Differentiating mental and physical stressusing a wearable ECG-device

Transcription

1 Department of Biomedical Engineering Linköping University Master Thesis Differentiating mental and physical stressusing a wearable ECG-device Linnea Nordvall June 16, 2015 LiTH-IMT/BIT30-A-EX 15/523 SE

2 Abstract The company Linkura uses ECG-measurements to detect stress in people s daily life. The term stress often refers to mental stress, which is caused by psychological factors and can (in case of long term exposure) cause several diseases. However stress can also refer to physical stress, such as exercise, which in principle prevents disease. This thesis aims to identify and differentiate these two states of stress in long-term ECG-recordings. This is done using different ECG-derived parameters and the deviation from their respective recovery state values. An algorithm that extracts one-minute segments corresponding to recovery, from long term ECG-recordings, was developed and evaluated. In the final algorithm a combination of the ECG-parameters - heart rate (HR), RMSSD (a measure of heart rate variability) and noise (in the ECG baseline), was used. In order to observe how mental and physical stress affect different ECG-parameters, a study using Linkura s wearable ECG-device and different stress provocations was performed. An accelerometer was included to investigate the possible improvements. After the study, different ECG-parameters were extracted, and the relative deviation from recovery was calculated for each of the different parameters. The parameters showing the greatest difference between rest, mental stress and physical stress were used in the final algorithm. This algorithm was divided into two steps. In the first step the algorithm used the parameters - HR and noise, to separate stress-free states from states of stress (mental or physical). In the second step, the parameters- RMSSD and respiration rate, were combined in order to separate mental stress and physical stress. The algorithm could separate mental stress and higher intensity physical stress, but had some trouble when separating mental stress from lower intensity physical stress. When adding accelerometer data to the algorithm a more distinct separation of the two stress types was obtained. However using accelerometers in real life applications can be troublesome since the relation between acceleration and physical stress varies between different activities. i

3 Acknowledgements This thesis work was executed at the company Linkura AB in collaboration with the Department of Biomedical Engineering at Linköping university. Special thanks to: Supervisor: Anders Johansson Examiner: Göran Salerud I would also like to thank Linkura for the opportunity to do an interesting thesis work, and also everyone that attended my study. ii

4 Contents 1 Introduction Background Aim and question formulation Limitations Background Mental stress Physiological responses to mental stress Physical stress Physiological responses to physical stress The heart and the ECG Measuring ECG ECG components Long term ECG-recordings Linkuras wearable ECG-device Stress and ECG Mental stress Physical stress Linkura s parameters related to stress Alternative ways to measure stress Mental stress Physical stress Automatic recovery state detection Method and materials Materials Selection of ECG parameters Combining the parameters Algorithm Validation of the algorithm Results Separating mental and physical stress Materials and method ECG-device Accelerometer Stress study Stress algorithm Results Stress study Stress algorithm iii

5 CONTENTS 5 Algorithm demonstration Performing the demonstration Processing the demonstration data Performance of the algorithm Discussion ECG parameters Stress algorithm Accelerometer measurements Noise and false detections Choice of method Recovery detection Stress differentiation Stress and the ECG Future improvements Conclusions 45 iv

6 List of Figures 2.1 Anatomy of the heart The dominant vector and ECG ECG-curve Linkura s wearable ECG-device ECG-recording chest strap Histogram heart rate Histogram RMSSD Histogram noise ECG with false detections ECG with noise Recovery ECG Attachment of the ECG Accelerometer used in the study Accelerometer attached on the ECG-device Structure of the stress study Box plot of heart rate for the provocations Box plot of RMSSD for the provocations Box plot of normed-noise for the provocations Box plot of noise for the provocations Box plot of respiration rate for the provocations Box plot of RT-distance for the provocations Box plot of RT-ratio for the provocations Box plot of the activity measure for the provocations Box plot of recovery measure Box plot of stress measure without accelerometer Box plot of stress measure with accelerometer Mental and physical stress during demonstration v

7 List of Tables 3.1 Validation of the rest-algorithm Means and standard deviation rest and not rest Means and standard deviation parameters and provocations Recovery activities during demonstration Physical stress activities during demonstration Mental stress activities during demonstration vi

8 List of abbreviations All abbreviations and acronyms are also explained in the text, but are listed here in alphabetical order. AV-node ECG EMG GAS GSR HR HRV RMSSD SA-node Atrioventricular node Electrocardiography, Electrocardiograph, Electrocardiogram Electromyography General adaption syndrome Galvanic skin response Heart rate Heart rate variability Root mean square of successive differences Sinoatrial node vii

9 Chapter 1 Introduction This report describes the work of a thesis performed as a collaboration between the company Linkura and the Department of Biomedical Engineering at Linköping University. 1.1 Background The most common reason for sick leave at Swedish workplaces today is mental stress. This stress has also been proven to lead to an increased risk for cardiac disease, diabetes and depression and is therefore one of the biggest health problems today [1]. The term stress often refers to mental stress, which is a physiological load on the body caused by psychological factors. More correct is to talk about stress as either mental or physical. Physical stress refers to a stress caused by any physical activity such as a load during physical exercises. While mental stress increases the risk of cardiac disease, regular physical stress is claimed to decrease the risk of cardiac diseases. As the society gets more sedentary many people do not get the required physical activity each day. A counter part to these two stress types is recovery. Any type of stress exposure needs to be followed by a period of recovery in order for the body to recover from the stress reaction. This process helps the body to restore the systems and in that way avoid any stress-related diseases. The lack of both recovery and physical activity is why stress is one of the largest health problems today. Linkura is aim to make people with high mental stress and low activity more aware of the downside of their lifestyle and assist them to change it to a more healthy lifestyle. One of Linkura s products is called, Lifestyle analysis, an analysis is performed during three days where the subjects are logging their food intake, sleep and perceived level of stress, both mental and physical. During these days the subjects are also wearing Linkura s wearable recording device. The device measures and stores the Electrocardiogram(ECG) from the heart. Based on the logged and recorded data Linkura analyzes the life style of each subject. The recorded ECG is analyzed in order to find periods of stress, both physical and mental during the three days. Since the ECG is affected in a similar way by both physical and mental stress. For this reason, it is of interest to develop an ECG-based algorithm that differentiate between these two types of stress. This thesis will focus on the possibility of using different ECG-derived variables and their respective deviation from a recovery state in order to differentiate mental and physical stress. 1

10 CHAPTER 1. INTRODUCTION 1.2 Aim and question formulation The aim of this thesis is to investigate a possibility of separating mental and physical stress in long term ECG-recordings, using different ECG-derived parameters. Further, the aim is to investigate which of these parameters that are affected by mental and physical stress respectively, and to use this to separate these to different stress states. This will be done by answering the following questions: 1. What parameters in the ECG are affected by mental stress? 2. What parameters in the ECG are affected by physical stress? 3. How do the different parameters deviate from their recovery values during mental and physical stress, respectively? 4. How can a combination of these normalized parameters be used for separating mental from physical stress? 5. Can this ECG-based approach be improved by adding accelerometer data to the ECG data? 1.3 Limitations This thesis will only focus on long term ECG-measurement with a wearable two-electrode system, and not clinical standard ECG, which usually involves five-or ten-electrode system. The thesis will treat physical and mental stress as separate states and not consider when the two states appear simultaneously. No other physiological measures than those possible to extract from an ECG will be measured to estimate stress. All the measurements during the thesis are performed on healthy people, meaning that no people suffering from stress related disorders have been included 2

11 Chapter 2 Background In this chapter the required theory for understanding mental and physical stress in relation to the ECG is presented. This chapter treats the basic stress physiology and the essential ECG theory, along with detecting stress using an ECG. Also different ways to measure mental and physical stress without ECG are discussed. 2.1 Mental stress Mental stress is according to the Swedish Social Insurance Agency[2] the most common reason for sick-leave in Swedish workplaces. However the term stress can have both a positive and a negative meaning. According to Tortora et al. [1] stress means the reaction to a stressor, which is any disturbance to the human body. In the case of mental stress these stressors are emotional or cognitive and the response is therefore strongly individual. There are mainly two types of mental stress, acute- and chronic. Where acute stress refers to a response to a stressor of short duration, such as running late to an appointment. Chronic stress is the response to a reoccurring stressor or when an individual is exposed to a stressor for a long time. Stress is necessary in order to prepare the body for different challenges. This stress is positive and is called eustress. However the term stress often refers to negative stress, distress, which can cause pathological changes and therefore be harmful for the body [1]. To understand why mental stress can cause harm, one first need to understand the physiological responses to stress, which are described below in section These physiological responses can be used for measuring the stress levels in an individual Physiological responses to mental stress The physiological response to stress is mainly controlled by the autonomic nervous system and the endocrine system. The autonomic nervous system is described extensively by Tortora et al. [1], and is the part of the nervous system that cannot be controlled by will. The autonomic nervous system can be divided into the sympathetic and the parasympathetic division, which cooperate in order to maintain balance in the body, homeostasis. During rest the parasympathetic, also called rest and digest, nervous system is dominating. The aim of the parasympathetic activity is to preserve energy during metabolism activities. When an individual is exposed to a stressor the sympathetic nervous system, called fight and flight,dominates instead. This results in an increased alertness and metabolism in the body. 3

12 CHAPTER 2. BACKGROUND The physiological response to stress starts when the brain by cognitive evaluation decides if a certain event can be classified as stressful or not. This classification is strongly individual and according to Goldstein [3] a balance between the observed or sensed event and expectations or programmed patterns. If the event is characterized as a stressor the body will answer with a stress response. An established model for the stress response is the general adaption syndrome (GAS), which was firstly described by Hans Selye in the fifties [1, 4]. Selye described stress as a three step process: 1. Fight or flight response 2. Resistance reaction 3. Exhaustion The fight or flight response is described as the initial response to stress caused by neurological triggering, which prepares the body to act against danger. The response means that the organs that are most active when acting against danger will need more oxygen, and therefore the cardiac activity is increased. In order to prioritize these organs, organs not active in the process will be inhibited, which among other things mean decreased blood flow to these parts of the body. The fight or flight reaction is followed by the resistance reaction, which is caused by hormones and lasts longer than the first response. The involved hormones enable faster metabolism by the liver which results in increased levels of glucose, fatty acid and amino acids that can be used by the involved organs to produce more energy. This phase makes the body continue to fight the stressor when the fast fight and flight response has ended. The most important actors during the stress response are described further below. Stress mechanisms: Sympathetic nervous system: The first response to a stress event, increases the heart rate and blood pressure and prepares the body for fight or flight. Epinephrine: Hormone that is produced during stress, that increases the heart rate and affects the dilation of the arteries in the brain. Norepinephrine: Hormone that is produced during stress. The hormone controls the blood pressure by affecting vasoconstriction of the arteries. Norepinephrine is more present during physical demands. Cortisol: Hormone that is released 30 minutes after the stress exposure. Cortisol affects the metabolism of the cells and controls the immune system and rises with emotional challenges and novel conditions, rather than positive stress. The events described, all increase the mental and physical capability of a human in order to cope with a stressful situation. According to Lundberg [5], these reactions are not dangerous as long as the individual has time for recovery after the stress event. Recovery is important in order to get all the body systems back to a normal state again, for example to decrease blood pressure and heart rate. Mental stress such as stress at work can last for a long time without recovery and this is an example of when stress becomes a health risk. The most important recovery activity is sleeping. Unfortunately high stress levels decrease the ability for good sleep, which can lead to a vicious circle. This state is the exhaustion state in the model, which can lead to physical exhaustion such as cardiac diseases or mental exhaustion such as depression. 4

13 CHAPTER 2. BACKGROUND 2.2 Physical stress The term physical stress refers to any physical activity such as standing up, walking or running. The response to physical stress is in many ways similar to the fight or flight response described in section One type of physical stress is aerobic physical exercise, activities that increases heart rate, such as running or biking. These stress activities are opposed to mental stress said to reduce the risk of high blood pressure, depression and anxiety [1] Physiological responses to physical stress Studies [6] have shown that the first reaction to a physical activity is a withdrawal of the parasympathetic activity. This withdrawal increases with the intensity of the activity. If the activity is low to moderate the parasympathetic system will still be activated, but if the activity is moderate to heavy the sympathetic nervous system will dominate [6]. The parasympathetic withdrawal results in increased ventilation and heart rate. Sympathetic activation results in the same fight or flight mechanism as described in section The heart rate will increase further and the vascular resistance in the parts of the body not active in the activity will also increase [7]. This is since the body prioritize to supply the muscles active in the exercise with oxygen and blood. Both Epinephrine and Norepinephrine are released in the blood during physical activity, but Norepinephrine is more sensitive to physical stress than mental stress [8]. The levels of these hormones increase with increased intensity. Unlike the mental stress response, Cortisol is not an active hormone during physical stress. According to the Swedish Stress Research Institute [9], physical stress is important in the recovery process, for those who not have a work that requires physical activity. This is since physical activity serves as a distraction from the daily work. Research has also shown that several endorphins are excreted during physical activity which increases the individuals well-being. Regular physical activity has also been proven to decrease the risk of different cardiac diseases [1]. It is though important to point out that a high level physical activity needs to be followed by a long recovery period to get the state of the different body systems back to normal, otherwise the exhaustion state can occur. This thesis will not go deeper into the differences in physiological effects over time of mental and physical stress, but aims to identifying the two stress types. 2.3 The heart and the ECG The ECG visualizes the different phases of a cardiac cycle by mirroring the electrical activity of the heart with electrodes placed on the chest. ECG-recordings are used to observe abnormalities and trends from an individuals heart. To understand ECGrecordings one first needs to understand the different phases of the cardiac cycle, along with the electrical conduction system of the heart. The primary role of the heart is pump oxygenated blood to the body and to pump deoxygenated blood to the lungs. The anatomy and the direction of the blood flow in the heart are shown in figure 2.1. The heart consists of four chambers: right and left atrium and right and left ventricle. In brief the right side of the heart is responsible for pumping deoxygenated blood to the lungs in order to get oxygenated, and the left side pumps the oxygenated blood to the rest of the body through the Aorta. A cardiac cycle describes an entire heart beat, from the generation until the next heart beat begins. 5

14 CHAPTER 2. BACKGROUND Figure 2.1: Anatomy of the heart: the four chambers and the arteries and veins that leads to them. The arrows shows the direction of the blood flow.[10] The mechanical force of contraction is created by the wall of the heart, the myocardium, which consists of specialized muscle cells. These cells are connected in a conduction system which enables electrical impulses to spread through the heart. These electrical impulses control the contraction of the heart. The activity of the conduction system creates an electrical field that can be measured on the body surface as an ECG-signal. The conduction is initiated by a cluster of pacemaker cells in the upper right atrium, called the sinoatrial node (SA-node). These cells spontaneously fire electrical impulses. The signal then propagates throughout the right atrium,to the left atrium which forces the atria to contract. The frequency of the SA-node is regulated by the autonomic nervous system. During rest the stimulation is less frequent than during stress. From the atria the signal is spread to the atrioventricular node(av-node), located in the lower of the right atrium. Here the signal is delayed due to slower muscle cells, which allows for more blood to fill the ventricles. From here the signal is spread through the bundle of His which connects the atrium and the ventricles. The signal is then spread to the ventricles by a divided pathway and then to the specialized fibres called purkinje fibers. These fibres make the signal spread through the ventricles, which forces the ventricles to contract. This contraction forces the blood to the pulmonary artery and the aorta. The electrical control of the heart can be divided into two phases, depolarization and repolarization which describe activation and recovery of the heart muscle cells respectively. The contraction of the ventricles is followed by repolarization of the ventricles. During the time when the ventricles are depolarized the atria are repolarized. During the time the muscle cells in the myocardium are in the recovery phase the cells can not be stimulated by any electrical impulses Measuring ECG When measuring ECG, electrodes that register the electrical activity from the conduction system are placed on the chest. During clinical ECG-measurements at least five electrodes are used. These electrodes are placed around the heart and measure the surface potential at the skin, which is affected by the electrical field induced by the conduction system during a cardiac cycle. 6

15 CHAPTER 2. BACKGROUND Laguna et al. [11] describe the electrical activity as a dominant vector. This vector is what is measured by the ECG. The theory of this explanation is based on treating every cell in the myocardium as a current dipole. Each cell will then have a vector with different magnitude, orientation and position, depending on the electrical field in that cell at a certain time. The sum of all these vectors will result in a dominant vector, which varies depending on the phase of the cardiac cycle. In figure 2.2 the dominant vector and the corresponding segment of the ECG-curve for a normal cardiac cycle can be seen. The electrical signal that is recorded by the ECG will then be dependent on the orientation, magnitude and position of the dominant vector in relation to the ECGelectrodes. Depending on the direction of the dominant vector in relation to the ECGelectrodes the recorded signal will either be positive or negative. This polarity is what is causing the characteristic components of the ECG-signal explained in the following section. Figure 2.2: The dominant vector during the different phases in a cardiac cycle. (a) Recovery of the heart muscle. (b) and (c) Atrial depolarization. (d-g) Ventricular depolarization. (h) Ventricular repolarization. (i) Recovery period.[11] ECG components A typical ECG curve, for an entire normal cardiac cycle, with its main components is shown in figure 2.3. As can be seen the curve consists of three characteristic waves, the P-wave, QRS-complex and the T-wave. The shape of these waves differs depending on the characteristics of the dominant vector in relation to the electrodes. Therefore the different phases of the cardiac cycle are described by different waves and intervals. 7

16 CHAPTER 2. BACKGROUND Figure 2.3: An ECG curve showing the most important components: The p-wave, QRS-complex and the t-wave and the different intervals.[12] The three different waves are listed and further described below. P-wave The P-wave is a small positive wave and describes the atrial depolarization. The P-wave normally lasts for about 120 ms [11] and has a low frequency content of about Hz. QRS-complex The QRS-complex is the most significant wave of the ECG. It starts with a deflection and continues with a high amplitude triangular wave. The QRScomplex is caused by the fast ventricular polarization and describes how the potential spreads through the ventricles. The QRS-complex also contains the atrial repolarization. The duration of the complex is usually ms. T-wave The T-wave is the last wave of the ECG and describes the ventricular repolarization, which is occurring more slowly than the depolarization. During high heart rate the T-wave is closer in time to the QRS-complex. In addition to the different waves, the intervals between the different waves are important when describing the characteristics of a cardiac cycle. The most important intervals are listed below. RR-interval The length between two R-peaks is called the RR-interval. This interval describes the length of a cardiac cycle. Due to the significant amplitude, the intervals between R-peaks are often used to identify heart beats when processing an ECG. QT-interval The QT-interval is the interval from the beginning of the QRS-complex to the end of T-wave, this interval is depending on the heart rate and decreases with increasing heart rate. The interval describes the length from the ventricular depolarization until the end of ventricular repolarization. A long QT-interval can be an indicator of different cardiac diseases [1]. All these waves and intervals are of great importance when detecting different characteristics in and ECG. 8

17 CHAPTER 2. BACKGROUND Long term ECG-recordings Technical developments have enabled for small ECG-devices that can be worn for long term recording, for instance one or several days. ECG-monitors should not be confused by systems such as heart rate monitors that only measures the heart rate. These are further described in section Long term ECG-recording has many advantages. Recording ECG for a long time, enables analyzing a persons lifestyle. The recordings can be used to detect different physical and psychological states of a person, such as mental stress and physical activity. The measurements can also be used for clinical purposes, such as detecting irregularities of the heart beats that can t be found during short hospital visits. Long term ECG-recordings are though entailed with some difficulties. Many of the difficulties arise from the unpredictability of a human during a day. The ECG-devices are sensitive to motion, which can cause artifacts. In the worst case these artifacts can lead to false detections, since the frequency characteristics of the motion artifacts overlap the essential frequencies in the ECG-curve and can therefore not easily be filtered out. Due to the lower sampling frequency and less electrodes, the ECG-signals that are recorded from commercial ECG-devices are not as accurate as those recorded by clinical systems Linkuras wearable ECG-device The company Linkura uses long term ECG-recordings as a step in measuring peoples lifestyle. Linkura has developed a wearable ECG-device, which can be seen in figure 2.4. The device is attached to a chest strap and can be worn for several days. The device consists of a chest strap with two electrodes and an ECG-device which is attached to the electrodes using snap fasteners. These electrodes measure the surface potential on the chest, which reflects the electrical activity of the heart.the placement of the chest strap is of great importance since a wrong placement of the electrodes leads to a distorted ECGsignal. The ECG-device detects the potential difference, of order 1 mv, between the electrodes and since this is very weak the device contains an amplifier for amplification. The ECG-device contains an HP-filter in order to separate the ECG-signal into an AC- and a DC-signal. The DC-signal is used to detect if the electrodes have contact with the chest and the AC-signal is the ECG-signal that is further analyzed. The sampling frequency of the ECG-device is 128 Hz, which is lower than in ECGsystems for clinical use. The device has a memory of one gigabyte which means that it can store data for 271 hours. The ECG-device also contains a battery that enables ECG-measurements for up to four days. When performing a Lifestyle analysis Linkura s customers wear the ECG-recorder during the waking hours. Linkura then processes the data and returns a report containing graphs which show the periods where the person was mostly stressed and the periods where the person was mostly recovered during the days. The device is also used for measuring and showing physical exertion due to activity. 9

18 CHAPTER 2. BACKGROUND (a) Front of the ECG-device (b) Back of the ECG-device Figure 2.4: Front and back of Linkura s ECG device. (a) Front of the chest strap, with the snap fasteners. (b) Back of the chest strap with the electrodes. Figure 2.5: Front and back of the chest strap, showing the attachment snap fasteners and the recording electrodes. 2.4 Stress and ECG The physiological stress response described in sections and 2.2, affects the heart. Different stress events can therefore be seen in the ECG. There have been many studies on how to identify stressed periods in an ECG [13, 14, 15]. The most common measure is the heart rate variability (HRV), which is a measure of how the distance between heart beats varies. HRV is of significance since it is supposed to mirror the autonomic regulation of the heart. During rest the regulation alters between being sympathetic and parasympathetic in order for the system to be prepared for any input. This regulation results in oscillations of the RR-intervals, which means a high HRV [16]. During stress exposure, the secretion of Norepinephrine and epinephrine results in a more monotonic regulation of the heart which results in a decrease in HRV. There are several measures used for HRV analysis, both time and frequency methods. The time domain methods are using the distance of different RR-intervals to measure the variability, while the frequency methods use the power spectral density and separate the result in different frequency components [16]. Below the most frequently used methods are described. Time domain measures [11]: SDNN: Standard deviation of RR-intervals, often measured over 24 hour-recordings. SDNN is not a good measure for long term recordings since the heart rate changes much between exercise and rest periods during a day. SDANN: Standard deviation of the average of RR-intervals calculated over short periods, often 5 minutes. This measure shows slow circadian variations of the heart rate. 10

19 CHAPTER 2. BACKGROUND RMSSD: Squareroot of the difference of successive RR-intervals, as described in equation 2.1. Different from SDNN and SDANN, RMSSD shows short term variations in the heart rate. Frequency domain measures:[16]: RMSSD = 1 N 1 (RR i+1 RR i ) N 1 2 (2.1) Very low frequency: Power in frequency ranges lower than 0.04 Hz. Unknown physiological correlate. Low frequency (LF): Power of RR-interval signal in frequency range Hz. Considered to mirror both the sympathetic and parasympathetic modulation of the heart. High frequency (HF): Power in frequency range Hz. Influenced by the parasympathetic regulation of the heart. LF/HF-ratio: The ratio between LF and HF, supposed to measure the sympathetic influence on the heart. Studies has shown that the frequency measures are more useful for short term recordings, and time domain measures more suitable for long term recordings [16]. How all these parameters are affected by stress is not yet fully understood and there are disagreements on how to interpret some of the parameters Mental stress Mental stress is associated with an increased heart rate which can be seen in an ECG. The instantaneous heart rate can be accessed using equation 2.2, an increase in instantaneous HR could therefore indicate mental stress. HR = i=1 60 [bpm] (2.2) RR interval[s] Studies regarding measuring mental stress often focuses on HRV. Mainly two types of stress is studied, chronic- and acute stress. Studies have been done where the HRV during a stress provocation is compared to HRV during rest. These studies have shown a decrease in the HRV measures: SDNN, RMSSD, LF and HF during stress compared to a resting state [15]. Such studies show the effect of acute stress. Chronically stressed individuals have shown different responses to stress due to the over activation of the physiological stress response [15]. There has also been a previous master thesis [17] regarding measuring mental stress using a wearable ECG-device. In this thesis the suggestion was to use the length of the QT-interval as a measure of mental stress. The motivation for this is that the hormones active during a mental stress response strengthens the contraction of the heart which should shorten the contraction time and therefore also the QT-interval Physical stress HR increases with increased physical activity and can therefore easily be extracted from an ECG. As described physical activity results in activation of the sympathetic nervous system, whereas the heart rate variability also will decrease. 11

20 CHAPTER 2. BACKGROUND Another measure of physical stress is the respiration rate, since the respiration rate increases during physical stress. There are many ways to extract the respiration rate from the ECG [11]. There have been attempts [18] of using the electromyographic (EMG) components of the ECG to separate mental and physical stress. There have also been studies where the effect of a physical activity is studied and then compared with the same physical activity performed with a mental load. The result of the study showed a decrease in all HRV parameters when the mental load was added to the physical stress [15] Linkura s parameters related to stress Linkura calculates seven different parameters for every ECG-recording, these parameters are calculated as averages of one-minute intervals. These parameters are: Heart rate (HR): Describes the rate of the heart beats. Calculated as the median of RR-intervals, the R-peaks are identified by a QRS-detector. RMSSD: A measure of how much the heart rate varies, described further in section 2.4. Calculated according to equation 2.1. Normed-noise: A measure of the amount of noise in the signal. Noise is defined here as components of the signal that not originate from the heart. The normed-noise parameter is calculated as the standard deviation of the second derivative of the signal in a window between the R-peak and the next heart beat, normalized by the amplitude of the R-peak. Noise: Describes noise in the signal calculated as the normed-noise parameter without the normalization. Respiration rate: Describes in what rate the subject breathes. Calculated as the frequency of the respiratory variations in the RR-interval signal. RT-ratio: A measure of the ratio between the amplitude of the R-peak and the amplitude of the T-wave. The interpretation of this parameter is not fully understood, but is interesting to investigate. RT-distance: A measure of the distance between the R-peak and the end of the T- wave. As mentioned in section 2.4 there have been studies of if this parameter can indicate stress. All parameters except RMSSD and respiration rate are calculated beat by beat. The generated parameter are the average of beat by beat values for a one-minute interval. The parameters RMSSD and respiration rate are calculated minute by minute. 12

21 CHAPTER 2. BACKGROUND 2.5 Alternative ways to measure stress In addition to ECG measurements there are several other ways to measure mental and physical stress. To set the perspective of the studied application, these measurements will be further described below Mental stress There are both subjective and objective methods to measure mental stress. The subjective methods consist of measuring stress using psychological questionnaires. The objective methods measure stress using physiological parameters. There are both invasive and non-invasive methods in this area. The non-invasive methods are for example ECG and galvanic skin response (GSR), and the invasive methods measure stress hormones using for instance blood or saliva samples. The most trustworthy result is probably given by using a combination of these three measurements [19]. Psychological questionnaires Psychological questionnaires use questions in different categories [19]. In the questions used the individual classifies whether he agrees or disagrees to a given claim. Based on the answer the individual is given an indicator on the stress level. There are many different questionnaires to measure stress, some focus on life changing events and some on for example work conditions. The questionnaires are easy to perform and can easily be done on a large part of the population. The main disadvantage is that they are subjective and the individuals might not be honest with the answers. Blood and saliva samples Samples of blood and saliva can be used to measure the levels of different stress hormones [19]. The levels of Norepinephrine, Epinephrine and Cortisol can be measured using blood samples. Only Cortisol can be measured using saliva. The hormone levels can be measured after a single stress provocation or to observe the levels during several days. To get a good result one first needs to have reference values during rest for all hormones. Levels of Epinephrine and Norepinephrine can be measured directly after the stress event, while a change in cortisol level, first can be measured after 30 minutes. Another way is to measure the cortisol levels during a couple of days to see if the levels follow the normal pattern or is deviating. A deviating behavior can indicate stress. The advantage with these measurements is the accuracy. A disadvantage is that the measurements require a high workload and are invasive and therefore not easy to perform. Another disadvantage is that the measurements can not be done continuously. Galvanic skin response Galvanic skin response (GSR), continuously measures the resistance of the skin using electrodes on areas with high amount of sweat glands for example the palm [19]. During a stress event the production from the sweat glands will increase which will decrease the resistance of the skin and increase the galvanic response. In this way the stress can be measured. GSR allows continuous measurements for a long time and is an objective way to measure stress. One disadvantage is that GSR only return a relative measure of the stress, which means that measurements during several activities are needed for comparison [19]. 13

22 CHAPTER 2. BACKGROUND Physical stress Today there are many ways to measure physical activity, both direct measurements which use sensors to measure the activity and indirect measures which use self reporting of physical activity. The main goal of physical activity measure is to evaluate if the individual is getting the recommended dose of physical activity. In this section the most common assessment tools for physical activity: self reports, accelerometers and heart rate monitors will be further described. Self reports Many different tools can be used for assessing physical activity using self report tools, all described by Ainsworth et al. [20]. There are mainly two types: questionnaires and physical activity diaries. The two types of questionnaires are short term and quantitative recall. Short term includes questions about physical activities during the latest month and quantitative is about physical activity during the latest year or even lifetime. The short term is used to see if the individual fulfill the recommendation and the quantitative is used to predict morbidity based on activity behaviors. The physical activity diaries are either of log type or diary type. The log type is a checklist where the individual checks the different physical activities done during the day. In a diary, the individual writes down detailed information about the physical activity during the day. Accelerometers Accelerometers measure the physical activity by measuring the acceleration of a movement in different directions. One type of accelerometer is a pedometer, which only measures the acceleration in the vertical direction. Such accelerometers are called unaxial accelerometers [21]. A more complex accelerometer is the tri-axial accelerometer which measures the acceleration in three directions, and therefore gives a more accurate measure of the activity. The accelerometers are either piezo-resistive or piezo-electric sensor, and since a piezo-resistive sensor is used in this project it will be the main concern here. The piezo-resistive sensor measures the acceleration by a change in resistance in the silicon resistors of the sensor. When the individual perform an activity the acceleration on the accelerometer will result in a change in resistance of the silicon resistors.this change results in an output voltage proportional to the amplitude and frequency of the movement. In this way the movement can be measured. The accelerometers can be used for monitoring duration, intensity and frequency of physical activity over long time periods. The disadvantage with using an accelerometer for assessing physical activity is mainly connected to the questionable accuracy of the accelerometer [21]. The function of the accelerometers also differs with the activity. Ambulatory activities such as walking is easy to detect with an accelerometer, while there are difficulties detecting non-ambulatory activities such as weight-lifting [20]. Heart rate monitors According to Ainsworth et al. [20] heart rate monitors are the most common tool for assessing physical activity. Heart rate monitors usually consist of a chest strap that measures electrical activity from the heart using two electrodes. A microprocessor on the chest strap then processes the signal and returns the heart rate. The heart rate monitor also usually include a wristwatch that receives the actual pulse from the sensor and gives it as continuous feedback to the user. The physical activity can then be divided into different levels depending on the percentage of the maximal pulse of the 14

23 CHAPTER 2. BACKGROUND user. The activity estimation therefore depends more on the individuals reaction the activity compared to accelerometer measurements. The difference between a heart rate monitor and a wearable ECG-device is that the heart rate monitor usually not measures and stores the entire ECG signal. Other wearable measurement systems The ongoing health trend has generated an increased development of technical systems that measure different physiological signals in order to quantify and evaluate the lifestyle of an individual. There are both sensors that measure the activity and sensors that warn you when you have been sedentary for too long. These systems are often multisensor systems, which means that two or more sensors are used to gain information about the individual. These systems can for example combine heart rate monitoring with accelerometer measurements to get more accurate measurements. There are also systems that use the optical properties of the body for detecting both heart rate and heart rate variability. These systems are often worn as bracelets or watches, instead of chest straps. An explanation for this increased use of measurement devices is that the use of smartphones have made these physiological measures more assessable than before. Most smartphones today have a built-in accelerometer, and there are many mobile-applications that uses the accelerometer to measure an individuals activity. There are also many wearable sensors that can be connected to smartphones, this enables for logging of activities, view progression and observe trends. This increment indicates that people are willing of measuring and quantifying their lifestyle. This is very interesting since the population also is getting more unhealthy due to the increased amount of stress. 15

24 Chapter 3 Automatic recovery state detection As described in section 2.3.3, long term ECG-recording enables detection of different physical states such as recovery, mental stress and physical stress. The basic idea of this thesis is to use different ECG characteristics to identify and separate these states in one-day ECG-recordings. The state where the characteristics have been studied most extensively is recovery. The recovery state will be the foundation when separating mental and physical stress. The idea is to use ECG parameters that describe recovery in order to identify an ECGsequence where the person is most likely in recovery. The parameter values for this period will then be used as a reference in order to see how the parameters deviate for the rest of the ECG. The different parameters deviation from the recovery state will then be studied in order to separate mental and physical stress. 3.1 Method and materials An algorithm based on ECG parameters that characterize recovery was developed and validated. The steps when developing the algorithm were to first identify ECG-parameters that are significant for a recovery period. These parameters were chosen from Linkura s ECG-derived parameters described in section The different parameters were combined in order to find ECG-segments with high probability for recovery. The different combinations were tested using one-day ECG-recording and when an algorithm was determined this was tested on different ECG-recordings Materials The algorithm was tested and validated using 35 different one-day ECG-recordings recorded on healthy individuals, which were randomly chosen from Linkura s database. When developing the algorithm, testing was done using 15 of these recordings. When the algorithm was considered finished, validation was done using the remaining 20 recordings. None of the recordings had any reference for the activities done during the recordings, and therefore manual identification of the characteristics of the ECG-signal was used to determine the likelihood of recovery. Matlab was used for the programming needed Selection of ECG parameters The first step when finding the recovery periods was to decide which parameters to include in the algorithm. This was done by studying the different parameters that Linkura calculates and stores for every ECG-recording, described in section

25 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION A recovery period is characterised by low heart rate, which therefore was natural to choose as the first parameter. Because of the parasympathetic activity during rest the heart rate decreases and the heart rate variability increases, therefore RMSSD was chosen as the second parameter. A recovery state is also associated with low motion, therefore the third and last parameter chosen for the recovery identification was the noise-parameter. Also the normed-noise-parameter was considered during the testing. Each parameter-value represents one-minute of the ECG-signal. Therefore the duration of the extracted recovery segments was decided to be one minute. These one-minute ECG-segments will from now be refereed to as the ECG-segment Combining the parameters After choosing the three parameters that were significant for the recovery state, different attempts of combining these were done before the final algorithm was set. Several attempts were done, in this section the three attempts that were studied most thoroughly will be described, together with the reason for the rejection of these methods. The first attempt was to normalize the three parameters with respect to the minand max-value, according to equation 3.1. The one-day histogram for each normalized parameter was calculated, and a small area in each of the histogram was chosen to indicate recovery. The area was determined based on the samples of the parameters from the ECG-recording. norm parameter = parameter min parameter max parameter min parameter (3.1) If for one time instance, all three parameters had values in the allowed areas, the normalized values were multiplied. The smallest product of all one-minute segments was chosen as recovery. However the normalization method is depending on the range of each parameter, and this range has a great impact of the size of the normalized value. Since the RMSSD and normed-noise parameter had a larger difference between the smallest and largest value, these parameters were of very different magnitude than HR. This had a major impact of the result, since all parameters were multiplied. Because of this the method was rejected. The problem with this method was that the normalized parameter values were depending on the range of parameter values. The next attempt was to use the same principle, but to use another normalization method. The parameter values were instead normalized with respect to the position in the allowed area in the histogram. Meaning that the first allowed value got the value zero and the last got the value one. This solved the magnitude problem but did not yield a satisfactory result, since the algorithm returned a lot of noisy periods that were probably not corresponding to recovery. The problem seemed to be that the normed-noise and RMSSD parameters were not so reliable when choosing the recovery period. To solve this problem, the final idea was to let one parameter be dominant when deciding the recovery segment. Since HR is an accepted indicator of recovery this was chosen as the dominant parameter. RMSSD was still used as a rest indicator, but due to the instability of the normed-noise parameter this was replaced with the noise parameter. The position dependent normalization method was replaced by the range dependent normalization method used in the first step. This method showed did generate ECG-segments that could be classified as recovery from all estimation data, and is described in detain in the next section. 17

26 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION Algorithm The three selected parameters: HR, RMSSD and noise, were used to develop an algorithm that extracted one-minute ECG-segments indicating recovery. In short the algorithm uses the one-day histograms for each of the three parameters, with thresholds to exclude values that may contain false detections or other noise. The algorithm uses the parameters extracted from a one-day ECG recording and returns an one-minute ECG-segment corresponding to a recovery segment. The algorithm is described step by step below. 1. Extraction of parameters. The three different parameters: HR, RMSSD and noise, together with the corresponding time instances were extracted minute for minute from the one-day ECG-recording. In order to get every value between 0 and 1 every parameter was normalized according to equation 3.1, where parameter is a vector containing all parameter values for the chosen ECG-recording. 2. Histogram calculation. One-day histograms for the normalized parameter values for each of the three parameters were calculated. 3. Thresholds. Thresholds with respect to the total histogram area were set for each of the three parameters. Meaning that a threshold of 0.8 yields a threshold at 80 % of the area. The thresholds were determined by testing different thresholds and observing the result. These thresholds are set to : HR: The upper threshold was set in order to remove too high HR that can indicate motion artifacts between heart beats, the lower threshold was set to remove too low HR that can indicate non-detected R-peaks. RMSSD: The upper threshold was set in order too remove high variability that can indicate non-detected R-peaks, and the low threshold was set in order to remove low variability that can indicate motion artifacts. Noise: The threshold was set in order to remove high values of noise that can indicate movements and therefore not should be classified as rest. Figure 3.1, 3.2 and 3.3 below are showing examples of histograms with thresholds for the three parameters Heart rate histogram Hr-data Thresholds: Figure 3.1: A one-day histogram for heart rate with thresholds at 0.1 and 0.8 of the total area, based on minute for minute values. 18

27 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION Figure 3.2: A one-day histogram for RMSSD with thresholds at 0.1 and 0.8 of the total area, based on minute for minute values. 60 Noise Histogram Noise-data Thresholds: Figure 3.3: A one-day histogram for the noise parameter with thresholds at 0.1 and 0.8 of the total area, based on minute for minute values. For every parameter the values between the thresholds and the corresponding time instances were extracted in order to be further analyzed. 4. Iterations. In this step the parameter-values that were extracted in the previous step were used. The algorithm iterates over the time instances for the extracted HR-values and compare these instances with the corresponding time instances for the extracted RMSSD-values and the extracted noise-values. If the same time instance exists for all three parameters, the absolute values for all three parameters are saved. If the time instance only exist for HR-values or for only one of the other parameters the values are ignored. This results in a matrix of HR-, RMSSD- and noise-values and the corresponding time instances. 5. Recovery values. The next step was to choose the time instances that corresponded to a recovery period. According to the theory, the parasympathetic 19

28 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION activity during rest results in low heart rate and high HRV. Therefore the time instances with the lowest heart rates are extracted from the matrix created in the previous step. These time instances are suppose to correspond to recovery periods. 6. ECG rest segment. The ten time instances extracted in the previous state were used to extract the one-minute ECG-segments corresponding to each of the time instances. These ten ECG-segments are the resulting selection of ECG recovery segments Validation of the algorithm The algorithm was validated by visual inspection, using the remaining 20 one-day ECG recordings. The algorithm extracted the ten one-minute ECG-segments with lowest heart rate that fulfilled the criteria described in the previous section. Meaning that totally 200 one-minute ECG-segments were automatically selected and analyzed. An ECGsegment selected by the algorithm was considered correct if the two following criteria were fulfilled: 1. The ECG-segment did not contain any false detections. 2. The ECG-segment did not contain any noise that could indicate motion. As a criterion to decide if an automatically generated segment was erroneously selected because of false QRS-detections, the QRS-detector was applied on each generated ECG-segment. The detected R-peaks were compared to the R-peaks in the ECG to see if any peak was missed, or if something that was not an R-peak had been classified as one. A false detection is referred to a detection of noise that is classified as an R-peak. If the ECG-segment contained false detections the segment was not further analyzed. This was since the non-recovery property depended on the performance of the QRSdetector rather than the algorithm. In figure 3.4 a zoomed ECG-segment containing false detections is shown. The red crosses in the figure indicate R-peaks detected by the QRS-detector False detections :06:20 10:06:21 10:06:22 10:06:23 10:06:24 10:06:25 10:06:26 10:06:27 10:06:28 Time Figure 3.4: A part of a falsely selected recovery segment because of fale detections. The red crosses indicate the detected R-peaks. Between the times 10:06:22 and 10:06:25 one can see that noise has been classified as R-peaks and one can therefore not be sure that this is right. 20

29 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION As a criterion to decide if a selected segment was erroneously classified as recovery because of motion induced noise in the ECG-segment, the ECG-segment was compared to another segment in the ECG-signal with low pulse. If this other segment also contained noise the noise was thought to depend on the overall quality of the ECG-signal and not motio, and the signal was classified as recovery. If the other segments did not contain any baseline noise the noise was classified as motion induced and the segment was not classified as recovery. In figure 3.5 a zoomed ECG-segment showing movement induced noise is shown. The noise can be seen by comparing the ECG at time 9:04:02 to the ECG at time 9:04:10 where the ECG contains more noise. 580 Baseline noise :04:02 09:04:04 09:04:06 09:04:08 09:04:10 09:04:12 09:04:14 09:04:16 09:04:18 Time Figure 3.5: A part of an ECG-segment containing noise that could indicate movement The ECG-segment in figure 3.6 is categorized as recovery since it neither contains false detections or baseline noise Rest segment :38:06 17:38:07 17:38:08 17:38:09 17:38:10 17:38:11 17:38:12 17:38:13 Figure 3.6: A ECG segment that neither contains baseline noise or false detections and should therefore be classified as recovery. 21

30 CHAPTER 3. AUTOMATIC RECOVERY STATE DETECTION 3.2 Results Of the 200 automatically selected ECG-segments, 158 were correctly classified as recovery. Of the 42 ECG-segments that were not classified as recovery, 22 contained false detections and 20 contained baseline noise. In every of the 20 studied at least one segment was correctly classified as recovery. Six of the recordings only generated segments that could be classified as rest, while fourteen of the recordings did generate at least one segment not classified as rest. The result of the validation is summarized in table 3.1. Table 3.1: Table of the result of the validation of the performance of the algorithm. One day ECG-recordings for validation Automatically selected ECG-segments Falsely classified recovery segments Recordings with all ten segments correctly classified as recovery Recordings with at least one segment falsely classified as recovery 6 (30 %) 14 (70 %) True recovery 158 (79 %) False recovery 42 (21 %) Because of false detections Because of motion induced noise 22 (52 %) 20 (48 %) The 22 ECG-segments that contained false detections were rejected since the nonrecovery properties were depending on the performance of the QRS-detector rather than the recovery algorithm. In order to observe trends, a comparison of the parameters HR, RMSSD and noise were done between the segments containing motion induced noise and the correctly classified ECG-segments, was done. The mean values and standard deviation for the three different parameters can be seen in table 3.2. Table 3.2: Table of the means ± the standard deviation for the three ECG-parameters HR, RMSSD and Motion in the ECG-segments classified as Res respective not rest. True recovery classification HR [bpm] 71.9 ± ± 8.92 RMSSD[ms] 29.0 ± ± 9.82 Noise ± ± False recovery classification 22

31 Chapter 4 Separating mental and physical stress When detection of the recovery states was enabled, the next step was to develop an algorithm that separates mental and physical stress. In order to do this ECG-measurements during mental stress and physical activity had to be made. 4.1 Materials and method ECG-measurements during mental and physical stress were obtained, by performing a stress study on ten healthy individuals. The ECG and the different ECG-parameters were then analyzed in order to observe which parameters that differed between mental and physical stress. Based on these parameters an algorithm that distinguished between mental and physical stress was developed. In this section the stress study and the development of the algorithm will be further described. In addition to Linkura s ECG-device, an accelerometer also included in the study. The purpose of adding the accelerometer was to see if the algorithm could be improved by adding accelerometer-data ECG-device During the study the subjects wore Linkura s ECG-device, which is described in detail in section The ECG-device was attached to the chest strap and placed in the middle of the chest, which also is illustrated in the figure

32 CHAPTER 4. SEPARATING MENTAL AND PHYSICAL STRESS Figure 4.1: Correct placement of the ECG-device on the chest, and the attachment of the ECG-recorder on the chest strap.[22] Accelerometer The accelerometer used in the study can be seen in figure 4.2. The accelerometer is of the type BMA250, and measures the acceleration in the x-,y- and z-directions. The accelerometer measured between -512 and 512 units, where 256 corresponds to an acceleration of 1g. The accelerometer therefore measured accelerations between -2g and 2g in each direction. The sampling frequency of the accelerometer was 15.2 Hz. Figure 4.2: The accelerometer used in the stress study. The accelerometer also delivered an activity measure, which estimated the activity based on the acceleration in each direction. This measure treated every measurement as a vector and calculated the difference in angle and absolute value between two adjacent measurements. This calculation was done continuously and estimated the activity of the subject. This measures were used when analyzing the accelerometer data. In order to only use one device in the study the accelerometer was attached on top of the ECG-device according to figure 4.3. Accelerometers are further described in section

33 CHAPTER 4. SEPARATING MENTAL AND PHYSICAL STRESS Figure 4.3: Arrangement of the accelerometer and the ECG used in the study. The accelerometer is attached on-top of the ECG-device Stress study A stress study on ten healthy individuals was performed in order to get controlled ECGrecordings during rest, mental stress and physical stress. Students at the biomedical engineering program and employees at Linkura were asked to participate in the study. Seven men and three women between ages 23 and 32 participated in the study. Before the study the participants signed an information of consent, where the essence and the risks of the study were described. Planning the study Planning the study involved finding provocations that reflected mental and physical stress. The planning also concerned writing a protocol and deciding the duration of each provocation. In figure 4.4 a flowchart describing the final design of the study can be seen. To find a proper mental stress test, a literature study on previous stress studies using ECG-recordings was carried out. The most frequently mentioned tests were: public speaking task [23], mental arithmetic task [13, 24] and the Stroop color word test [18]. Since the mental arithmetic test has shown significant effects in the ECG signal [25] and is easy to perform, this test was chosen as the mental stress provocation. To mirror everyday physical activity walking at different speeds was chosen as the physical stress provocation. A lower speed of two km/h was chosen as the lower intensity physical activity and a higher of four km/h was chosen as the higher intensity physical activity. The lower one was chosen to mirror slow walking such as walking around at home or at the office, and the higher one walking with higher intensity. In order to get the heart rate and the other parameters to a steady level the time was chosen to seven minutes for each physical provocation. In the beginning of the test and between the mental arithmetic and the physical task rest periods were placed, to get the parameters to a baseline recovery- level. Another purpose with this period was to get reference values for all parameters 1. The first rest period lasted for 20 minutes and the second was set to 15 minutes. Figure 4.4: Structure of the stress study. 1 Since the study included a known recovery period, the recovery algorithm was not needed when analyzing the data from the study. 25