Classification of orthostatic intolerance through data analytics

Abstract

Objective:

Imbalance in the autonomic nervous system can lead to orthostatic intolerance manifested by dizziness, lightheadedness, and a sudden loss of consciousness (syncope); these are common conditions, but they are challenging to diagnose correctly. Uncertainties about the triggering mechanisms and the underlying pathophysiology have led to variations in their classification. This study uses machine learning to categorize patients with orthostatic intolerance.

Methods:

We use random forest classification trees to identify a small number of markers in blood pressure, and heart rate time-series data measured during head-up tilt to (a) distinguish patients with a single pathology and (b) examine data from patients with a mixed pathophysiology. Next, we use Kmeans to cluster the markers representing the time-series data.

Results:

We apply the proposed method analyzing clinical data from 186 subjects identified as control or suffering from one of four conditions: postural orthostatic tachycardia (POTS), cardioinhibition, vasodepression, and mixed cardioinhibition and vasodepression. Classification results confirm the use of supervised machine learning. We were able to categorize more than 95% of patients with a single condition and were able to subgroup all patients with mixed cardioinhibitory and vasodepressor syncope. Clustering results confirm the disease groups and identify two distinct subgroups within the control and mixed groups.

Conclusion:

The proposed study demonstrates how to use machine learning to discover structure in blood pressure and heart rate time-series data. The methodology is used to classification of patients with orthostatic intolerance.

Significance:

Diagnosing orthostatic intolerance is challenging, and full characterization of the pathophysiological mechanisms remains a topic of ongoing research. This study provides a step toward leveraging machine learning to assist clinicians and researchers in addressing these challenges. Autonomic nervous system imbalance can lead to orthostatic intolerance (OI). Uncertainties about the pathophysiology makes classification difficult. This study categorize OI using machine learning.

1689780_Figa

1. Introduction

Cerebral hypoperfusion is usually caused by a decrease in arterial blood pressure (BP) to a level below the autoregulatory capacity giving rise to symptoms including dizziness, nausea, blurred vision, and eventually syncope (or fainting – loss of consciousness from 10 to 20 seconds with spontaneous recovery [9]). Arterial BP falls when cardiac output and/or systemic vascular resistance are reduced – usually by a sudden pause in the heart rhythm or loss of arteriolar tone. Both can cause syncope, but in many patients, symptoms result from a combination of the two pathophysiological responses [16, 35]. It is estimated that 25 percent of the population experience syncope and even more feel lightheaded or dizzy at some point in their life and these symptoms accounts for over 1 million visits to emergency departments per year in the US alone [31]. A standard diagnostic procedure includes the head-up tilt (HUT) test continuously measuring heart rate (HR) and arterial BP.

Several studies advocate the HUT test as the diagnostic method of choice due to its high diagnostic yield [7, 15], but correct analysis of recorded signals requires expertise in the field, not always available in emergency clinics. The HUT test provides information about the integrated cardiovascular response via measurements of HR and BP. However, analysis of these signals are often limited to simple steady-state quantities including mean BP and HR, pulse pressure, and the HR change between rest and HUT, omitting critical information embedded in the signals’ dynamics interaction.

The most prevalent form of orthostatic intolerance is reflex activation of the autonomic nervous system, sub-classified phenomenologically based on induced changes in BP or HR [4, 26, 27, 40]. For patients with orthostatic intolerance, the HUT test is performed to reproduce their neurally mediated reflex in a controlled setting. The HUT test induces pooling of blood below the diaphragm and a subsequent decrease in venous return. For healthy controls, homeostasis is re-established by triggering the baroreflex, which increases HR, cardiac contractility, and peripheral vascular resistance. Prolonged time in the upright position (> 40 min) without the ability to activate the muscle pump will induce reflex syncope (often referred to as vasovagal syncope). If patients cannot be diagnosed within the first 20 min of HUT, to distinguish controls and patients with orthostatic intolerance additional stimulation is induced by administering nitroglycerine, which increases HR above the regulatory limit.

Patients with chronic orthostatic intolerance and a HR increase of more than 30 bpm (> 40 bpm in individuals aged 12–19 years) within the first 10 minutes of HUT with little or no change in BP, and typically no presyncope conditions with postural orthostatic tachycardia (POTS)[39]. Patients whose BP falls before HR and where HR falls to a ventricular rate of less than 40 bpm for more than 10 s, without asystole of more than 3 s do are classified as a cardioinhibitory (type 2A) [5], and patients experiencing a rapid fall in BP with no or minor HR changes are classified as vasodepressive [6, 27]. If both conditions are present, patients are classified as mixed. For each group, these signal characteristics are often combined with frequent episodes of light-headedness, palpitations, tremulousness, generalized weakness, blurred vision, exercise intolerance, or fatigue in the upright position [10, 39].

Figure 1 shows examples of each subject type: control, POTS, cardioinhibition, vasodepresson, and mixed. The figure includes two patients experiencing a mixed syncope; one patient primarily experiencing vasodepression and one patient primarily experiencing cardioinhibition.

Fig. 1.

Example data from the HUT test for six subjects: a healthy control and patients with POTS, cardioinhibitory syncope, vasodepressive syncope, and mixed cardioinhibition and vasodepression syncope. For the latter patient group, we show two examples, a patient primarily experiencing cardioinhibition and one primarily experiencing vasodepression. The red line lines denote the start and end of the tilt from the supine position, up to 60° and back to the supine position. Heart rate (beats per minute) and BP (mmHg) are displayed as functions of time. The light blue lines denote time for administration of nitroglycerine, a vasodilator.

Several pathways could be responsible for the reflex responses, but little is known about the precise hemodynamic maladjustments causing these reactions, and hence it is challenging to generate targeted treatments. The Identification of precise pathways requires a more detailed classification of patient groups.

This study demonstrates how machine learning can be used to assist clinicians in classifying patients with orthostatic intolerance providing more insight into what features are the most important for identifying different disease subgroups. We focus on four types of abnormal responses, including POTS, cardioinhibition, vasodepression, and mixed vasodepression and cardioinhibition (in the remainder of this study, we refer to the latter group as “mixed”). These groups were chosen since they comprise most patients referred to the syncope center at Bispebjerg and Frederiksberg Hospitals, Denmark, from which data has been made available.

2. Data and Methods

This study analyzes anonymized retrospective data from 186 subjects distributed in five groups: controls and patients with POTS, cardioinhibition, vasodepression, and mixed syncope. All subjects were diagnosed using a HUT test, examining their ability to control BP and HR. Data extracted from health records were collected between 2004 to 2015 at the syncope clinic at Bispebjerg and Frederiksberg Hospitals, Denmark. All subjects were referred to the clinic after experiencing dizziness, lightheadedness, or syncope. Subjects were diagnosed in the clinic and those who did not show any symptoms were categorized as controls. For all subjects, cardiovascular disease was ruled out prior to testing, and no patients received medication at the time of the study. The data-handling committee at Bispebjerg and Frederiksberg Hospitals approved the analysis of the anonymized retrospective data.

2.1. Head-up tilt test

After arriving at the hospital, all subjects were instrumented with BP and electrocardiogram (ECG) sensors. BP was measured using a Finapres device (Finapres Medical Systems B.V.) in the index finger of the non-dominant hand. The hand was placed in a sling at the level of the heart. ECG was recorded using standard precordial leads. Continuous ECG and BP signals were sampled at a rate of 1000 Hz and saved digitally using an A/D-converter communicating with a computer via LabChart 7 (ADInstruments). This program allows the extraction of HR from the ECG measurement. After steady signals were detected, following recommendations by Bartoletti et al. [1] the subjects rested for 10 minutes in the supine position before being tilted head-up to an angle of 60 degrees at a speed of 15 degrees/second measured by way of an electronic marker. The table remained tilted up during the initial passive phase of the test. In the case of a negative passive response, after approximately 20 minutes in the upright position, a provocative drug–nitroglycerine-–was administered to induce a neural reflex (an example is shown in Fig. 1). For POTS patients, their diagnosis was established within the first 10 minutes of head-up tilt, and therefore they were not given nitroglycerine. All patients and controls were returned to the supine position at the same tilt speed after a total of 30 minutes or earlier if they presented signs of syncope or presyncope.

2.2. Data and clinical classification

For each subject, time-series measurements of HR and BP are available over the course of the HUT test. Our analysis is based on data starting two minutes before the tilt-up and lasting until two minutes after the tilt-down. The duration of the test varies for each subject, and thus so do the lengths of the time-series. We denote by $p_i$ the number of samples taken for subject $i$, $i=1,\dots ,186$. The time-series data for the $i^{th}$ subject are denoted by

$$h^i=\left(h_1^i,h_2^i,\dots ,h_{p_i}^i\right),$$

$$b^i=\left(b_1^i,b_2^i,\dots ,b_{p_i}^i\right),$$

where $h$ and $b$ refer to HR and BP, respectively. Each subject was classified by a clinician as either control or symptomatic, experiencing POTS, cardioinhibitory syncope, vasodepressive syncope, or mixed syncope (see Fig. 1 and text below). The distributions are reported in Table 1, and summary statistics for the HR and BP measurements are given in Table 2. This table reports the mean HR and BP plus/minus one standard deviation calculated at rest (before HUT), during the first 50% and last 25% of the HUT. Patients not displaying a response to the tilt in less than 20 minutes, nitroglycerine was administered sublingually at a dose of 0.4 mg; it was given to 94% of the control subjects and 0%, 64%, 78%, 76% of the POTS, cardioinhibitory, vasodepressive, and mixed patients.

Table 1.

Summary of subject distribution.

class	subjects	age range	age mean/median	% female
Control	89	14–92	50/49	67
POTS	13	16–38	24/22	85
Cardioinhibition	28	15–80	33/31	63
Vasodepression	27	67–91	58/63	67
Mixed	29	17–92	50/51	70

Table 2.

Mean values plus/minus one standard deviation of HR and BP at rest and during the beginning (phase I) and end (phase II) of the HUT test.

Mean heart rate (HR)
class	rest	tilt phase I	tilt phase II
Control	70±12	78±13	93±17
Cardioinhibition	67±12	78±14	83±21
Vasodepression	70±10	79±13	87±17
POTS	69±15	98±14	105±19
Mean blood pressure (BP)
class	rest	tilt phase I	tilt phase II
Control	76±16	89±17	88±18
Cardioinhibition	76± 7	89±10	83±14
Vasodepression	82±17	90±17	81±16
POTS	67±8	74± 7	77± 7

Phase I includes averages over the first 50% of the tilt interval and Phase II includes averages over last 25% of the tilt interval

Controls include subjects admitted to the syncope clinic at Bispebjerg and Frederiksberg Hospitals, Denmark, who had a typical outcome when testing their autonomic nervous system.

POTS patients may experience dizziness and lightheadedness, but typically do not faint [39] (none of the POTS patients included in this study fainted). These patients may have a reduced central blood volume causing blood pressure regulation to be challenged by changes in intrathoracic pressure due to respiration. This causes pathological fluctuations in blood pressure with phase-shifted heart rate changes elicited by the baroreceptor control system. POTS patients typically have excessive vagal withdrawal leading to inappropriate increases in heart rate by more than 30 beats per minute (>40 bpm in individuals aged 12–19 years), which further reduces cardiac filling due to a shortening of the diastolic filling time. As a result, the heart rate increases while blood pressure oscillates [11].

Cardioinhibitory syncope is caused by excessive pooling of blood in the extremities with few presyncopal conditions, combined with an extreme slowing of HR (bradycardia) or temporary cardiac pause (asystole) of at least 5 seconds [6, 26]. The drop in HR is followed by a rapid fall in BP occurring within a few missed heartbeats for most patients. The result is decreased venous return, reducing the filling of the left heart. The reduced filling of the left heart likely causes the ventricular walls to touch in systole, eliciting the so-called Bezold Jarisch reflex [25, 38], stimulating pressure receptors in the ventricular wall.

Vasodepression syncope is also caused by excessive pooling of blood in the extremities. These patients experience a withdrawal of sympathetic activity, leading to peripheral vasodilatation and a more gradual fall in BP, giving presyncopal symptoms. For patients with this condition, the drop in BP is slower, allowing prominent presyncopal symptoms, including lightheadedness, nausea, dizziness, and visual blurring. For these patients, the Bezold Jarisch reflex likely inhibits sympathetic vasoconstriction resulting in a significant drop in BP, which may or may not be followed by a drop in HR, eventually inducing the syncope.

Mixed syncope patients typically experience both cardioinhibition and vasodepression.

Figure 1 shows two examples of patients with mixed symptoms. One patient primarily exhibits cardioinhibition, and one patient primarily exhibits vasodepression.

The classification corresponding to the above expert diagnosis is denoted $Y^i$, $i=1,\dots ,186$. For any patient, $Y^i$ takes values in the five classes introduced in the previous paragraph. The complete data is thus contained in the set

$${\left\{h^i,b^i,Y^i\right\}}_{i=1}^{186}.$$

The time-series are first subsampled at 20 Hz, down from 1000 Hz in the original signal. Second, the signals are preprocessed using a moving average window with a width of 1000 points (or equivalently 50 seconds). Finally, we normalized each signal by subtracting its global mean for each subject. We denote the preprocessed normalized time series by ${𝓗}^i$ and ${𝓑}^i$ where

$$\left({𝓗}^i,{𝓑}^i\right)\in {ℝ}^{N^i}\times {ℝ}^{N^i},i=1,\dots ,186,$$

with $N^i$ referring to the number of retained sample values for the $i^{th}$ subject.

2.3. Random Forest classifier

We use Random Forests [2, 17] (also called Random Decision Forests) an ensemble learning method for classification. This method constructs a large number of decision trees operating as an ensemble. Each tree outputs a class prediction, and the algorithm selects the best solution by means of voting. We use the implementation encoded in the R RANDOMFOREST function.

The goal of our classification is twofold: (i) to determine a subset of the data that can distinguish the patients, and (ii) to explore the similarity between the mixed syncope and the other three classes. To achieve our first goal, we focus on the 157 patients with a single pathology (i.e., we do not include patients with mixed cardioinhibition and vasodepression symptoms) and choose the markers that maximize the classification rate. Below, we show that high classification rates can be obtained by restricting these markers to simple time sampling of both the normalized HR and BP signals ${\left({𝓗}^i,{𝓑}^i\right)}_{i=1}^{157}$.

As illustrated in Figure 2, for each patient, one marker is placed one minute before the head-up tilt and another is placed one minute after the tilt-down. We parameterize the placement of the remaining markers by partitioning the “tilt-up to tilt-down intervals” into three subintervals

$$\left[0,x_1{T}_i\right],\left[x_1{T}_i,x_2{T}_i\right]\text{and}\left[x_2{T}_i,T_i\right],$$

where $0<x_1<x_2<1$ and $T_i$ denotes the elapsed time between tilts for the $i^{th}$ patient. Next we place $z_j$ markers uniformly spaced in the $j$-th subinterval, $j=1$, 2, 3. For each interval, we retain the sampled values which are the closest in time to

$$\mathbf{\text{interval}}\mathbf{1}:T^{up}+𝓁\frac{x_1}{z_1-1}T,𝓁=0,\dots ,z_1-1,$$

$$\mathbf{\text{interval}}\mathbf{2}:T^{up}+\left(x_1+𝓁\frac{x_2-x_1}{z_2}\right)T,𝓁=1,\dots ,z_2,$$

$$\mathbf{\text{interval}}\mathbf{3}:T^{up}+\left(x_2+𝓁\frac{1-x_2}{z_3}\right)T,𝓁=1,\dots ,z_3,$$

with the conventions that if $z_1=0$, there is no node in the first interval and if $z_1=1$, the first interval only contains the node corresponding to the tilt up time, $T^{up}$.

Fig. 2.

Marker parameterization: the red lines represent the onset of tilt-up and tilt-down. Markers are placed before and after the tilt, and within the tilt interval (between the red lines). The tilt intervals is split into three subintervals $\left[0,x_{1}T\right]$, $\left[x_{1}T,x_{2}T\right]$ and $\left[x_2,T\right]$, where 0 refer to the beginning of the data and $T$ denotes the length of the signal. Each subinterval contains $z_i$ nodes, $i=1$, 2, 3. The nodes (sample times) are marked by red circles. The first and last interval. The optimal numerical values for these parameters are $x_1=0.4999$, $x_2=0.9588$ and $z_1=5$, $z_2=7$, $z_3=3$.

We seek an optimal sampling strategy whereby, within a predefined range, the relative size of the intervals defined by $x_1$ and $x_2$ and the number of points $z_1$, $z_2$ and $z_3$ in each interval are chosen to maximize the classification rate. More precisely, each choice of $\xi =\left(x_1,x_2,z_1,z_2,z_3\right)$ defines a subset of the available data ${𝒟}_{\xi }$ with

$${𝒟}_{\xi }={\cup }_{i=1}^{157}{𝒟}_{\xi }^i,$$

where ${𝒟}_{\xi }^i$ is the subset of the data for the $i^{th}$ subject corresponding to $\xi$. We construct a cost function through 10-fold cross validation, namely, ${𝒟}_{\xi }$ partitioned as follows

$${𝒟}_{\xi }={\cup }_{k=1}^{10}{𝒟}_{\xi }^{{\sigma }_k}\text{with}{𝒟}_{\xi }^{{\sigma }_k}={\cup }_{i\in {\sigma }_k}{D}_{\xi }^i,$$

where the ${\sigma }_k$’s partition $\{\text{1},\dots ,\text{157}\}$. Algorithm 1 defines a function $F$ which inputs $\xi$ and returns the successful classification rate computed through cross validation on the Random Forest model. Note that $F$ inherits the Random Forest model’s stochasticity: two calls to $F$ with the same input parameters may lead to two different outputs. However, the stochastic aspect is mostly negligible as numerical experiments indicated that classification rates for the same parameterization have small variations when the model is run multiple

times. We used 10-fold cross-validation, which gave a good approximation of the classification rate attained, with leave-one-out cross-validation while allowing for a 20-fold speed-up.

We find the optimal markers by solving the maximization problem

$$\underset{\xi }{\mathrm{argmax}}F(\xi )\text{subject}\text{to}\left\{\begin{array}{l}0<x_1<x_2<1,\hfill \\ z_i\text{integer},i=1,2,3,\hfill \\ z_i\ge 0,i=1,2,3,\hfill \\ 12\le z_1+z_2+z_3\le 16,\hfill \end{array}\right.$$ (1)

where the last constraint is chosen through trial and error; the retained choice balances the amount of information and the associated cost. To maximize $F$, we first fix $z_1$, $z_2$, $z_3$ and consider the function mapping from $\left(x_1,x_2\right)\mapsto F\left(x_1,x_2,z_1,z_2,z_3\right)$ as the objective function. We optimized it using the L-BFGS-B option in the R optimx function. This is repeated for every possible combination of $z_1$, $z_2$, $z_3$ satisfying the constraints. The initial iterate is taken as (.5, .75); for all data sets numerical convergence was reached in 10 iterations or less in all cases. The resulting optimal parameterization is given is illustrated in Fig. 2 with optimal numerical values $x_1=0.4999$, $x_2=0.9588$ and $z_1=5$, $z_2=7$, $z_3=3$.

Results showed that we need 17 nodes (including one pre-tilt and one post-tilt node), i.e., HR and BP data analysis requires 34 nodes (17 BP and 17 HR markers). Moreover, we found that most of the critical information is concentrated immediately before the tilt down. Using the optimal classification rates corresponding to each choice of $z_1$, $z_2$, $z_3$ we compared 605 parameterizations with mean 93%, median 94%, min 69% and max 97%. Since half of the potential parameterizations gives a classification rate between 94% and 97%, we conclude that the classification rate is not sensitive to perturbations in the parameterization.

2.4. Clinical metrics

To determine if machine learning improves classification, we compare results with common clinical metrics used to discern the patient populations. Several of these rely on frequency analysis; therefore, we computed all metrics using data from the first 5 minutes of the HUT (if available). Five data sets had less than 3 minutes of HUT data and were not analyzed. To quantify heart rate variability (HRV), we compute the standard deviation of successive differences of RR intervals (SDSD). This measure is one of the most widely used metrics of HRV and is closely related to the standard Poincaré map metrics (SD1 and SD2), which provide a geometric representation of patient heart function in clinical settings [3]. To assess beat to beat baroreflex sensitivity, we use the spontaneous baroreflex method (SBR) [29]. This technique has been shown to provide similar insight as pharmacological methods of baroreflex sensitivity [30]. The above traditional metrics assume stationarity of the signal. However, tachycardia and/or syncope in many of the patient’s HUT data suggests that the signals are highly nonstationary. For this reason, we also compute metrics using the uniform phase empirical mode Decomposition [14], which is based on nonstationary signal processing. With this method, we compute the amplitude of the 0.1 Hz component of the HR signal, and the phase difference metric $M_h$.

3. Results

3.1. Classification

First, we classify the controls and the POTS, cardioinhibitory, and vasodepressive, patients. Using the optimal sampling strategy from the optimization problem (1) with leave-one-out cross-validation, we obtain a classification rate of 96%. This result demonstrates the predictive power of the markers and the potential value of machine learning as a complementary tool for syncope diagnosis.

Next, we train the Random Forest on the 157 (non-mixed syncope) subjects and predict the classification of the 29 mixed patients. The Random Forest model classifies these patients using a majority vote over 500 classification trees. The proportion of votes provides a measure of confidence the model has in its classification. Fig. 3 shows the subjects plotted using barycentric coordinates of the proportion of votes. The color legend identifies the classification from expert clinicians. All but two controls (98%) were classified correctly. The two “wrongly” classified controls were both termed vasodepressive, likely due to experiencing a slow drop in BP following NG administration; neither patient experienced syncope before being tilted back down. One POTS patient was characterized as a control (a success rate of 92%). This patient experienced an increase in HR of about 20 bpm, below the threshold for POTS. Closer examination of this patient reveals that both HR and BP are oscillating, a characteristic of patients in this group [14]. One patient with cardioinhibition was characterized as a control (a success rate of 96%). As expected for patients in this group, HR fell before BP, but the lag between the drop in HR and BP was shorter for this patient than for other patients in this group. Two patients diagnosed (by the expert) as vasodepressive were classified as control, a success rate of 93%. Both patients experienced a fast drop in BP following NG administration, similar to the control group, but fainted - characteristic for patients with vasodepression. Finally, one patient with mixed response was characterized as a control (success rate of 97%). Examining this patient’s data reveals that neither HR nor BP drops following the tilt; therefore, the HUT is inconclusive for this patient. An important note is that the classification algorithm was able to sub-classify patients with mixed responses in two groups, with 11 patients belonging to the vasodepression group and 17 patients belonging to the cardioinhibition group, picking up differences between patients in this group (illustrated in Fig. 1).

Fig. 3.

Barycentric coordinate representation of the classification of four classes and mixed patients (denoted by a ◊). Misclassified subjects are denoted by a Δ. Point tightness indicates how well-defined a specific class is.

To further analyze classification, we compare heart rate and blood pressure using ANOVA between groups comparing the various parts of the recorded time -series signal. Results (Table 3) show that this analysis can only detect statistically significant differences between the POTS group and other groups. Comparisons of age reveal a difference in age between POTS and cardioinhibitory patients and the control, vasodepression, and mixed patients. We also used a multivariate ANOVA analysis (via the R function manova) comparing the mean HR and BP features reported in Table 2. Results (see Table 4) show that patients in all groups differ from the control group (p < 0.01). This analysis also clearly distinguishes POTS patients from cardioinhibition and vasodepression (p < 0.01), while it cannot distinguish cardioinhibition from vasodepression.

Table 3.

Pairwise ANOVA analysis (p < 0.01) between groups comparing age, the mean blood pressure (BP), and heart rate (HR) at rest, during the first 50% (Phase I) and the last 25% (Phasee II) of the tilt.

Metric	Group I	Group II
BP - Phase 0 (Rest)	-	-
BP - Phase I (First 50% of tilt signal)	POTS	Control
BP - Phase II (Last 25% of tilt signal)	-	-
HR - Phase 0	-	-
HR - Phase I	POTS	Healthy
		Vasodepression
		Cardioinhibitory
		Mixed
HR - Phase II	POTS	Cardioinhibitory
		Mixed
Age	POTS	Vasodepression
		Mixed
		Control
	Cardioihibitory	Vasodepression
		Mixed
		Control

Phase I includes averages over the first 50% of the tilt interval and phase II includes averages over last 25% of the tilt interval

Table 4.

MANOVA analysis comparing all the statistics in Table 2 across groups. The Table lists comparisons with $p$-values smaller than the 0.01 threshold.

Group I	Group II
Control	Cardioinhibition
Control	Vasodepression
Control	POTS
Control	Mixed
Cardioinhibition	POTS
Vasodepressor	POTS

Finally, pairwise ANOVA comparisons of standard markers described in Section 2.4 and reported in Table 5 reveal that some of the metrics can distinguish POTS and cardioinhibitory from the other patient types, but that none of the metrics can differentiate control, vasodepression, or mixed patients or POTS and cardioinhibitory patients.

Table 5.

Statistically significant comparisons (p < 0.01) from pairwise ANOVA comparisons among all five groups via the following methods: Standard deviation of successive differences (SDSD), Spontaneous Baroreflex sensitivity (SBR), and the nonstationary metrics $M_h$ and $a_{HR}$.

Metric	Group I	Group II
SDSD	Cardioinhibitory	Control, Vasodepression, Mixed
SBR	Cardioinhibitory	Control, Vasodepression, Mixed
$M_h$	POTS Cardioinhibitory	Control, Vasodepression, Mixed Control, Mixed
$a_{HR}$	POTS	Control, Vasodepression, Mixed

3.2. Clustering

The clustering algorithm, K-means, takes the desired number of clusters, $K$, as an input argument. To explore possible data structures, we repeat the clustering analysis with various values of $K$. Clustering results, shown in Fig. 4, are displayed using Silhouette plots [37]. Silhouette values greater than 0 indicate that the patient fits best in its cluster; values less than 0 indicate that it fits better in another cluster. When clustering with $K=3$, left panel of Fig. 4, we note that the POTS patients are clustered with the controls. This is likely explained by the observation that HR and BP changes following the HUT display similar features (BP and HR do not change significantly, as these patients did not faint). Clustering of patients with single pathology disease with $K=4$ gives four groups; see the center panel of Fig. 4. Next, for $K=5$, two different groups of controls emerge, see the right panel of Fig. 4, while the three single pathology groups (patients with POTS, cardioinhibition, and vasodepression) form their own cluster. Finally, taking $K<3$ or $K>5$ did not yield any informative cluster structure. We also observed, in agreement with the classification results, see Fig 3, that vasodepression is the most challenging pathology to characterize.

Fig. 4.

Silhouette representations of the clustering of the population in 3 (left), 4 (center) and 5 (right) clusters.

with mixed patients characterized either as cardioinhibition or vasodepressive

Further investigation reveals that the two “control” clusters (obtained for $K=5$) include patients with different features. Fig. 5 displays the average (over subjects in the cluster) HR and BP plotted as a function of time. Since the experiment’s duration differs slightly for each subject, the horizontal axis in Fig. 5 is a scaled time where the interval between the start and end of head-up tilt is mapped to the interval [0,1].

Fig. 5.

Normalized BP and HR evolution for the two healthy clusters from Fig. 4; normalized signals are obtained by subtracting the individual global temporal mean from the original signal. Left: BP; the healthy subjects demonstrate two different behaviors; right: HR; the healthy subjects display a similar behavior. Group A and group B are represented by the blue curves and green curves, respectively.

The vertical lines indicate the position of the markers. We refer to the two clusters as group A and group B. While HR is similar for the two groups, BP differs significantly. Controls in group A, represented by the blue curves, experience a drop in BP following nitroglycerine, while patients in group B, represented by the green curves, do not. It should be noted that the two controls and the two vasodepression patients who have wrongly classified all express features similar to patients in group A

For each group, a representative subject is chosen (A and B). Their HR and BP is displayed in Fig. 6. The differing trends in BP are observed comparing the left and right panels of Fig. 6.

Fig. 6.

HR (top) and BP (bottom) for two representative patients from control group A (left) and B (right).

3.3. Traditional metrics

To better judge the results discussed above, we test if data can also be categorized using standard autonomic regulation measures. The specific focus is on analyzing heart rate variability (HRV) and baroreflex sensitivity. These methods can distinguish patients with POTS and cardioinhibition, but not control subjects, patients with vasodepression, or mixed pathologies.

4. Discussion

This study uses machine learning to classify and cluster patients with orthostatic intolerance manifested as POTS, cardioinhibition, and vasodepression. Data analyzed include HR and BP measurements from a control group and four groups of patients, three single pathology groups, and one mixed group with patients exhibiting cardioinhibition and vasodepression.

Results of our study show that supervised machine learning–here in the form of Random Forests–can differentiate controls and patients with orthostatic intolerance; more specifically, we show that our method can classify more than 93% of all single pathology patients. We also showed that our classification scheme can identify patients with mixed cardioinhibition/vasodepression pathology, a feature we could not determine by simple ANOVA analysis of BP and HR features. Finally, it should be noted that these results were achieved without adding a marker identifying if patients received nitroglycerine.

Of the 186 datasets, only seven patients were classified incorrectly. Of these, we were unable to distinguish between control and vasodepression for five patients. Both pathologies cause a drop in BP after the administration of nitroglycerine. In addition, one patient experiencing cardioinhibition was characterized as a control. For this patient, the HUT did look normal, but the diagnosis may have been based on other tests not analyzed here. One POTS patient was classified as normal, but closer inspection revealed that HR only increased by 20 bpm following HUT, less than the conventional characterization of POTS. However, this patient did exhibit oscillations in both HR and BP, which is common in POTS patients [14, 32, 33] and the patient is therefore likely diagnosed correctly. Since the machine learning algorithm used in this study is based on an analysis of average BP signals sampled at only 17 points during the tilt-protocol and is therefore likely not able to capture this feature.

Findings reported here advances previous studies [19, 20, 21, 41] differentiating controls and patients with syncope. Various degrees of success have been reported depending on the type of markers/features considered, e.g., time-domain versus frequency domain, the population size (large versus small), the methods (linear versus non-linear analysis), and the amount of information taken into account. These studies were set up (at least in part) to investigate early prediction of syncope, to determine if it is possible to identify patients susceptible to syncope before they faint, i.e., as early as possible during the HUT. However, to our knowledge, no studies were able to identify markers for this purpose. The study reported in this article was not conducted with this question in mind. Nonetheless, our results shed light on this problem.

Analysis of optimal marker location, discussed in Subsection 2.3, reveals (not surprisingly) that markers placed immediately before the onset of syncope are the most important, i.e., the data collected just before tilt-down, corresponding to interval 3 above, are essential to characterize the disease type. This observation is confirmed by [21], who noted that data analysis from the first 15 minutes following tilt-up is inadequate for a specific classification of the patient, as the data “miss” that critical time. The results in [41] are encouraging, though they use data in the last minute before syncope in over half of their results.

The focus of this study is on multi-class classification and clustering of syncope data. We are not aware of similar published studies that show similar results. A possible explanation for the dearth of closely related work might be the difficulty of defining these specific classes, as it requires close collaboration with clinicians examining data from a significantly large cohort of patients.

Unlike other recent analyses of syncope data, such as [19], we do not retain as features quantities explicitly dependent upon the time-frequency analysis of the two signals BP and HR; instead, we sample the signals at optimized times. We tested and observed that the inclusion of “variation dependent features” does not lead to higher classification rates, though we expect that properly chosen quantifiers based on local spectral properties likely can improve our analysis; this is the topic of ongoing investigations. The classification results validate our choice of features and facilitate our understanding of the mixed syncope patients. While clustering largely confirms the validity of our initial clinical classification, it does uncover the existence of two distinct control groups. The two control groups differentiate patients who can maintain BP in response to nitroglycerine versus those who experience a BP drop, though not sufficient to experience presyncope or syncope; all patients in the control group were non-symptomatic (they did not faint). One explanation is that the control subgroup experience a BP drop following nitroglycerine administration caused by their sympathetic system operating near or at its maximum (before vasodilation induced by nitroglycerine). As a result, these patients may not be able to maintain a high BP through vasoconstriction in response to nitroglycerine.

Overall, our analysis of HR and BP time-series give a more accurate classification than traditional methods, including HRV (via the standard deviation of successive differences of RR intervals), BRS (baroreflex sensitivity), $M_h$ the phase between heart rate and BP, and $a_{HR}$ the amplitude of the 0.1 Hz heart rate frequency. These methods can distinguish patients with POTS and cardioinhibition, but not control patients, patients with vasodepression, or mixed pathologies. There are many other useful metrics that we did not compute, such as those reported in [12, 13]; however, we chose to examine the above metrics due to their widespread clinical use or their ability to handle nonstationarity.

This study focuses on analyzing HR and BP time-series data measured during HUT, but other physiological signals also impact autonomic function. Examples include respiration [22, 28, 42], cardiac output [24], age [23], and gender [18]. Respiration can be measured directly or determined from the ECG signal by extracting the height of the RR intervals. This approach was used in the study by Randall et al. [32], analyzing data from patients undergoing the Valsalva maneuver. We did not add respiration in the analysis presented here, since it is not measured, and we do not want to add a second identifier extracted from a signal already analyzed. Besides, several modeling studies (e.g., [43]) have shown that adding cardiac output improves prediction. The Finapres can approximate cardiac output, but this estimate is not reliable [34]; and therefore, it was not saved in the patient records. As a result, we could not estimate peripheral resistance, as it relates to BF and BP. Cardiac output can be measured, e.g., using echocardiography, and future studies should examine if adding any of these estimates would improve the classification. While it is not likely that these measures improve the classification success, they may affect marker location and help use fewer data to classify the disease categories.

Tables 2 and 3, show that POTS and cardioinhibition patients are younger than controls, and patients with vasodepression, and mixed a pathophysiology, and that age can be used as a marker to distinguish some pathologies. However, patients included in this study are not age-matched, as data represents the cohort of patients referred to the syncope clinic, and we do not know if the age groups examined are representative for each pathophysiology. Therefore, we did not include this marker in our clustering analysis. Our objective was to include as little data as possible, and even without accounting for age, our algorithm was able to classify most patients correctly. Future work should test if adding age or any of the standard clinical markers would improve the classification.

Finally, the data analyzed are skewed as all patient groups have more female than male participants. This could result from two factors a) that females faint more frequently than males (as reported by [8, 18, 36] or b) that females may be more likely to seek care when they faint.

Future work will involve the clustering analysis of patients with symptoms that do not fit the pathologies considered here, including patients with dysautonomia and postural hypotension. Further research is also necessary to investigate possible pathophysiological characterizations of the above two control groups and potentially include frequency information to identify patients in the POTS group. Moreover, we propose to compare machine learning methodology (discussed here) with direct mathematical modeling accounting for mechanics in the system. The latter promotes a better understanding of underlying causes leading to observed quantities and facilitate testing of different root causes. Finally, to enable faster diagnoses, we propose to determine how much data is needed for effective classification.

Fig. 7.

Additional types of syncope pathologies are not included in this project. Left: dysautonomic response. HUT does not lead to a significant increase in HR, likely due to reduced vagal response; sympathetic system regulation may be intact but cannot keep up with the progressive drop in central blood volume due to capillary filtration of fluid from the intra- to the extravascular compartment. Middle: postural hypotension. HUT causes an excessive drop in BP with some change in HR. This could be due to reduced sympathetic vasoconstriction. The small change in heart rate could be caused by intact vagal stimulation. Right: orthostatic intolerance. During HUT, reduced central blood volume causes BP regulation to be challenged by intrathoracic pressure changes due to respiration. This causes pathological fluctuations in BP with phase-shifted changes in heart rate elicited by the baroreceptor control system.

Biography

Steven Gilmore, PhD Student in Mathematics at NC State University. Contributed to writing the manuscript and was responsible for clustering and classification analysis.

Joseph Hart, PhD, Researcher at Sandia National Laboratories. Developed the machine learning classification methods for time-series data while he was a graduate student at NC State University.

Justen Geddes, PhD Student in Biomathematics at NC State University. Was responsible for ANOVA analysis, computed standard indices for baroreflex regulation, and contributed to clustering and classification analysis.

Christian H Olsen, PhD, Senior Statistical Programmer at Novo Nordisk, Denmark. Extracted and analyzed data from patient records and contributed to physiological interpretation of results from classification algorithms.

Jesper Mehlsen, MD, Surgical Pathophysiology, Rigshospitalet, University of Copenhagen. Was responsible for collecting data, providing expert classification, and interpretation of classification results.

Pierre Gremaud, PhD, Professor of Mathematics at NC State University. Expert in mathematical modeling and data analysis. Helped design the machine learning classification protocols used in this study.

Mette Olufsen, PhD, Professor of Mathematics at NC State University. Responsible for completing the manuscript, designing the study and interpreting the machine learning classification results.