spherical, the above mean basically is the average of the means of population and patient pdf weighted according to their variance. Therefore, the personalised Visensia index is defined by:
which can be used as a measure for the wellness of the patient.
Thus to obtain the second numerical index the incoming multi-parameter observation yt (HR, BP, BR, SpO2. Temp), is compared to the population model of equation (1) to read off its probability density p(xt;) with respect to the population and to the patient-specific model of equation (5) to obtain its probability density p(xt|yt;
) with respect to the patient's own status inserting this into equation (17) This is illustrated in step 135 in
Artefacts such as motion of probe detachment, or movement can provide spurious measurements and would result, generally, in a high instantaneous VSI value in the population model (which would normally be smoothed out to prevent false alerts). To stop updating the model based on the noisy measurements, the artefact-free data may be selected by clinicians or automatically. Automatic selection of the acquired data would be on the basis of the population model. For instance, acquired vital sign sets can be excluded by the following rules:
It may be the case that a patient's state is stable, but the vital signs are so abnormal that they would always generate a VSI alert, for example, if the heart rate is over 150, but the patient is deemed by a clinician to be “stable”. With the learning system proposed above, the specific model would never learn. However, the acceptance for online learning criterion can be modified by allowing the operator (e.g. clinician) to treat certain vital signs (in this case the Heart Rate) to be treated as “missing variables” for the purpose of accepting the measurements. A simple method for doing this is to replace that variable's value by the population mean value.
Therefore, in this approach when the vital signs are considerably small or large, the system stops updating the weights of kernels, as these parameters could be because of the artefact in the data set. This is not however a problem, in the updating example based on the particle filtering, the particles are sampled from the prior distribution.
In this section, we provide some of examples of applying the invention to real vital signs measurements.
The first two principal components of the centre of the kernels modelled by particles are shown again in
The novelty scores calculated with respect to the population based statistical model (equation 11) patient (equation 13) and combined (equation 17) are shown in
The invention may be embodied in a dedicated vital signs monitor or may be embodied in a computer system. In both cases the main parts of the system are, as illustrated in
Number | Date | Country | Kind |
---|---|---|---|
1408469.3 | May 2014 | GB | national |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/GB2015/050993 | 3/31/2015 | WO | 00 |