US20100016743A1

US20100016743A1 - Identifying Groups of Patients with Similar Physiological Characteristics and Risk Profiles

Info

Publication number: US20100016743A1
Application number: US12/504,543
Authority: US
Inventors: Zeeshan H. Syed; John V. Guttag; Collin M. Stultz
Original assignee: Massachusetts Institute of Technology
Current assignee: Massachusetts Institute of Technology
Priority date: 2008-07-17
Filing date: 2009-07-16
Publication date: 2010-01-21
Also published as: US20100016748A1; US8340746B2

Abstract

The invention relates in part to methods for partitioning a plurality of patients into risk profile groups comprising the steps of: recording a physiological signal from each patient of a plurality of patients; segmenting the physiological signal into a plurality of components for each patient of a plurality of patients; grouping the components into a plurality of information classes for each patient of a plurality of patients; assigning a representation to each information class for each patient of a plurality of patients; and grouping the patients in response to the representations of their respective information classes.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application claims priority to and the benefit of U.S. Provisional Patent Application No. 61/081,445, filed Jul. 17, 2008 and U.S. Provisional Patent Application No. 61/081,437, filed Jul. 17, 2008, the entire disclosures of each of which are hereby incorporated by reference herein.

BACKGROUND

Many different techniques exist for the assessment of cardiac risk using information in the ECG signal. Measures such as heart rate variability, heart rate turbulence, t-wave alternans and morphologic variability have been shown to be good risk-stratifiers for future cardiovascular events and focus on calculating a particular feature from the raw ECG signal to rank patients. However, current techniques are typically insufficient for predicting future heath and risk of mortality.

SUMMARY OF THE INVENTION

In one aspect the invention relates to an approach to estimate the difference in electrocardiographic activity recorded during the long-term monitoring of two patients. In one embodiment this metric is termed the electrocardiographic mismatch (EM) between the patients. In one embodiment EM is computed by first creating a symbolic representation of the electrocardiographic signal for each patient, and by then carrying out a weighted inter-patient comparison of the symbol distributions. The resulting electrocardiographic mismatch value serves as an indicator of how different the patients are electrocardiographically.
In one aspect the EM is used to partition a population of patients into sub-groups comprising individuals with similar cardiac characteristics and risk profiles. This is done by using clustering to group together patients with low EM values relative to each other.
When evaluated on a population of 686 patients who had suffered from non-ST elevation acute coronary syndromes, hierarchical clustering with EM was able to identify abnormal patients at increased risk of death and myocardial infarction over a 90 day follow-up period. The 20% patients most different from the majority cluster were at 5.3 times increased risk of death (p-value: 0.003) relative to other members of the population, and had a 2.8 times increased risk for the combined endpoint of death and myocardial infarction (p-value: 0.003).
In one aspect, the invention relates to a method of partitioning a plurality of patients into risk profile groups. The method can include the steps of recording a physiological signal from each patient of a plurality of patients; dividing each physiological signal into a plurality of equivalent time portions for each patient of the plurality of patients; assigning a representation to each portion of the plurality of equivalent time portions for each respective physiological signal for each respective patient of the plurality of patients; and grouping the patients in response to the representations of their respective physiological signals. Physiological signals include, for example, physiological signals. In some embodiments of the method, the representation is a numerical value. In some embodiments of the method, the representation is a symbol. In some embodiments, the physiological signal is an ECG and the equivalent time portion is a heartbeat.
In one aspect, the invention relates to a method of partitioning a plurality of patients into risk profile groups. The method includes the steps of recording a physiological signal from each patient of a plurality of patients; segmenting the physiological signal into a plurality of components for each patient of a plurality of patients; grouping the components into a plurality of information classes for each patient of a plurality of patients; assigning a representation to each information class for each patient of a plurality of patients; and grouping the patients in response to the representations of their respective information classes. In some embodiments of the method, the representation is a numerical value. In some embodiments of the method, the representation is a symbol. In some embodiments, the physiological signal is an ECG and the equivalent time portion is a heartbeat. In some embodiments, the representation is a waveform such as, for example, a prototype (archetype) waveform or a centrotype waveform.
In some embodiments, the grouping step includes measuring an electrocardiographic mismatch between each pair of patients in the plurality of patients; assigning each patient of the plurality of patients to a respective cluster of a plurality of clusters; grouping clusters in response to the electrocardiographic mismatch until the number of clusters reaches a predefined minimum; and assigning risk outcomes to each of the clusters. In some embodiments, the grouping of clusters is performed hierarchically. In some embodiments, the grouping of clusters is performed using, for example, fuzzy clustering, max-min clustering, k-means clustering, or svm clustering.
In one aspect, the invention provides a method of partitioning a plurality of patients into risk profile groups. The method includes the steps of recording an ECG signal from each patient of a plurality of patients; dividing each ECG signal into a plurality of heartbeats for each patient of the plurality of patients; grouping heartbeats into clusters for each patient of the plurality of patients; assigning a representation to each cluster for each patient of the plurality of patients; assigning a value to differences in the representation of clusters for each respective ECG signal for each respective patient of the plurality of patients; and grouping the patients in response to the values of their respective differences in ECG signals.
In one aspect, the invention relates to a method of partitioning a plurality of patients into risk profile groups. The method includes the steps of recording an physiological signal from each patient of a plurality of patients; dividing each physiological signal into a plurality of equivalent time portions for each patient of the plurality of patients; grouping time portions into clusters for each patient of the plurality of patients; assigning a representation to each cluster for each patient of the plurality of patients; assigning a value to differences in the representation of clusters for each respective physiological signals signal for each respective patient of the plurality of patients; and grouping the patients in response to the values of their respective differences in physiological signals.
In another aspect, the invention relates to a method of assigning a risk score to new patients by matching them to an existing database of patients. The method includes the steps of grouping patients in the database using the methods described herein; segmenting the physiological signal into a plurality of components for each new patient; grouping the components into a plurality of information classes for each new patient; assigning a representation to each information class for each new patient; matching the representations of new patients with representations of groups of patients from the database; and assigning new patients the risk characteristics of patients in the groups of patients in the database with matching representations.

BRIEF DESCRIPTION OF THE DRAWINGS

The aspects, embodiments, and features of the invention can be better understood with reference to the drawings described herein. The drawings are provided to highlight specific embodiments of the invention and are not intended to limit the invention, the scope of which is defined by the claims.

FIG. 1 shows Kaplan-Meier survival curves for (a) Death, (b) MI and (c) Death/MI. The survival curves of the low risk group (n=460) shown as a dashed line with the survival curves of the high risk group (n=226) shown as a solid line, in accordance with an illustrative embodiment of the invention.

FIG. 2 is a graph showing that event rates in high risk population as cutoff is varied, in accordance with an illustrative embodiment of the invention.

DESCRIPTION OF A PREFERRED EMBODIMENT

These and other aspects, embodiments, and features of the invention are further described in the following sections of the application, which are provided to highlight specific embodiments of the invention and are not intended to limit the invention. Other embodiments are possible and modifications may be made without departing from the spirit and scope of the invention. In addition, the use of sections in the application is not meant to limit the invention; each section can apply to any aspect, embodiment, or feature of the invention.
It should be understood that the order of the steps of the methods of the invention is immaterial so long as the invention remains operable. Moreover, two or more steps may be conducted simultaneously or in a different order than recited herein unless otherwise specified.
There is considerable evidence to suggest that information in the electrocardiographic (ECG) signal may have prognostic value for cardiac patients. Techniques such as heart rate variability, heart rate turbulence, t-wave alternans and morphologic variability, have all been shown to be good risk-stratifiers for future cardiovascular events in the setting of a prior acute coronary syndrome. The focus of these methods is to calculate a particular feature from the raw ECG signal, and to use it to rank patients along a risk scale.
This invention discussed herein relates to a comparative approach to identify abnormal patients at increased risk of adverse cardiovascular outcomes. In contrast to deriving a feature from the ECG for each patient, we directly compare the signals for every pair of patients to determine how different they are, i.e., how much electrocardiographic mismatch exists. Patients with abnormal cardiac characteristics then correspond to those individuals whose long-term electrocardiogram did not match a dominant group in the population.
The invention relates to a means to obtain a quantifiable comparison of how different two patients are electrocardiographically. The focus of the invention is to use this information to partition patients into similar groups, with matching long-term electrocardiograms. The underlying hypothesis here is that patients with ECG signals that match in morphology and dynamics will have consistent risk profiles. This allows one to obtain a more fine-grained understanding of how a patient's health will evolve over time, and more accurately assign a risk score to the patient for events such as death and myocardial infarction.
The application is organized as follows. Section 2 describes how electrocardiographic mismatch is computed through comparisons of symbolic distributions of ECG. Section 3 details a hierarchical clustering approach that is able to partition patients into groups with different risk profiles for future cardiovascular events. Section 4 discusses evaluation and Section 5 presents the results of this study. Section 6 compares the results to related efforts. Section 7 presents a discussion and conclusions.

Section 2: Electrocardiographic Mismatch

The electrocardiographic mismatch (EM) between two patients, p and q, is calculated using a two-step process.

Symbolization

As a first step, the ECG signal for each patient is symbolized using the techniques previously described (Syed Z, Guttag J, Stultz C. Clustering and symbolic analysis of cardiovascular signals: discovery and visualization of medically relevant patterns in long-term data with limited prior knowledge. EURASIP Journal on Applied Signal Processing, 2007 b). Symbolization involves segmenting the original ECG signal into heart beats, and then separating the beats into different groups based on their morphology and assigning a representation to each beat. In one embodiment the representation is a symbol. In another embodiment the representation is value. This method is disclosed in detail in provisional application No. 61/081,437 (attorney docket number MIT-184PR) by Syed et al., entitled Motif Discovery in Physiological Datasets: A Methodology for Inferring Predictive Elements; filed on Jul. 17, 2008; assigned to the same party of record as this application; and incorporated herein in its entirety.
The process of comparing morphology between beats is carried out using a dynamic time-warping (DTW) algorithm. (Myers C, Rabiner L. A comparative study of several dynamic time-warping algorithms for connected word recognition. The Bell System Technical Journal. (1981) 60:1389-1409.) Given two beats, x₁and x₂, of length l₁and L₂respectively, DTW produces the optimal alignment of the two sequences by first constructing an l₁-by-l₂distance matrix d. Each entry (i,j) in this matrix represents the square of the difference between samples x₁[i] and X₂[j]. A particular alignment then corresponds to a path, φ, through the distance matrix of the form:
φ(k)=(φ₁(k),φ₂(k)), 1≦k≦K (1)
where φ₁and φ₂represent row and column indices into the distance matrix, and K is the alignment length.
The optimal alignment produced by DTW minimizes the overall cost:
$\begin{matrix} C (x_{1}, x_{2}) = \min_{ϕ} C_{ϕ} (x_{1}, x_{2}) & (2) \end{matrix}$
where C_φ is the total cost of the alignment path φ and is defined as:
$\begin{matrix} C_{ϕ} (x_{1}, x_{2}) = \sum_{k = 1}^{K} d (x_{1} [ϕ_{1} (k)], x_{2} [ϕ_{2} (k)]) & (3) \end{matrix}$
The search for the optimal path is carried out in an efficient manner using dynamic programming. (Cormen T, Leiserson C, Rivest R, Stein C. Introduction to Algorithms. MIT Press and McGraw-Hill. 2001; 2^nded.) The final energy difference between the two beats x₁and x₂, is given by the cost of their optimal alignment, and depends on both the amplitude differences between the two signals, as well as the length K of the alignment (which increases if the two beats differ in their timing characteristics).
In this way, the DTW approach described here measures changes in morphology resulting from amplitude and timing differences between the two beats. Using this information, beats with distinct morphologies are placed in different groups and each group is assigned a unique label or symbol. Additional description of the symbolization process is provided in Syed Z, Guttag J, Stultz C. Clustering and symbolic analysis of cardiovascular signals: discovery and visualization of medically relevant patterns in long-term data with limited prior knowledge. EURASIP Journal on Applied Signal Processing, 2007 b. The final result of this step is that the original electrocardiogram is transformed from raw samples to a sequence of symbols.

Comparing Symbol Distributions

Denoting the set of symbols for patient p as S_pand the set of probabilities with which these symbols occur in the electrocardiogram as P_p(for patient q an analogous representation is adopted), we calculate the EM between these patients as:
$\begin{matrix} {EM}_{p, q} = \sum_{a \in S_{p}} \sum_{b \in S_{q}} D T W (a, b) P_{p} [a] P_{q} [b] & (4) \end{matrix}$
In (4), DTW(a,b) corresponds to the dynamic time-warping cost of aligning symbols a and b.
Intuitively, the electrocardiographic mismatch between patients p and q corresponds to an estimate of the expected dynamic time-warping cost of aligning any two randomly chosen beats from these patients. The EM calculation in (4) achieves this by weighting the cost between every pair of symbols between the patients by the probabilities with which these symbols occur.

Section 3. Hierarchical Clustering of Patients Using EM

For every pair of patients in a population, the electrocardiographic mismatch between them is computed using the techniques described in Section 2. The resulting divergence matrix, D, relating the pairwise electrocardiographic mismatches between all the patients is used to partition the population into groups with resembling cardiac characteristics. In one embodiment this process is carried out by means of hierarchical clustering. (Duda R, Hart P. Pattern Classification. Wiley-Interscience. 2000; 2^nded.)
Hierarchical clustering starts out by assigning each patient to a cluster of its own. It then proceeds to combine two clusters at every iteration and terminates when all the patients in the population have been amalgamated within a single cluster. A number of different criteria can be used to determine which two clusters should be combined at each step. In our work, we choose the unweighted average linkage (UPGMA) criterion, which corresponds to merging the two clusters A and B for which the mean electrocardiographic mismatch between the elements of the clusters is minimized, i.e., we merge clusters A and B such that they minimize:
$\begin{matrix} \frac{1}{\langle A \rangle \cdot \langle B \rangle} \sum_{x \in A} \sum_{y \in B} {EM}_{x, y} & (5) \end{matrix}$
To obtain a clustering at a specific precision, the hierarchical clustering process can be terminated when the original patient population has been reduced to a given number of clusters. Other clustering methods known to the prior art are also contemplated for this purpose. For example, the grouping of clusters can be performed performed using fuzzy clustering, max-min clustering, k-means clustering, or svm clustering.

Section 4. Evaluation

The population used for this work comprised patients in the TIMI DISPERSE2 trial (Cannon C, Husted S, Harrington R, Scirica B, Emanuelsson H, Peters G, Storey R. Safety, tolerability, and initial efficacy of AZD6140, the first reversible oral adenosine diphosphate receptor antagonist, compared with clopidogrel, in patients with non-ST-segment elevation acute coronary syndrome: primary results of the DISPERSE-2 trial. J Am Coll Cardiol. 2007;50:1844-51), who were admitted to a hospital with non-ST-elevation (NSTE) acute coronary syndromes. Three lead continuous ECG (cECG) monitoring (LifeCard CF/Pathfinder, DelMar Reynolds/Spacelabs, Issaqua Wash.) was performed for a median duration of 4 days at a sampling rate of 128 Hz. The endpoints of death and myocardial infarction were adjudicated by a blinded Clinical Events Committee for a median follow-up period of 60 days. The maximum follow-up was 90 days. Data from 686 patients was available after removal of noise-corrupted signals.
To evaluate the ability of electrocardiographic mismatch to identify patients at increased risk of future cardiovascular events, we first separated out the patients into a dominant normal sub-population and a group of abnormal patients. This was done by terminating hierarchical clustering one iteration before it placed more than 80% of the patients in the same cluster. This dichotomized the patients into a low risk group containing less than 80% of the population, and a high risk group containing the rest.
Kaplan-Meier survival analysis (Machin D, Cheung Y, Parmar M. Survival Analysis: A Practical Approach. Wiley) was used to study the event rates for death and myocardial infarction (MI). The outcomes were studied both separately, as well as after being combined to create a composite endpoint of death or MI (death/MI).
Patient physiological signals are separated into discrete components, and each variation of that component is assigned a unique representation (e.g., a number or symbol). For example, ECGs from one or more patient are separated into a plurality of discrete waveforms which correspond to individual heart beats. Waveforms corresponding to normal heartbeats are each assigned a unique representation (e.g., the letter N). Waveforms corresponding to abnormal contractions originating from ventricular regions are each assigned a different representation (e.g., the letter V). Further classes of abnormal heart beats are each assigned their own unique symbols. To accommodate for minor variations in individual waveforms, all N waveforms are grouped together and a characteristic N (i.e., normal) waveform is extrapolated therefrom. Characteristic waveforms are also extrapolated for each type of abnormal heartbeat. The characteristic waveforms are then used to evaluate heartbeats in the ECGs of new patients.
The characteristic waveform can be a prototype (archetype) waveform or a centrotype waveform. The difference between the prototype and the centrotype is as follows—the prototype is a waveform we construct that is the ‘average’ waveform; the centrotype is the waveform of the average element. For example, if we want the average of the numbers 1, 4, 10, the prototype approach would be to use 1+4+10 divided by 3 (i.e., we compute the average). The centrotype approach would be to say that 4 is the middle element. In some embodiments, the representation is a waveform and the probability of the information class.

Section 5. Results

The results of univariate analysis for death, MI and the combined outcome are shown in Table 1. The corresponding Kaplan-Meier curves are presented in FIG. 1.

TABLE 1

Results of univariate analysis for the outcomes of death, MI and
death/MI.

Endpoint	Hazard Ratio		95% CI	P Value

Death	5.28	1.64-17.02	0.003
MI	1.81	0.85-3.87	0.150
Death/MI	2.84	1.45-5.56	0.003

As seen in Table 1, patients who were electrocardiographically mismatched with the dominant group of the population, were at increased risk of adverse cardiovascular events. Patients placed automatically in the high risk group had a much higher rate of death during follow-up than patients in the low risk group (4.42% vs. 0.87%; p=0.003). A similar trend was seen for MI (5.75% vs. 3.26%) although in this case the relationship was not statistically significant (p=0.149). For the combined death/MI endpoint, i.e., the occurrence of either of these adverse outcomes, the cumulative incidence in the high risk group was 9.29% as opposed to 3.48% in the low risk group (p=0.003).
In all, hierarchical clustering produced 31 clusters, i.e., one dominant cluster that constituted the low risk group of patients and 30 clusters that collectively formed the high risk group. Of the high risk clusters, 15 had only a single element. No death or MI events were observed in these groups, and it is likely that these isolated singletons corresponded to noisy electrocardiograms that were not removed during the noise rejection stage described in Section 4. Conversely, 5 of the high risk clusters had 10 or more elements. Table 2 presents the risk of events for these individual high risk clusters. The data suggests that patients in different clusters have distinct risk profiles. For example, in cluster B patients, the risk of death is 17.39% relative to a risk of 0.87% in the low risk population. The overall risk of death/MI is also correspondingly elevated (21.74% vs. 3.48%). Similarly, in cluster D, there are no deaths but the risk of MI is 14.29% as opposed to 3.26% in the low risk population.

TABLE 2

% of patients with events in five largest clusters in high risk group.

	Cluster	# of Patients	% Death	% MI	% Death/MI

A	101	2.97	4.95	7.92
B	23	17.39	4.35	21.74
C	21	9.52	4.76	9.52
D	14	0.00	14.29	14.29
E	10	10.00	10.00	10.00

The percentage of events in the combined population comprised by the high risk clusters with 10 or more members (i.e., clusters A to E) is shown in Table 3. This data suggests that improved noise removal techniques, or disregarding small electrocardiographically mismatched clusters, could allow for a further focus on high risk cases.

TABLE 3

% of patients with events in aggregate of five largest clusters
in high risk group (n = 169) compared to low risk
group (n = 460).

	# of Patients	% Death	% MI	% Death/MI

169	6.25	8.13	13.13
460	0.87	3.26	3.48

The results presented so far use a cutoff of 80% to separate out the low risk and high risk groups. The effect of varying this cutoff for hierarchical clustering is shown in FIG. 2. Specifically, for each decile, hierarchical clustering was terminated before it placed a corresponding percentage of patients in a single dominant cluster. The event rate between the high risk and low risk groups was then calculated. As shown in FIG. 2, increasing the clustering threshold (i.e., choosing a high risk population that is more electrocardiographically mismatched from the dominant group) generally leads to an increased percentage of events in the risk group. This effect tapers off at the final decile, most likely due to the effect described earlier, where some of the exceedingly dissimilar electrocardiograms correspond to noise.
This invention relates to an alternative approach of identifying abnormal patients by searching for population outliers. A comparative framework is adopted that is able to discover patient groups that are at an increased risk of adverse cardiovascular outcomes. In contrast to deriving a feature from the ECG for each patient, signal morphologies for every pair of patients to determine how different they are, i.e., how much electrocardiographic mismatch exists are directly compared. Patients with abnormal cardiac characteristics then correspond to those individuals whose long-term ECG did not match the dominant group in the population. A more fine-grained risk profile for patients based on the specific cluster they fall within is also developed.
The process of separating out patients into different groups essentially clusters individuals with similar ECG morphology and dynamics together. The concept of clustering ECG signals based on morphology has been proposed earlier, e.g., (Lagerholm M, Peterson C, Braccini G, Edenbrandt L, Sornmo L. Clustering ECG complexes using Hermite functions and self-organizing maps. IEEE Trans Biomed Eng. 2000;47:838-48) and (Cuesta-Frau D, Perez-Cortes J, Andreu-Garcia G. Clustering of electrocardiographic signals in computer-aided Holter analysis. Computer Methods and Programs in Biomedicine. 2003;72: 179-96). The focus of these techniques has typically been to cluster individual ECG beats together based on their morphology. The current methodology describes a method that is able to cluster together patients based on the morphology of the entire electrocardiogram, i.e., inter-patient comparisons based on ECG morphology as opposed to inter-beat comparisons.
The present invention develops an approach to obtain a quantifiable comparison of how different two patients are electrocardiographically. This information may be used to partition patients into similar groups, with matching long-term electrocardiograms. The hypothesis underlying the work is that patients with ECG signals that match in morphology and dynamics will have consistent risk profiles. This allows one to obtain a more fine-grained understanding of how a patient's health will evolve over time, and more accurately assign a risk score to the patient for events such as death and myocardial infarction.
An experimental study shows that patients who are electrocardiographically mismatched from the majority patient population are at an increased risk of cardiovascular events such as death and MI over a 90 day follow-up period.
It is important to point out that the technique makes almost no a priori assumptions as to how these high risk patients are different from the rest of the population. In other words, no set of specific morphology classes occurring exclusively in high risk patients are assumed, nor are patients who have more or less variability in the distribution of symbols sought to be identified. One of the strengths of the method is that it is able to find a wide variety of abnormalities that would be difficult to describe along an ordinal scale. One limitation of this approach, however, is that patients whose electrocardiograms have been corrupted by noise will also appear as outliers. This does not lead to any important cases being missed but adds false positives to the high risk group.
Another significant aspect of the work is that symbolization is used as an intermediate step to calculate EM. The ECG from patients is first symbolized, and the distribution of symbols is compared as described in Section 2. The use of symbolization can be considered an optimization step. For example, EM can be calculated directly from the raw data by treating each beat as a distinct symbol. However, the use of symbolization greatly reduces the number of comparisons between the beats of two patients, allowing us to simply compare the representative elements of each symbol cluster and weight the differences by the probabilities of these symbols. For example, comparing the raw electrocardiograms between two patients may involve comparing 100,000 beats from the first patient with 100,000 beats from the second (i.e., 100,000²comparisons). However, using symbolization to reduce the data to 50 symbols for each patient would result in 50²comparisons being needed to estimate EM.
One of key goals of EM with hierarchical clustering is to partition patients into smaller groups with consistent risk profiles. The results in Section 5 show that different clusters may be at varying risk for subsequent cardiovascular events. This allows for a more fine-grained assessment of individual patients. Specifically, the idea of using a nearest neighbor system to assign patients to one of the previous clusters determined by EM with hierarchical clustering is currently being investigated. The approach could allow one to more precisely state what the expected risk of an individual patient is, as opposed to merely placing them within a group with elevated risk.
Importantly, the formation of groups with smaller risk profiles permits new patients to be assigned a risk score based on which patient group this new patient best matches to. Thus, the invention provides methods of assigning a risk score to new patients by matching them to an existing database of patients, comprising the following steps: grouping patients in the database as described herein; segmenting the physiological signal into a plurality of components for each new patient; grouping the components into a plurality of information classes for each new patient; assigning a representation to each information class for each new patient; matching the representations of new patients with representations of groups of patients from the database assigning new patients the risk characteristics of patients in the groups of patients in the database with matching representations
In this application the invention is discussed generally in terms of a method for clustering patients according to various physiological states. The invention can be implemented as a physiological (e.g., electrophysiological) monitor (e.g., ECG) in communication with, for example, a general purpose computer. The physiological signal data is received and stored in a data storage device for subsequent analysis by the program modules of the computer. Individual program modules include but not limited to: dividing the signal data into a plurality of time portions; assigning a representation to each time portion; and grouping the patients in response to their representations or symbols. It is contemplated that in another embodiment such program modules may in fact be incorporated into the ECG monitor itself. The data storage device can be a central database or it can be a local memory device (e.g., computer readable medium) located on, for example, a computer. The data storage device can be in bidirectional communication with the computer such that the computer can retrieve physiological data from the data storage device and the computer can save physiological data (e.g., new patient data) and analytical results to the data storage device. The computer optionally can be in communication with a display for displaying physiological data, time portions, risk profiles, risk scores, representations, symbols, numbers, waveforms, clustering and other features as described herein.
Variations, modification, and other implementations of what is described herein will occur to those of ordinary skill in the art without departing from the spirit and scope of the invention as claimed. Accordingly, the invention is to be defined not by the preceding illustrative description, but instead by the spirit and scope of the following claims.

Claims

1. A method of partitioning a plurality of patients into risk profile groups comprising the steps of:

recording a physiological signal from each patient of a plurality of patients;

segmenting the physiological signal into a plurality of components for each patient of a plurality of patients;

grouping the components into a plurality of information classes for each patient of a plurality of patients;

assigning a representation to each information class for each patient of a plurality of patients; and

grouping the patients in response to the representations of their respective information classes.

2. The method of claim 1 wherein the representation is a numerical value.

3. The method of claim 1 wherein the representation is a symbol.

4. The method of claim 1 wherein the representation is a waveform.

5. The method of claim 4 wherein the waveform is a prototype (archetype) waveform.

6. The method of claim 4 wherein the waveform is a centrotype waveform

7. The method of claim 1 wherein the physiological signal is an ECG and the equivalent time portion is a heartbeat.

8. The method of claim 7 wherein the step of grouping comprises:

measuring an electrocardiographic mismatch between each pair of patients in the plurality of patients;

assigning each patient of the plurality of patients to a respective cluster of a plurality of clusters;

grouping clusters in response to the electrocardiographic mismatch until the number of clusters reaches a predefined minimum; and

assigning risk outcomes to each of the clusters.

9. The method of claim 8 wherein the grouping of clusters is performed hierarchically.

10. A method of partitioning a plurality of patients into risk profile groups comprising the steps of:

recording an ECG signal from each patient of a plurality of patients;

dividing each ECG signal into a plurality of heartbeats for each patient of the plurality of patients;

grouping heartbeats into clusters for each patient of the plurality of patients;

assigning a representation to each cluster for each patient of the plurality of patients;

assigning a value to differences in the representation of clusters for each respective ecg signal for each respective patient of the plurality of patients; and

grouping the patients in response to the values of their respective differences in ECG signals.

11. A method of partitioning a plurality of patients into risk profile groups comprising the steps of:

recording an physiological signal from each patient of a plurality of patients;

dividing each physiological signal into a plurality of equivalent time portions for each patient of the plurality of patients;

grouping time portions into clusters for each patient of the plurality of patients;

assigning a value to differences in the representation of clusters for each respective physiological signals signal for each respective patient of the plurality of patients; and

grouping the patients in response to the values of their respective differences in physiological signals.

12. A method of assigning a risk score to new patients by matching them to an existing database of patients, comprising the following steps:

grouping patients in the database using the method of claim 1;

segmenting the physiological signal into a plurality of components for each new patient;

grouping the components into a plurality of information classes for each new patient;

assigning a representation to each information class for each new patient;

matching the representations of new patients with representations of groups of patients from the database; and

assigning new patients the risk characteristics of patients in the groups of patients in the database with matching representations.