GB2537844A

GB2537844A - Electrode-skin contact quality assessment

Info

Publication number: GB2537844A
Application number: GB1507139.2A
Authority: GB
Inventors: Ang Su-Shin; Hernandez-Silveira Miguel
Original assignee: SENSIUM HEALTHCARE Ltd
Current assignee: SENSIUM HEALTHCARE Ltd
Priority date: 2015-04-27
Filing date: 2015-04-27
Publication date: 2016-11-02
Anticipated expiration: 2035-04-27
Also published as: GB2537844B; GB201507139D0

Abstract

A method of identifying poor electrical contact between a patients skin and the electrodes of a patch used for real-time monitoring of a time-varying physiological signal such as an ECG. The method comprises extracting features from a physiological signal that can be used to classify a time portion of the monitored time-varying physiological signal for the purpose of providing estimated probabilities that poor electrical contact exists between the patients skin and the electrodes. The method further requires receiving the monitored time-varying physiological signal and, for each sequential time portion, determining a value for each feature. Past and present feature values are then used to sequentially update a probability model to determine the likely cause of the artefacts. A notification is further provided when one or more of said probabilities is above a pre-determined threshold.

Description

ELECTRODE-SKIN CONTACT QUALITY ASSESSMENT

Technical Field

The present invention relates to a method and apparatus for assessing the quality of electrical contact between human skin and an electrode used for monitoring a time-varying physiological signal.

Background

The quality of skin-electrode contact is crucial in clinical monitoring of physiological signals including, for example, measurement of heart rate obtained through electrocardiography and respiration derived from impedance pneumography. When the electrical contact is poor, the impedance between the skin and the electrodes is high (on the order of mega ohms). High impedance not only affects the shape and magnitude of the signals, but also increases the sensitivity of the monitoring system to artefacts resulting from motion of the human subject, which can lead to further corruption of the measured signals. Factors such as improper electrode application, regrowth of skin cells following electrode application, and electrode decline over time are responsible for bad skin-electrode contact.

Typically, the quality of skin-electrode contact is monitored by periodically injecting a small test current through the electrodes, thereby allowing direct determination of the impedance. However, this requires additional electronics, including a current source, which becomes impractical in settings where space for components and available power is limited, e.g. within a patient monitoring patch.

There exists a need for a reliable method of and apparatus for assessing the quality of skin-electrode contact based only on analysis of the recorded physiological signals.

Summary

According to the present invention there is provided a method of identifying poor electrical contact between a patient's skin and electrodes of a patch used for real-time monitoring of a time-varying physiological signal f(t). The method comprises defining one or more features that can be used to classify a time portion of the monitored time-varying physiological signal for the purpose of providing estimated probabilities P that poor electrical contact exists between the patient's skin and the electrodes, receiving the monitored time-varying physiological signal f(t) and, for each sequential time portion t, determining a value Ft for the or each feature, using the past and present feature values F in a computer algorithm to provide updated estimated probabilities P, and providing a notification when one or more of said probabilities is above a predetermined threshold.

Brief Description of the Drawings

Figure 1 illustrates a system for monitoring a physiological signal; Figure 2 shows a concept graphical user interface; Figure 3 illustrates four states describing the combined patient-patch system; Figure 4 shows an example ECG trace and its frequency spectrum; Figure 5 shows an example ECG trace and its frequency spectrum where the ECG trace contains significant artefacts, due to motion of the patient for example; Figure 6 is a block diagram of a method according to an embodiment of the invention; Figure 7 is a block diagram of an algorithm for using classifiers to calculate the likelihood of each state of the combined patient-patch system; and Figure 8 illustrates the topology of an artificial neural network; and Figure 9 illustrates an alternative topology of a method according to an embodiment of the invention.

Detailed Description

It is desirable to be able to assess the quality of skin-electrode contact by analysing one or more physiological signals, denoted f(t), recorded as part of a normal patient monitoring process, without relying on a separate, dedicated system designed specifically to perform this function, e.g. by directly measuring the impedance of the skin-electrode contact over time. Relinquishing the requirement for a dedicated system for monitoring the skin-electrode contact is advantageous in scenarios where it is sought to miniaturize a patient monitoring apparatus. In such scenarios, each additional electronic component is an additional drain on a limited power supply and must be carefully budgeted for.

Figure 1 illustrates a patient monitoring system including a wearable patch 2. The patch monitors physiological parameters such as heart rate, respiration rate and temperature. An example patch may be based on the SensiumVitalsTm ultra-low power wireless system for monitoring patient vital signs. The patient monitoring patch transmits this data wirelessly 3 to a base station 4. The data is then transmitted over a network 5 to a server computer system 7 for further processing. Results of the data analysis are displayed on a computer terminal 6, for viewing by a healthcare professional, for example. Performing the processing on a server computer has the advantage that it incurs no additional drain on the limited resources within a portable patient monitoring system. Alternatively, the processing may be done directly on the portable patient monitoring system, using for example digital signal processing (DSP).

The present approach provides a computer algorithm which captures and processes physiological signals, f(t), in order to determine the quality of the skin-electrode contact. Physiological signals, such as those obtained through electrocardiography or impedance pneumography, are captured and stored in a continuous manner over time.

In the case of electrocardiography, the captured physiological signals correspond to an electrocardiogram (ECG). Features extracted from captured physiological signals are fed to a computer algorithm which sequentially updates a model based on the feature history and certain initial assumptions to determine the likely cause of artefacts present within the physiological signal.

The likelihood that artefacts present in the physiological signal are due to poor skin-electrode contact is calculated, as well as the likelihood that the artefacts are due to motion of the patient. If the likelihood that the artefacts are due to poor skin-electrode contact is above a pre-determined threshold, action may be taken such as sending a notification to nurses who can then check if electrode application has been done properly on the patient, and then re-apply if necessary. The likelihood that artefacts are due to a pathological condition exhibited by the patient may also be calculated.

An example of a notification displayed on a graphical user interface (GUI) 8 is shown in Figure 2, where a "traffic light" representation is used to display the status of the skin-electrode contact. A red light 10 indicates to a nurse that attention is required. On the other hand, if the signal corruption is brief and most likely due to motion of the patient, then no notification should be sent, e.g. a green light 9 may be displayed. This has the advantage of avoiding the potential frequent false alarms, i.e. notifications being sent to nurses that a patch is incorrectly applied which are in fact triggered by human motion momentarily causing artefacts in the physiological signal. The GUI may also display in real time the captured physiological data such as heart rate 11, respiration rate 12 and patient temperature 13.

Considering in more detail the monitoring algorithm, let M denote the state corresponding to the patient undergoing motion and M' denote the state where the patient is not undergoing motion. Furthermore, let P denote that the electrical contact between the skin and electrode is good and P' denote that it is bad. At any point in time there are four possible combinations which describe the combined state of the patient-patch system within this framework, as illustrated in Figure 3: M'P 14, M'P' 15, MP' 16, and MP 17. In general one is interested in cases where the skin-electrode contact is poor, irrespective of the motion state of the patient, since in this case the electrodes require reapplication which must be performed by a trained person. The states of interest for taking further action are thus MP' and M'P', denoted by the box 18 in Figure 3. If the state is either M'P or MP then no action need be taken. As well as considering motion of the patient and application quality of the patch, additional states may be considered -such as whether the patient is exhibiting a physiological condition.

At any one time it is difficult to determine the current combination taken from the set {M'P, MP', MP, M'P') based only on the current characteristics of the physiological signal, e.g. the electrocardiogram. The hidden nature of these states implies that this scenario can be modelled as a Hidden Markov Model (HMM). At each step the algorithm calculates the conditional probability for each of the four combinations, based on the observed feature history, denoted F. The algorithm thus learns from past events to make predictions about the current state of the system.

Having briefly outlined the conceptual operation of the algorithm, an example will now be given based on an ECG. After acquisition of the ECG signal using an analogue front-end sensor, the signal is discretized using an analogue-to-digital converter (ADC) module. The algorithm is applied to discrete blocks of ECG signals, each 30 seconds in duration. A number of features are computed from the raw ECG signal. These include, but are not restricted to, the harmonic energy, signal-to-noise ratio, saturation ratio, and the energy of the QRS complex.

An exemplary filtered ECG signal 19 and the spectrum of the signal 20 are shown in Figure 4. The original ECG signal has been filtered using infinite impulse response (IIR) filters designed to attenuate noise outside the 8 to 16 Hz band, as well as other modules such as the first order differential and moving average operators. The spectrum of the filtered signal consists of the superposition of multiple harmonics 21a, 21b, 21c, 21d, where the frequency of the first harmonic 21a corresponds to the heart rate. The aggregate spectral amplitude of the inter-harmonic peaks 22a, 22b, 22c etc. are good indicators of the amount of ECG signal artefacts.

Figure 5 illustrates a very noisy ECG signal 23, for which the harmonics are often indistinguishable from the spectral peaks resulting from noise. The spectrum 24 of this signal is characterised by a lack of uniformity in inter-harmonic widths and a large aggregate spectral amplitude between inter-harmonic peaks.

An ECG signal includes a sequence of QRS complexes, each of which corresponds to the electrical activity associated with a heart beat. The frequency band of the QRS complex within the ECG signal is typically between 1 and 40 Hz. In order to determine changes in signal characteristics, the original ECG signal is segmented into two sub-blocks of equal lengths (15 seconds in duration). The energies of the in-band and out-of-band spectral components are separately quantified. Subsequently, changes in QRS characteristics are quantified by taking the ratio of the QRS energies between the first and second halves of the signal (QRS ratio), the Signal-to-Noise ratio, and the sum of the spectral components corresponding to the QRS complex in the first and second halves of the signal (QRS energy). These parameters are thus "features" or characteristics, which are calculated and recorded in a sequential manner and contribute to the observed feature history, denoted F. The algorithm for calculating the most likely state of the patient-patch system will now be described in detail. The forward probability a that the algorithm is at a given state Sk={M'P, MP', MP, M'P) is given as follows: cro(Cio) = PSioFc, at(Sit) =Icrt_i(Skt_i)Fskt_isitPS Ft k =0 By similar means, the backward probability can be derived. However, this requires knowledge of future feature values and outputs, which are unknown. Consequently, the backward probability is left out of the calculation of the conditional probability involving past and present features, and the conditional probability can be described as follows: a(Si) Ft(SitiFo, *** Ft) vc_o at(Skt) Subsequently, the most probable state can be identified by selecting the state with the highest estimated probability according to equation (3). In order to obtain these parameters, the transition probability,P -sict-isiti as well as the posterior probability, PSitFt, needs to be known. The default values for transition probabilities can be obtained using annotated data, and updated in each time step using the Baum-Welch algorithm Figure 6 shows an overview flow chart of the exemplary method. First, the algorithm is initialized, e.g. by populating the transition matrix with default values 25. Once this step is complete, the algorithm can commence processing of physiological data, in this example ECG epochs. The method broadly comprises four distinct steps which are performed in a loop 31. First an ECG epoch is acquired and converted to an analogue signal using an analogue-to-digital converter (ADC 26). The acquired ECG epoch, which typically has a duration of 30 seconds, is then processed on a computer to extract the desired feature values 27, e.g. ORS ratio. The extracted feature values are then fed to a classification module 28 to calculate the forward state probabilities and hence determine the most likely state 29 of the patient-patch system. The calculated forward probabilities are subsequently used to update the transition probability matrix (e.g. using the Baum-Welch algorithm) 30. The updated transition probability matrix then serves as the starting point for the next iteration 31. The transition probability (3) matrix contains approximations of the probabilities of transition at any one instance from one state to another state in the subsequent time step. With each iteration of the algorithm the transition probability matrix becomes more accurate and hence predictions of the state become more accurate.

The classification module 32 is shown in more detail in Figure 7. Given an annotated dataset with N observations and N corresponding labels, the prior distribution, i.e. P(M'P), P(MP), P(MP), and P(MP), can be simply derived by analysing the distribution of the labels in the annotated dataset. The posterior distribution, P(SkIF), can be similarly derived by analysing the labels in conjunction with the feature values.

In theory, a multi-dimensional histogram can be derived from this dataset and stored in a look-up table. During run-time, the relevant 'slices' of the histogram corresponding to the appropriate state and features can be retrieved to work out the conditional probabilities. However, the storage capacity and complexity of this approach does not scale well with the number of features and its corresponding resolution, and practical implementation in real-time may be difficult.

A less complex solution can be achieved through the use of a classification function (i.e. classifier 33a, 33b, 33c, 33d) that is capable of producing a continuous output between 0 and 1, in order to approximate these conditional probabilities. In doing so, the function parameters, rather than the posterior probability distribution, need only be stored. A one-layer Artificial Neural Network (ANN) with M hidden units is described according to equation (4).

yk(F,w) = if I w0-(1w,(Fi)1, {ic E ZIO k 3} (4) J=0 I=0 o-(a) = 1 + e -a 1 (4a) The sigmoidal function of equation (4a), a(a) , is a non-linear and differentiable function, that produces an output between 0 and 1. The back-propagation algorithm can be used to determine the weight matrix w in order to minimise the objective error function. In this case, yk(F,w) = (SkiF). The topology of the network is shown in Figure 8.

An artificial neural network (ANN) consists of 3 types of layers: the input layer 35, the hidden layers 35, and the output layer 36. An optional transformation step is included in order to change the dimensionality of the feature space, and improve the separability of the classification problem. Otherwise, the raw features are first normalized before being fed to the network. Range normalization or Gaussian normalization (equations (5) and (6) respectively) are normally used. The normalization step is carried so as not to favour any of the features used in the classification problem. Normalization may be performed according to equation 5 or 6 below: f' = -fmin /max -rain f-[mean I' In (5) and (6), the terms f fmea" and a refer to the raw feature value, minimum feature value, maximum feature value, mean feature value, and the standard deviation respectively.

Four posterior probabilities are required in the present example, namely P(WPIF), P(MP'IF), P(MPIF), and P(M'P'IF). Two possible topologies are proposed: four ANNs with one output node each, or a single ANN with four output nodes. A single ANN implies that weights will be shared across the derivation of different posterior probabilities.

The basic node of an ANN node consists of a linear sum of weights, w', and inputs (these inputs are features if the node is resident in the input layer, or from all other nodes in the previous layer if the node is resident in the hidden layer). The subsequent output is fed to a sigmoid function (4), which results in a value between 0 and /. In order to 'train' the network/s to become an effective classifier, input features with known labels (or outputs) are fed to the network. Error back-propagation using the gradient descent algorithm is then used to adjust the weights in order to minimize the classification error. There are various ways to organize the training procedure. One possible method is k-fold cross validation.

Although methods for the quantification of signal corruption are likely to be similar in different states, motion and improper patch application may corrupt the signal in subtly different ways. As such, it makes sense to make use of different classifiers in different states in order to calculate the posterior probabilities in the given context. Therefore, separate networks can be maintained in different states, rather than combining them together in the manner shown in Figure 8.

The characteristics of corrupted biomedical signals are similar regardless of the source of noise, e.g. motion artefacts or improper application of the electrodes, saturation and low signal-to-noise ratio. For that reason, classifiers used in different states are likely to share many features, some of them listed below. Note that this algorithm is not restricted to this particular set of features, and the feature set can be easily extended to improve classification accuracy. A non-exhaustive list of potential features is given below.

* Saturation proportion: the proportion of the raw ECG signal that is saturated.

* Absolute QRS energy: the sum of the spectral components corresponding to the ORS complex.

* ORS ratio: the ratio of the ORS energies between the first and second halves of an acquired ECG signal (see, for example, Figure 4).

* Frequency-based heart rate variability features: whereby a power spectral density analysis is applied to a tachogram (a set of peak-to-peak intervals derived from the ECG signal).

* Impedance leads-off notification proportion.

* Proportion of invalid heart rates.

* Coefficient of variation among valid heart rates.

* Time-based metrics for quantifying variability in RR intervals including Mean Absolute Difference (MAD), Average Absolute Difference (AAD), RMSSD (Root Mean Square of Successive Differences).

Besides the use of features, other traditional feature mapping methods for improving separability in the feature space, such as dimensionality reduction (e.g. Sammon Mapping, principal component analysis and linear discriminant analysis) or different kernel methods can be applied.

An alternative topology is shown in Figure 9. Two sets of features are calculated based on filtered 37 and un-filtered 38 ECG signals. After normalization, the calculated feature values 39 are passed on to an ANN 40 for the estimation of the four posterior probabilities -P(A4'9F), P(MP'IF), P(M'P'IF), and P(MPIF).

A subset of the current calculated feature sets, and previous posterior probabilities are fed into a second ANN 41 network in order to calculate the conditional probabilities of these four states, based on past and present feature values -P(MPIFt P(MP1Ft Frm), P(M'ElFt FrA, and P(MPIFt an). These estimated probabilities present a better estimation for the persistence and extent of these artefacts. More specifically, these probabilities are stored in a 'register bank' 41 for use in subsequent cycles. A multiplexer 43 is used to switch between the outputs of the first and second ANNs in order to take into account the initial latency of the second ANN.

Although the invention has been described in terms of preferred embodiments as set forth above, it should be understood that these embodiments are illustrative only and that the claims are not limited to those embodiments. Those skilled in the art will be able to make modifications and alternatives in view of the disclosure which are contemplated as falling within the scope of the appended claims. Each feature disclosed or illustrated in the present specification may be incorporated in the invention, whether alone or in any appropriate combination with any other feature disclosed or illustrated herein.

Claims

CLAIMS: 1. A method of identifying poor electrical contact between a patient's skin and electrodes of a patch used for real-time monitoring of a time-varying physiological signal f(t), the method comprising: defining one or more features that can be used to classify a time portion of the monitored time-varying physiological signal for the purpose of providing estimated probabilities P that poor electrical contact exists between the patient's skin and the electrodes; receiving the monitored time-varying physiological signal f(t) and, for each sequential time portion t, determining a value Ft for the or each feature; using the past and present feature values F in a computer algorithm to provide updated estimated probabilities P; and providing a notification when one or more of said probabilities is above a pre-determined threshold.
2. The method according to claim 1 wherein the combined patient-patch system is described by a state Sk, k={0,1,2,3), wherein at any one time the state is one of: patient undergoing motion and skin-patch contact good (S0=MP); or patient not undergoing motion and skin-patch contact good (S1=M'P); or patient undergoing motion and skin-patch contact poor (S2=MP); or patient not undergoing motion and skin-patch contact poor (S3=M'P).
3 The method according to claim 2 wherein each state Sk is additionally combined with the possibility that the patient is either exhibiting a pathological condition (C) or not exhibiting a pathological condition (C), such that at any one time the state is one of: patient undergoing motion, skin-patch contact good, no pathological condition (MPC) or patient not undergoing motion, skin-patch contact good, no pathological condition (M'PC); or patient undergoing motion, skin-patch contact poor, no pathological condition (MP'C); or patient not undergoing motion, skin-patch contact poor, no pathological condition (M'P'C).patient undergoing motion, skin-patch contact good, pathological condition (MPC); or patient not undergoing motion, skin-patch contact good, pathological condition (M'PC); or patient undergoing motion, skin-patch contact poor, pathological condition (MP'C); or patient not undergoing motion, skin-patch contact poor, pathological condition (M'P'C).
4. The method according to claim 2 or 3 wherein the estimated probabilities P are estimated conditional probabilities P(SkIF) for states Sk given past and present feature values F.
5. The method according to claim 4 wherein the computer algorithm comprises an artificial neural network for estimating the conditional probabilities P(SkIF) of different states Sk of the patient-patch system.
6. The method according to any preceding claim wherein the time-varying physiological signal is an electrocardiogram (ECG). 20 7. The method according to claim 6 wherein the features used to classify a time portion of the monitored time-varying physiological signal are selected from the group comprising: saturation proportion (as herein defined); absolute ORS energy; ORS ratio; frequency-based heart rate variability features; impedance leads-off notification proportion; proportion of invalid heart rates; coefficient of variation among valid heart rates.8. The method according to any one of claims 1 to 5 wherein the time-varying physiological signal is an impedance pneumograph.9. A computer algorithm executed on a server computer system for carrying out the method according to any one of the preceding claims.Amendments to the Claims have been filed as follows: CLAIMS: 1. A method of identifying poor electrical contact between a patient's skin and electrodes of a patch used for real-time monitoring of a time-varying physiological signal f(t), the method comprising: defining one or more features that can be used to classify a time portion of the monitored time-varying physiological signal for the purpose of providing estimated probabilities P that poor electrical contact exists between the patient's skin and the electrodes; receiving the monitored time-varying physiological signal f(t) and, for each sequential time portion t, determining a value Ft for the or each feature; using the past and present feature values F in a computer learning algorithm to provide updated estimated probabilities P; and providing a notification when one or more of said probabilities is above a pre-cr) 15 determined threshold; wherein the combined patient-patch system is described by a state Sk, CO k={0,1,2,3}, wherein at any one time the state is one of: patient undergoing motion and skin-patch contact good (S3=MP); or r 20 patient not undergoing motion and skin-patch contact good (S1=M'P); or patient undergoing motion and skin-patch contact poor (S2=MP); or patient not undergoing motion and skin-patch contact poor (S3=M'P); and wherein the estimated probabilities P are estimated conditional probabilities P(SkIF) for states Sk given past and present feature values F. 2 The method according to claim 1 wherein each state Sk is additionally combined with the possibility that the patient is either exhibiting a pathological condition (C) or not exhibiting a pathological condition (C), such that at any one time the state is one of: patient undergoing motion, skin-patch contact good, no pathological condition (MPC) or patient not undergoing motion, skin-patch contact good, no pathological condition (M'PC); or patient undergoing motion, skin-patch contact poor, no pathological condition (MP'C); or patient not undergoing motion, skin-patch contact poor, no pathological condition (M'P'C).patient undergoing motion, skin-patch contact good, pathological condition (MPC); or patient not undergoing motion, skin-patch contact good, pathological condition (M'PC); or patient undergoing motion, skin-patch contact poor, pathological condition (MP'C); or patient not undergoing motion, skin-patch contact poor, pathological condition (M'P'C) 3. The method according to claim 1 or 2 wherein the computer learning algorithm comprises an artificial neural network for estimating the conditional probabilities P(SkIF) of different states Sk of the patient-patch system.(r) 4. The method according to any preceding claim wherein the time-varying physiological signal is an electrocardiogram (ECG).COCD 5. The method according to claim 4 wherein the features used to classify a time Nt 20 portion of the monitored time-varying physiological signal are selected from the group comprising: saturation proportion (as herein defined); absolute QRS energy; ORS ratio; frequency-based heart rate variability features; impedance leads-off notification proportion; proportion of invalid heart rates; coefficient of variation among valid heart rates.6. The method according to any one of claims 1 to 3 wherein the time-varying physiological signal is an impedance pneumograph.
7. A computer algorithm executed on a server computer system for carrying out the method according to any one of the preceding claims.