Keywords

1 Introduction

During the last few years has occurred a rapid technological evolution in the scientific field of the Internet of Medical Things (IoMT) and Wireless Body Area Network (WBAN). The main demands for these systems can be summarised as follows: (i) reducing the healthcare costs while keeping the quality of the services and (ii) promoting wellness programs to shift the health expenditure from treatment to prevention [3].

All the efforts in this field by the scientific research communities has made it possible to obtain electronic devices of minimal size and wearable [31]. This has created a fertile ground for telemedicine. Telemedicine can be commonly defined as the use of advanced telecommunications technologies for the purpose of supporting many medical activities. In the last years, this industry has grown and most US health institutions and hospitals are currently employing such kind of technology [12].

In this context, medical activities become responsible—beyond the knowledge and clinical skills—in handling an ample amount of data related to the patient health. Thus, appropriate elaboration of the clinical data are required to facilitate the work of experts and promote a policy of welfare. A Decision Support System (DSS) is the key component of an effective telemedicine system. Such a component is basically a layer of software that latently and continuously analyze the acquired data aimed at providing recommendations, even at patient-level, to the medical experts for the identification of a risky situation or for the diagnosis of a specific pathology [38].

In this paper we present an approach that could be integrated in a DSS of a telemedicine system aiming at supporting the identification of atrial fibrillation (AF) episodes through the analysis of ECG. AF is the most common sustained arrhythmia and is associated with significant morbidity and mortality [10]. We decided to focus on AF detection because of the incidence statistics of such a pathology. Indeed, around one third of all ischemic strokes are caused by AF [14] and the early phase of appearance is a particularly high-risk period for the development of stroke [34]. In addition, AF is often asymptomatic. Thus, it is crucial to detect onset episodes of AF with high accuracy to allow a proper intervention of cardiologist [18].

The detection of AF episodes generally involves two electrocardiogram (ECG) sources of information: (i) beat morphology, because during an AF episode, it is possible to observe fluctuating wave forms instead of P waves and (ii) rhythm, because during an AF episode it is possible to observe an irregularity of heart rate. The fact that an ECG recording of the episode is a diagnostic criterion can make the process cumbersome, especially if the arrhythmia is paroxysmal and not easily provoked during a recording session. In order to capture the episode, an extended recording time (at least 24 h) through an Holter monitoring is required [2]. These recordings from wearable ECG devices introduce an amount of data which results complicated for the physician to inspect and analyze. This recall the need of semi-automatic approaches to determine onset and duration of AF episodes.

A lot of effort in the research community has been devoted to the definition of methods to automatically detect AF. These are often based on Machine Learning techniques and—most of them—are based only on the analysis of R-R intervals (RRI), i.e., they just exploit the rhythmic source of information. Even if the accuracy of such approach is generally very high in terms of accuracy (more than 95%), the proposed approach still misclassifies fibrillant heart beat signals as non-fibrillant [41]. This suggests that there is still room for improvement. Especially, our conjecture is that by combining morphological and rhythm features is possible to improve the accuracy of approaches based on just one of the two source of information.

Based on the above consideration, in a previous work we have presented Morphythm [23], a new approach based on machine learning techniques where morphological and rhythmic information are fused together. Morphythm showed surprising results, especially for what concerns two vital aspects of the medical classification: increment of true positives and reduction of false negatives.

In this paper we present an extension of Morphythm aiming at further improving its accuracy. We first performed a rigorous feature engineering process in order to identify the features that contribute the most to the prediction of AF events. Then, we experimented most advanced machine learning techniques, including artificial neural network and deep learning techniques. Finally, we integrated in Morphythm the concept of"local" prediction, successfully used in other context [28]. Especially, instead of producing a single prediction model, the new version of Morphythm, called Local Morphythm, automatically build several prediction models based on the characteristics of the ECGs in the training set. In particular, the training set is clustered in order to put together ECGs that exhibits similar characteristics. Then, for each cluster, Local Morphythm builds a prediction model. When a new data point is provided, Local Morphythm first selects the most suitable model based on the characteristics of the new data point, and then it performs the prediction applying the selected model.

The rest of the paper is structured as follows: Sect. 2 provides details on AF and on automatic detectors of AF. Section 3 presents Local Morphythm, our novel approach for AF detection, while Sect. 4 reports the design and the results of the empirical study we conducted to evaluate Local Morphythm. Finally, Sect. 5 concludes the paper and provides suggestions for possible future research directions.

2 Background and Related Work

2.1 Atrial Fibrillation

Normally, the heart contracts and relaxes to a regular beat. In atrial fibrillation, the upper chambers of the heart (the atria) beat irregularly (quiver) instead of beating effectively to move blood into the ventriclesFootnote 1. If the ECG recording is available, AF is diagnosed by whenever an irregular heartbeat presents the following characteristics: the absence of P waves (with disorganized electrical activity in their place) and irregular R–R intervals due to irregular conduction of impulses to the ventricles [15].

The prevalence of atrial fibrillation (AF) is increasing all over the world and it is becoming one of the most important clinical issues for industrialised countries [15, 42]. AF is a crucial risk factor for the occurrence of stroke. Beyond stroke, AF can lead also to congestive heart failure. Furthermore, hypertension, diabetes and heart failure are some of the most common comorbidities [24, 39]. In addition, AF presents a sever influence on the global health conditions of individuals who contract it [20].

To produce a diagnosis of AF, a cardiologist checks the clinical history of the patient and the ECG signal, by at least observing a single lead during the revealing of the episode [15]. Unfortunately, AF is often paroxysmal, i.e., there are recurrent episodes that stop on their own in less than seven days [15], and asymptomatic. For these reasons, the screening of such a pathology needs to become a priority.

2.2 Automatic Detection of Atrial Fibrillation

In recent years, the scientific research has provided several works aiming at automatically detect AF episodes. Most of them have shown important results by exploiting only the analysis of heart rhythm, assumed as the observation of the distances between two successive R peaks (RRI, RR intervals) [8, 30, 37, 40]. Indeed, the detection methods based on RRI produce relatively more precise identification of AF since the R-wave peak of QRS complex is the most prominent characteristic feature of an ECG recording and the least susceptible to various kinds of noise [19, 21, 25, 26].

In the work by Hochstadt et al. [18], around 18 thousand consecutive RR interval measurements were recorded in 20 patients, including about 12 thousand RR intervals during AF and 6,087 RR intervals during sinus rhythm. The automatic algorithm—based on Lorenz-plot—used by the authors distinguished AF from sinus rhythm with a sensitivity of 100\(\%\) and specificity of 93.1\(\%\).

In the study by Andersen et al. [2], a novel approach for AF detection based on Inter Beat Intervals (IBI) extracted from long term electrocardiogram (ECG) recordings is presented. For this purpose, five time-domain features have been extracted from the IBIs and a Support Vector Machine (SVM) has been used for classification. The proposed approach has shown a significantly reduced computation time without loss of performance, if compared to a consolidated baseline.

Afdala et al. [1] test the ability of simply involving the Shannon entropy in the detection of Atrial Fibrillation episodes. In their research study, they used data from a well-known public data set (Physionet MIT-BIH AFDB) and, as performance, they observed that Shannon entropy has the highest accuracy if a threshold of 0.5 is set.

In the work by Chen et al. [5], a new feature extraction method based on RR interval is proposed with the aim at describing an heart rhythm which will be submitted to a classification experiment. As descriptors, they used the robust coefficient of variation (RCV), the distribution shape of RR interval is described with the skewness parameter (SKP), and the complexity of RR interval is described with the Lempel-Ziv complexity (LZC). Finally, the feature vectors have been used as input into the support vector machine (SVM) classifier model to achieve automatic classification and detection of atrial fibrillation. Also in this case, the MIT-BIH atrial fibrillation database was used to verify the data. The final classification results showed a sensitivity of 95.81\(\%\), a specificity of \(\%\) and an accuracy equal to 96.09\(\%\).

In the next subsection, the method chosen as baseline—and embedded in Morphythm and consequently in the new approaches proposed in this paper—is described by providing the main ideas and highlighting the computational steps.

The Method Proposed by Zhou et al. [41]. This section provides details on the method proposed by Zhou et al. [41], i.e., our baseline in the evaluation of Local Morphythm. Such an approach consists in the following steps:

  • the HR sequence is converted to a symbolic sequence in a fixed interval;

  • a probability distribution is constructed from the word sequence which is transformed from the symbolic sequence;

  • a coarser version of Shannon entropy is employed to quantify the information size of HR sequence using the probability distribution of word sequence;

  • discrimination of the heart beat type (AF or no-AF) using a threshold.

Step 1: Converting the HR Sequence. The first step of the method regards the generation of a symbolic dynamic starting from the analysis of a sequence of heart beat (\(hr_{n}\)). Especially, the authors encode the information included in \(hr_{n}\) to a sequence of fewer symbols, where each symbol aims at representing an instantaneous state of heart beating. The mapping function is the following:

$$ sy_{n} = {\left\{ \begin{array}{ll} 63, &{} \text {if }n\text { hr} \ge \text {315} \\ \lfloor hr_{n} \rfloor , &{} \text {other cases} \end{array}\right. } $$

where [\(\lfloor \cdot \rfloor \)] represents a floor operator.

Step 2: Building the Symbolic Sequence. The authors apply a 3-symbols template in order to explore the entropic properties of the symbolic series sy\(_{n}\). Thus, to examine the chaotic behavior, the word value can then be calculated as:

$$ wv_{n} = (sy_{n-2} \times 2^12) + (sy_{n-1} \times 2^6) + sy_{n} $$

Step 3: Computing the Entropy. The authors define a coarser version of Shannon entropy \(H^{''}\)(A) to quantitatively calculate the information size of wv\(_{n}\). In this study, the dynamic A comprises of 127 consecutive word elements from wv\(_{n-126}\) to wv\(_{n}\), as proposed in the function below:

$$\begin{aligned} H^{''}(A) = - \frac{k}{Nlog_{2}N} \sum _{i=1}^{k} p_{i}log_{2}p_{i} \end{aligned}$$

where N and k are total number of the elements and characteristic elements in space A, respectively.

Step 4: Classification. Based on the obtained entropy value, a final beat-to-beat classification (fibrillant or not-fibrillant) is presented by applying a threshold discrimination. The optimal threshold was empirically identified at 0.639.

3 The Proposed Approach: An Overview

This section describes Local Morphythm, an evolution of the approach recently proposed by Laudato et al. [23], called Morphythm. Local Morphythm is able—given a heart beat signal—to classify it as fibrillating or not fibrillating.

As well as Morphythm, Local Morphythm uses supervised machine learning techniquesFootnote 2 to combine rhythmic and morphological features extracted from an ECG and predict whether or not a heart beat is fibrillating or not fibrillating. However, in Local Morphythm, (i) a rigorous feature engineering process and (ii) a local prediction strategy have been adopted in order to identify respectively the features that best contribute to the prediction of AF episodes and to evaluate if a local approach may be preferred instead of a global one.

3.1 Pre-processing

Before extracting features, the ECG data have to be pre-processed according to [33] and [6]. The main steps involved in this phase are: (i) the detrend of the ECG signal, (ii) the application of a filtering stage (where a low and high pass filters have been applied to get rid of baseline wander and discard high frequency noise, respectively) and (iii) the normalization of the samples.

Once executed the previous steps, the Pan-Tompkins [33] QRS-detection method has been applied with the aim at segmenting the ECG in heart beat signals. In this work, as heart beat, it is intended the signal included between two successive R peaks. Such an interpretation is very suitable for AF detection, because it highlights the atrial activity.

3.2 Feature Extraction and Selection

As the name suggests, Local Morphythm embeds features extracted from both the heart rhythm and the morphology of a heart beat. Rhythmic features are based on one or more heart beats and they aim at capturing aspects that mostly regard the regularity of the heart beat signal. Zhou et al. [41] state that the detection methods based on RRI are more useful to produce a precise and accurate identification of AF because the R-wave peak of the QRS complex is the most prominent characteristic feature of an ECG recording. Such a characteristic is less subject to noise [26].

Even if the acquisition of rhythmic features can be very reliable, such features can only help detecting arrhythmia, which is just one of the possible signs of AF. Thus, morphological features are necessary to detect anomalies in the shape of a single heart beat signal and could be particularly useful to corroborate the warnings raised by analysing the rhythm.

Thus, similarly to Morphythm, also in Local Morphythm we consider both rhythm and morphological features. Especially, we consider the same set of features used in Morphythm [23]:

  • Rhythmic Features: we used two features based on the observation of a single heart beat signal, i.e., Heart Beat Length (HBL) and Heart Beat Discrete Length (HBDL), and two additional rhythmic features that consider the information of a sequence of consecutive heart beats, i.e., Heart Beat Regularity (HBR) and Entropy, as defined by [41]. HBL represents how long a single heart beat signal lasts. HBDL is a classification of the heart beat signal in three classes, based on its length: a beat is (i) short if it takes less than 0.5 s, (ii) long if it takes more than 1.2 s, and (iii) regular otherwise. HBR is based on HBDL. It considers a rhythmic pattern of 10 consecutive discrete heart beats lengths. Once obtained the pattern, we compute HBR simply counting the number of regular heart beats.

  • Morphological Features: given a sequence of samples provided for a heart beat signal, we computed several features: (i) the Mean Signal Intensity (MSI), (ii) the Signal Intensity Variance (SIV), (iii) the Signal Intensity Entropy (SIE). MSI, SIV and SIE are features obtained by measuring respectively the mean, the variance and the entropy [29] of all the samples acquired in a heart beat signal. To try to enrich the knowledge of classifiers, we also used the segmented version of these last features: we divided proportionally the heart beat signal in 10 segment and for each portion we evaluated the MSI, SIV and SIE. Finally, we included the features obtained by (i) the application of the Fast Fourier Transform on 32 points and (ii) the estimation of the coefficients of the Auto-Regressive model of order 16.

Table 1. Features ranking using Information Gain.
Full size table

For each heart beat signal we extract a total number of 76 different features (eight rhythm features and 68 morphological features). In order to select the most appropriate features for the detection of AF events we used the Weka InfoGainAttributeEval as Attribute Evaluator and Ranker as Search Method. The former basically evaluates the worth of an attribute by measuring the information gain with respect to the class, while the latter ranks attributes by their individuals evaluations.

The feature selection process has been conducted on the MIT-BIH AF Database [16], a commonly used benchmark which contains recordings of 25 patients. Each recording in the data set lasts 10 h and contains two ECG signals sampled at 250 samples per second (12-bit resolution). Due to the embedding of morphology descriptors, the feature selection process has been performed on the AFDB\(_2\), i.e., the AFDB without records 00735 and 03665 because, for such records, only information on the rhythm is available [16] and without 04936 and 05091 because—as others have shown [25]—the records 04936 and 05091 include many incorrect manual AF annotations.

The outcome of the features selection process is reported in Table 1. From the analysis of the results achieved, we observe that:

  • rhythmic information in AF episodes detection represent the main contribution in terms of information gain;

  • morphological features of an ECG can provide a contribution in terms of information gain for the automatic classification of heart beats. Specifically, these features refer to the middle and the last part of the signal, where the fibrillating rhythm appears and where the P-wave can exhibit its changes.

By selecting a fixed threshold of 0.12, we obtain a selection of a group of six features containing a balanced number of morphological and rhythmic features. Thus, we decided to incorporate in Local Morphythm the first six features reported in Table 1.

3.3 Making the Prediction

The main difference between Local Morphythm and Morphythm regards the way as the prediction is performed. In Morphythm, as in any canonical approach based on supervised machine learning techniques, a training set is used to build a (global) prediction model. Such a model is used on all the new data points where a prediction is required. Especially, when a new heart beat signal is provided, Morphythm first computes the features on this new heart beat signal and then uses the prediction model to determine whether or not the heart beat is fibrillating or not fibrillating.

Fig. 1.
figure 1

Workflow of Local Morphythm.

Full size image

However, the heart beat signals in the training set could be quite different each other. The heterogeneity of the training set might negatively impact the accuracy of the prediction model [28]. In order to mitigate such a problem, in Local Morphythm we integrated a local prediction strategy [28].

Local Morphythm first clusters the training set into homogeneous sets of heart beat signals. Then, it builds for each cluster a specific prediction model using a supervised machine learning technique. In this way, Local Morphythm does not have just one global prediction model, but it has a set of prediction models that are particularly suitable for specific heart beat signals.

When a new heart beat signal is provided, Local Morphythm first computes the features on this new heart beat signal and then it identifies the cluster of heart beat signals more similar to the new heart beat signal. Once identified such a cluster, Local Morphythm uses the model associated to the identified cluster of heart beat signals to noindent predict whether or not the new heart beat is fibrillating or not fibrillating. The workflow of Local Morphythm is depicted in Fig. 1.

Fig. 2.
figure 2

Results of the Calinski-Harabasz score in order to determine the best value of k for the k-means clustering algorithm. The higher the value of the score the higher the overall quality of the clustering.

Full size image

In order to cluster the training set, we have exploited the k-means clustering algorithm [27]. This method follows a simple way to classify a given data set through a certain number of clusters fixed a priori. The main idea is to define k centroids, one for each cluster. The main steps are described below:

  • Place K points into the space represented by the objects that are being clustered. These points represent initial group centroids.

  • Assign each object to the group that has the closest centroid.

  • When all objects have been assigned, recalculate the positions of the K centroids.

  • Repeat Steps 2 and 3 until the centroids no longer move. This produces a separation of the objects into groups from which the metric to be minimized can be calculated.

We have determined the optimal value of k using the Variance Ratio Criterion (also known as Calinski-Harabasz score) [4]. Especially, we have performed the clustering of the heart beats by using different values of k from 1 to 25. For each cluster we have computed the Calinski-Harabasz score in order to determine the value of k that determines the clustering with the highest score. The plot in Figure shows that the highest Calinski-Harabasz value occurs with \(k=16\). This number has also been confirmed by involving the Silhouette method [36], an alternative method for the identification of the best k value (Fig. 2).

4 Empirical Evaluation

This section reports the empirical evaluation we conducted to evaluate the accuracy of Local Morphythm.

4.1 Design of the Study

The goal of this study is to evaluate the accuracy of Local Morphythm is classifying AF events in a patient. The perspective is both (i) of a researcher who wants to understand if a local prediction strategy to combine rhythmic and morphological features is worthwhile for detecting AF events, and (ii) of a practitioner who wants to use the most accurate and precise approach in a telemedicine application for the detection of AF events. Thus, the study is steered by the following research question:

To what extent, a local prediction model—based on the combination of rhythmic and morphological information—improves the automatic detection of AF episodes?

The context of this study is represented by the MIT-BIH AF Database [16], and specifically the AFDB\(_2\), i.e., the AFDB without records 00735 and 03665 because, for such records, only information on the rhythm is available [16]. Also, records 04936 and 05091 were excluded due to many incorrect manual AF annotations [25].

In the context of our study, we also experimented a large set of machine learning techniques. Indeed, for the classification performances, we have involved in our experiments—beyond the Random Forest [17], J48 [35], Logistic [9], AdaBoost M1 [13] and RepTree [11] already used by Laudato et al. to evaluate Morphythm [23]—Neural Networks [22], Multi Layer Perceptron [32], JRip [7] and SGD (which implements stochastic gradient descent for learning various linear models)Footnote 3 .

As validation technique, we have chosen the Leave One Person Out Cross Validation (L1PO-CV). L1PO-CV means that one person at a time is left out from the training set, so that the training set contains no data specific to the individual who is being tested (the classifier was not tuned with the test data of that person). This is possible since each data segment is associated with an anonymous label corresponding to an individual.

To answer our research questions we compared:

  • True Positives (TP), i.e.,, the number of instances classified as fibrillating by the approach and that were actually fibrillating;

  • True Negatives (TN), i.e.,, the number of instances classified as not fibrillating by the approach and that were actually not fibrillating;

  • False Positives (FP), i.e.,, the number of instances classified as fibrillating by the approach and that were actually not fibrillating;

  • False Negatives (FN), i.e.,, the number of instances classified as not fibrillating by the approach and that were actually fibrillating.

In the context of telemedicine a high number of TP is desirable, because it indicates the number of AF episodes correctly detected. Also, it is desirable to have an approach that does not lose any AF episode: thus, keeping the number of FN low is very important.

4.2 Analysis of the Results

Table 2 compares the prediction accuracy, in terms of TP, TN, FP, and FN, achieved by Local Morphythm, Morphythm, and the approach proposed by Zhou et al. [41], the most accurate approach in the literature for the detection of AF events.

Table 2. Comparison of Local Morphythm with Morphythm (with the same features selection strategy used in Local Morphythm) and the approach proposed by Zhou et al. [41]. In boldface the best results achieved by these methods.
Full size table

From the analysis of the results emerges that for both the approaches Morphythm and Local Morphythm the best overall accuracy is achieved when SGD is used as machine learning techniques.

Using such a technique, Local Morphythm is able to achieve the best results in terms of both TP and FN. Specifically, Local Morphythm is able to identify 8,340 TP more than the baseline (approach by Zhou et al.) and 1,114 TP more than Morphythm. Also, Local Morphythm is able to retrieve less FN with respect to both the baseline and Morphythm, i.e., −5,533 and −569, respectively.

However, the approach proposed by Zhou et al. [41] is still the best in terms of TN and FP. Specifically, Local Morphythm and Morphythm generate 6,052 and 6,064 FP more than the approach by Zhou, respectively. In terms of TN, instead Local Morphythm and Morphythm retrieves less TN as compared to the baseline, i.e., −8,859 and −8,326, respectively.

By looking at the results achieved at patient level, i.e., by considering a single recording, we observe that Local Morphythm sensibly outperforms—in terms of every metrics—both the baseline and Morphythm for 5 out of 21 recordings (around 24\(\%\)). Examples of such an improvement is reported in Table 3, where it is possible to observe the classification performances of Local Morphythm with respect to the baseline and Morphythm.

In addition, if we focus the attention on just TP and FN, Local Morphythm outperforms both the other approaches baselines in 8 out of 21 recordings (around 38\(\%\) of the data set).

For the remaining recordings, the value of all the evaluation metrics are almost balanced, in the sense that no significant improvement can be observed.

Table 3. Example of records on which Local Morphythm outperforms both Morphythm and the approach by Zhou et al. [41] in terms of all the considered evaluation metrics.
Full size table

The only recording with abnormal classification performances is the recording 08378 where Local Morphythm presents a significant loss in terms of TP and FN with respect to the other two approaches. This suggests that on this particular recording the local prediction strategy is not worthwhile because very likely such a recording exhibits characteristics that are quite different from the other recordings in the data set.

In order to validate such a conjecture we compare the average distance between each recording and all the others but 08378 and the distance between each recording and recording 08378. In order to compute the distance between two recordings we considered them as mono dimensional vectors (by selecting the first ECG channel available for each recording) and then compute the Euclidean distance between the two vectors.

The analysis is depicted in Fig. 3. As we can see, the distance between the recording 08378 and a generic recording i is much higher that the average distance between the recording i and all the other recordings but 08378. Such a result confirms our conjecture that the recording 08378 is quite different from the others recording; thus, in this specific case, the local prediction strategy does not provide any benefits as compared to the other two approaches.

Fig. 3.
figure 3

Average distance between a generic recording i and all the other recordings but 08378 compared to the distance between recording i and 08378.

Full size image

Once this recording is excluded from the data set, the classification accuracy of Local Morphythm improves even more. Indeed, Local Morphythm—especially when using the Logistic and the SGD algorithms—avoids a loss of around 1,5 thousands heart beats classified as TP and FN.

5 Conclusion and Future Work

In this paper we presented an extended version of the approach proposed by Laudato et al. [23], named Morphythm, where rhythmic and morphological features are combined together in order to improve the classification accuracy of AF episodes. The new approach, called Local Morphythm integrates a more rigorous feature engineering process as compared to Morphythm and more advanced machine learning techniques, including artificial neural networks. We also extended Morphythm by integrating in the approach the strategy of “local" prediction, successfully used in other contexts [28]. Especially, instead of producing a single prediction model, Local Morphythm automatically builds several prediction models based on the characteristics of the ECGs in the training set. In particular, the training set is clustered in order to put together ECGs that exhibits similar characteristics. Then, for each cluster, Local Morphythm builds a prediction model. When a new data point is provided, Local Morphythm first selects the most suitable model based on the characteristics of the new data point, and then it performs the prediction applying the selected model.

An experimentation conducted on the MIT-BIH AF Database [16] indicates that Local Morphythm is able to increase the TP and reduce the FN as compared to Morphythm and the approach by Zhou et al. [41], one of the best approaches in the literature for the detection of AF episodes. Future work will be devoted on the one hand on the replication of the experimentation on other data sets in order to corroborate the results achieved on the MIT-BIH AF Database and on the other hand on the application of a local prediction technique in the context of automatic detection of other types of arrhythmia.