4 Patient cohorts and data
Our goal is to create a research cohort consisting of obese patients with and without genetic factors, who will share genetic data, biological samples, diet/lifestyle information, and in particular sensor/mobile data (see section 4.2).
4.1 Wireless Body Area Network
As one of the tasks, there is a need for tools such as sensor/mobile data that allow people to measure their physical activity intensity and intake over the course of a day to make informed decisions, and perform energetic trade-offs. Daily information about conditions of health, physical activity level, energy expenditure, and energy intake is central to weight control since energy balance is defined by the interaction between these variables. Wireless Body Area Network (WBAN) is a special kind of autonomous sensor network evolved to provide a wide variety of services. WBAN has emerged as a technology for e-healthcare that allows data concerning a patient’s vital body parameters and movements to be collected by small wearable or implantable sensors and communicated using short-range wireless communication techniques and has shown great potential in improving healthcare quality. Today, WBAN has become an integral component of health care management system where a patient needs to be monitored both inside and outside of home or hospital. Unfortunately, existing technologies in the area of physical activity are mostly designed for individuals who have already achieved a high level of physical fitness and are in good health. Most of them are equipped with heart rate sensor and accelerometer and provides means to track the heart rate, hours of sleep, speed, distance, pace and calories burned. But this is not enough for a medical application. The accurate measurement of physical activity, energy expenditure, and energy intake is challenging and, at present, there is no technology that allows people to measure these variables comfortably, accurately, and continuously over the course of a day and obtain real-time feedback. One of the challenges for this project will be to define technologies (system of multi devices and set of algorithms) that can measure and analyze many physical parameters (such as galvanic skin response. blood oxygen saturation, calorie intake, blood pressure, blood glucose) and that are adapted for use by patients who have difficulty maintaining a healthy weight and minimum levels of physical activity every day. As we will be collecting human patient-related data, safety issues need to be taken into account to avoid any risk to life, as well as security and privacy protection issues for the data collected from a WBAN, either while stored inside the WBAN or during their transmission outside of the WBAN. Controlling congestion and security is a major unsolved concern, with challenges resulting from stringent resource constraints of WBAN devices, and the high demand for both security/privacy and practicality/usability. In this work, we will address two important data security issues: secure and dependable distributed data storage, and fine-grained distributed data access control for sensitive and private patient medical data, as well as considering the diverse practical issues that need to be taken into account while fulfilling the security and privacy requirements.
4.1.1 Health trackers selection
The objective classification of HT is a important condition for the understanding of the complex relationship between health and physical activity. Many different Health Trackers (HTs) were considered during this study. Generally we can divide them by commercially availability, manufacturer ( big brands / start-ups or future projects) and area of applying ( medical / fitness). Further details about these sensors are provided in Table 1.
Table 1 How can we know which one is the best? Various HTs have been tested and validated for field-based research in both healthy and chronically diseased populations [21]. There are many studies about the validity of energy consumption algorithms embedded in different devices [22] of medical area rather than fitness trackers [26]. According [26] the resultats in 2014 of energy expenditure estimation during a 69-minute protocol were following: the mean absolute percent error values performed 9.3%, 10.1%, 10.4%, 12.2%, 12.6%, 12.8%, 13.0%, and 23.5% for the BodyMedia FIT, Fitbit Zip, Fitbit One, Jawbone Up, ActiGraph, DirectLife, NikeFuel Band, and Basis B1 Band, respectively.
The aim of this section was to compare four of the most suitable for our case of study commercially available activity monitors in terms of variety of functions and accuracy of mesurement:
GoBe and AngelSensor are Startups. Jawbone UP24/UP3 and Microsoft Band 2 are two of the most popular fitness trackers in the world.
All of these gadgets have three things in common:
- The main aim is to mesure physical activity and energy expenditure;
- They use sensors to determine things about your body, including your heart rate, breathing, steps taken, and calories burned;
- They connect to computer applications to show you the results pulled from the sensors and to help you track your progress.
These HTs use accelerometry which is combined with physiological sensors such as heart rate, temperature and galvanic skin response for increasing their accuracy of predicting energy expenditure and discriminating activity types. Activity type-specific equations are generally implemented into HTs to model energy expenditure [23]. These equations based on mesurement of step detection, but only a few studies have focused specifically on the accuracy of this estimate. In article [24] authers compared the accuracy of a pedometer and a triaxial accelerometer in their step count estimates in patients with Parkinson’s disease, using video recordings as reference. The error of the pedometer was speed-dependent, ranging between 4.5%- 17.2%, whereas the error of the accelerometer was around 7%. During another work [25] medical HT (Sensewear Armband Mini) was tested in group of patients with chronic pulmonary disease and the results at slow speeds (1.4-1.8 km/h) was not adequately accurate.
There are other things to consider as well, such as data extraction (SDK, API), price, set of sensors, intuitive application interface design, capability to motivate.
4.1.1.1 GoBe
GoBe is a bracelet which measure the wearer’s heart rate, calories burned, sleep, and stress levels. That’s all conceivable, given what the Jawbone UP3 and other body trackers already measure.
But it is interesting by its “patented flow technology”. Click here to learn more about this technology. According to it, GoBe promises to something a little more sensational: Automatically tracking the calories of everything the wearer eats, through his or her skin.
The automatic calorie-tracking, which GoBe claims to do by reading glucose levels in cells, would revolutionize dieting—even the best calorie-counting apps today rely on manual food logging. Fig The premise seems to be lofty, in fact. Let’s say GoBe does measure glucose levels without piercing the skin, as it claims to do. That would be a godsend to diabetics, who, as it stands, must regularly prick their fingers to test blood sugar. The less-invasive technology is probably coming soon, Michelle MacDonald ( a clinical dietician at the National Jewish Health hospital in Denver) told ( read full text) PandoDaily: but when it does it will be the size of a shoebox … It will come from a big lab, will be huge news and make a lot of money. But on top of that, blood glucose is only a rough measure of total energy intake. Eat a tablespoon of olive oil, and you’ve consumed 119 calories, but your blood sugar would barely rise. A very thin slice of white bread, meanwhile, would send blood sugar soaring and only yields 40 calories.
Developer API will be available later. Now data extraction is not posible.
4.1.1.2 Test-dreve AngelSensor M1
AngelSensor M1 is an open source fitness tracker that provides a lot of health data. The main advantage is consept of providing an open platform for developers to create own applications, relying on the raw data streams sent by the sensor via Bluetooth low energy. It opens up varied opportunities, because being able to accurately quantify physical activity is important for epidemiological research so that relations between physical activity, health status, environmental factors, and so on, can be determined. But it’s extremely hard to find open platforms with decent usability & low cost, specially when it comes to physiological data (ECG, PPG, etc.), there is basically nothing on the market. So in this context AngelSensor is outstanding from the authers sensors. However, this is more prototype than final device, some functions like blood oxygen or sleep quality are not ready yet, it will be available later with firmware update. No site to store data – data is sent to mobile application. Temperature measurements are inaccurate -2 – 3 C than usual, but reflect well temperature changes. The same with heart rate parameter. Over 20% of steps are missed, but there is a possibility to read raw data directly to implement other algorithms for step calculation later. So this is a prototype that mostly works, but a dedicated application or API is needed to provide all implemented parameters. It has not only 3-axis accelerometer but alse a gyroscope. Concerning usability of current applications - not all parameters monitoring and no history viewing capability in the current apps version. It is in alfa-version for Android and has more features for iOS. But the project is in active development. Also it provides SDK for developers - quite good for customisation and creation of our owns applications. This is not a device for consumers, but it is a device for developers.
The sensor comes with an open source project, hosted on github, which is quite useful as a starting point for new applications. The documentations is not nearly enought, but support staff is provide some help. We did small example of getting and using data from this sensor for determine if we could applay it in this work for estimating health parameters and measure energy expenditure. The sensor transmit data by using the standard bluetooth low energy protocol. By default it send only heart rate data, but also possiable reciving raw data from two channels PPG and accelerometer. Our application only store these signals to a csv file. A description of the protocols can be found here. After observing data we found that it has enough noise. Here we propose data example from accelerometer: 
The image of PPG raw data is much more difficult, because it seems to be very close to noise. It is explained by fact that PPG signal quality is depended by sensor’s factors such as the layout of the optical sensor, which the wavelength of light is used and the quality of the amplifiers. We need avoid to warping puls of real signal by filterning, in view of the risk of mixing some of the broad-spectrum noise with the signal, while removing some of the higher harmonics of the true pulse waveform. Anyway before feature extraction we have to applay data filter, taking into account that there might be important information in the low and high frequency spectra of our time series. R-package “signal” allow to split them providing butterworth digital filter. In conclusion, using AngelSensor make us free to develop our own system of energy expenditure estimation, but at other side, PPG data seems too noisy for making profond heart rate variability analysis. Concerning accelerometer data, it seems useable. As we had not enought time to collect big dataset by ourself, for activity recognition analysis (important step in measure energy expenditure, see 4.1.2) we will utilize dataset from smartphones [27], which consiste analogical physical activity information from gyroscope and accelerometer.
4.1.1.3 Test-drive Jawbone UP 24
In my informal testing (during several months) Jawbone UP 24 handles burned calories and steps count quite well (sometimes there are +/- 10 % of steps).It provides REST API to collect list of parameters: steps, distance, burned calories, sleep (see below parameters for new UP 3 version). Sleep duration is measured well (but by accelerometer i.e. number of hours in horizontal position), so sleep quality is not accurate, but anyway it attempts to give you additional information such as how heavy or light users sleep was, and how long waking periods were. And this is all that current version provides. In 3.0 version there is also heart rate parameter. Concerning usability of current applications - it is user-friendly and also allows users to log food consumption to help with their dietary goals. The user can set the device to vibrate under certain circumstances. A good option might be to have it vibrate after users haven’t moved for a certain period of time. Let’s look at the approache of data gathering that Jawbone API proposes. API discription can be found here. This is full list of possible scopes for Jawbone U3:
- basic_read
- extended_read
- location_read
- friends_read
- mood_read/write
- move_read/write
- sleep_read/write
- meal_read/write
- weight_read/write
- cardiac_read/write
- generic_event_read/write
Developers have not access Bluetooth or the raw sensors on the devices directly. All information are delived after processing, handling and transformations from the cloud via REST API. The only advanced sensor data available through the UP API is Resting Heart Rate. This provides one value per day, captured when the user wakes up while wearing the band, and is only available from users that have an UP3 band. However they are currently exploring opportunities to unlock access to more sensor data in the future, but do not have a set timeline yet. The information about there based step-classification concept is described in this article [https://jawbone.com/blog/classifying-steps-machine-learning/].
We tested data access via API and confirmed the comfortability of system.
4.1.1.4 Microsoft Band 2
This HT was not tested, but it it deserves attention and may be present the best desition. The Microsoft Band is a commercial product available to the public without any delays and problems. It features big number of important sensors: optical heart rate sensor, 3-axis accelerometer/gyrometer, ambient light sensor, UV sensor, capacitive sensor, galvanic skin response, baromete, haptic vibration motor, microphone, GPS. But the big positive point is that Microsoft has provided a SDK that allows access of the band’s sensor readings. The band communicates via Bluetooth to a synchronized smart phone. So it can present a good solution as it takes access to sensors data like Angel Sensor and has user-friendly application like Jawbone, which is also important for motivating. But the price is higher approximately on 100$.
4.1.2 Energy expenditure estimation
Using parameters from HT we can measure and control the physical parameters, intencity of activity and energy expenditure (EE). And as calories burnt (EE) estimation present the key aspect in obesity determination, this section we dedicate to consept of EE measurement. We quickly investigate existing problems of EE estimation. Then the relation between accelerometer, physiological data and EE will be establised. And at the end we propose models which will be able to increase accuracy level and to take into account variability in physiological signals between individuals applaying method of personalization.
Definition EE
EE = Calories burnt make up half of the energy balance equation:
Energy Balance = Energy Expenditure - Energy Intake.
We can determine three elements of EE:
- Basal metabolic rate (Energy required to maintain body functions - 60-75% of TEE)
- Diet Induced Thermogenesis (Energy used in digestion - 10% of TEE)
- Physical Activity (15-30% of total TEE)
According our method (see section “Method description”) we will take deal with AEE.
How we can measure EE
One of the most standard and practical method of estimate EE is indirect calorimetry - VO2max test. It analyzes inhale and exhale gases concentration and the rate of oxygen consumption (VO2, mL/min) is used as the representation of energy expenditure according Weir’s equation:
Resting EE=3.9×VO2+1.1×VCO2,
where: VO2 = oxygen uptake (ml/min); VCO2 = carbon dioxide output (ml/min).
Metabolic carts offer invaluable insight into exercise metabolism, allowing interrogation of each breath. There are a couple of portable indirect calorimeters on the market, like AeroSport TEEM 100 Metabolic Analysis System, but the participant must use a facial mask during all data acquisition.
Below, we demonstrate the exmaple of perfoming substrate utilisation using metabolic cart data from VO2 R-package collected during:
- a fast-ramp, maximal, graded exercise test;
- a slow-ramp, submaximal, graded exercise test.
For the estimation of substrate utilisation by way of gas exchange we have to prepare oxygen uptake data. Firstly we filtered this data form unwanted noises. Then according to Robergs and col.[45] removed ventilatory and cardiovascular frequency components (leaving theoretically only the muscular oxygen uptake kinetics) via the application of a 0.04 Hz, using third-order Butterworth filter. And finally, we used the equations of Jeukendrup & Wallis[44] which are implemented in the VO2 - package for perfoming energy expenditure (ee), carbohydrate oxidation rates (choox), and fat oxidation rates (fatox). 
It would be interesting to estimate the rate of maximal fat oxidation. Taking into account that the rate of fat oxidation during graded exercise presents as a quadratic function, it’s maximum can be extracted as the vertex of a fitted polynomial:
## x
## 167.4839
Caption
And at the end, we can try to evalue another important parameter - efficiency during exercise. The two dominating methods are calculations of gross efficiency and delta efficiency, were described in this work[46].
In fact, in the begining of research it will be very useful to create for each paitant metabolic profile provides detailed information on there caloric burn rate and individualized heart rate zones by indirect calorimetry. Because it shows the ability of patient to absorb oxygen. And exactly the level of oxygen consumption by the muscles influences to, for example, the ability to run fast to keep this speed long time. It helps to personalize workout plan and to determine the algorithms of EE estimation giving higher accuracy. Moreover, scientists have come to the conclusion that the VO2max indicator is inherited. And the main trouble is that the ability to improve this value is also inheruted. In 2007-2008, the Norwegian researchers conducted big number of testes on the participants of the experiment revealed the dynamics of VO2max. The resultats shown that despite of certain genetics limits everyone improved his physical forme and reached a good level of this indicator in condition of the regular thaining progress. This indicator could help not only to create the accurate system of EE estimation, but also present interesting phenotype for study in this work.
Problems of EE estimation
We have already talked that many HT in the consumer market provide EE estimation in these applications. They applay regression models using accelerometer data, HR, or a combination of the two, across all activities with including anthropometric characteristics in the equation, given the relation between body size and EE. Some of them are claiming higher accuracy due to a combination of accelerometer, gyroscope and physiological data (e.g. Basis or the Apple watch). But the resultats is also not nearly enought, partly because these devices are simply combining multiple signals without personalization. As it has been explained in section “Method description”, the personalization of data is very important and helping to avoid suboptimal results. This fact can be demonstrated very easily in the following example: we want to estimate EE from HR and there are two participants with similar anthropometric characteristics (e.g. similar body weight, etc.), but with different sport level in our experiment; Normaly individuals with similar body size expend similar amounts of energy during a certain activity, however their HR differs during moderate to vigorous activities explaining by the participant’s fitness level. It means we can measure EE from HR if only HR is properly normalized. However, there is a strong link between HR, oxygen consumption and EE, which motivates the use of physiological data for EE estimation. Let’s look at another example: If we use only accelerometer, for which the rationale is that body motion close to the body’s center of mass, we will see the wrong relation between accelerometer and EE for the participant over different activities. Single regression models are unable to fit all the activities,since the slope and intercept of the regression model changebased on the activity performed while data is collected [40]. So the relation gets very weak for non-weight-bearing some types of activities, such as biking, while it’s reasonably good for walking and running.
We can denote two points:
- Accelerometer data alone is unable to fit all activities, biking is modeled as a sedentary activity, due to the lack of motion, and therefore results are inaccurate for both high intensity activities without body motion (i.e. biking) and sedentary activities (lying, sedentary behavior);
- HR could show increased value in low intensity activity after high intensity activity for people without regular sports training.
Improving accelerometer data
For improving accelerometer data almost all modern devices and many researches (like [40], [41]) try to use activity recognition algorithms, where activity recognition is performed in different way. For example, it depends on sensor location, the number and types of clusters in set: it could classify 5 or 50 activities with different levels of accuracy, as well as the set can be mutually exclusive. In this work we propose to create own system of activity recognition using data from HT throw SDK. For that we have to build the machine learning model which will give the resultats closed to VO2max-test. In section “Activity recognition” we make simple example of classificator based on classical machine learning algorithms and show that applaing activity recognition algorithms makes system much more flexiable, for example it gives the ability to selectively use HR data - only when the activity is of a certain type.
Improving physiological data
Activity recognition method brought many improvements in EE measurement, and are now widely used. However we have another problem with HR showing huge variability across individuals and it seems to be not resolved. We have to develop models that fully exploit the potential of physiological data across different subjects, without requiring individual calibration. For resolving this problems some researches propose the following possiable ways:
- Explicit HR normalization [42]
We could normalize physiological data using a person’s physiological signals under high intensity activities. But for avoiding to re-perform individual calibrations according changes the relation between physiological data and physical form of person, we can use low intensity activities of daily living, automatically recognized using different machine learning models, to contextualize HR or any other physiological signal, and then predict what the physiological data during a high intensity activity would be - without having to actually perform that activity [42]. Knowing the “intense” value, we can normalize the signal across individuals. 
- Hierarchical model including estimated fitness as group level predictor [43]
Alternatively, we can use a more sophisticated solution, described in the work [43], which focused on using hierarchical Bayesian models. Bayesian hierarchical approach fits very well in tasks of EE estimation, because parameters are naturally interposed: belong to different levels. It meens that features at the second level influence to features at the first level. We will distinguish two levels - individual (first) level and group (second) level of a hierarchical structure. The group level parameters impact the relation between predictors at the first level and the outcome variable. In this method we have to understand the cause behind the necessity of individual calibration. For example, in case of HR, it is differences in fitness level of participances. So we could measure this level and then create a hierarchical model, using it as a group level parameter. In this case fitness level directly impacts the relation between HR and EE, without the need for explicit HR normalization. How to estimate fitness level is other question and we talked a little bit on this topic in subsection “How we can measure EE”. The principle is used by many sub maximal fitness tests (like Rockport Walk Test), where HR at a certain intensity is used as predictor - together with other variables - to estimate VO2max. The difference is that by using context recognition algorithms, there is no need to perform lab tests or specific exercises. We can take the HR while walking at certain speed as our sub maximal HR and then predict VO2max. Finally fitness level is included as group level parameter for prediction of EE.
Conclusion
- Presence of HR sensor in HT does not meen more accuracy. The approach of estimation is important.
- Better using the accelerometer which worn on the chest, because results would be much worse at the wrist, due to the lack of agreement between whole body motion and wrist movement;
- Combining physiological and accelerometer (+ gyroscope see section “Activity recognition”) data gives the best results;
- Using activity recognition makes system more accurate.
- Method of personalization by normalizing physiological data for individual improves accuracy for moderate and vigorous activities.
- Hierarchical Bayesian models propose a flexible alternative solution to the problem of physiological signals normalization, by installing the relation between HR and fitness level, without HR normalization.
4.1.3 Activity recognition
Physical activity recognition is one of the key problem for determining accuracy of energy expenditure - as one of the main parameters in obesity classification and treatment. Previous subsection describes the components needed to gather the relevant data from a person in context of everyday activity. This subsection shows simple example how machine learning techniques are able to determine the activity from the sensors data for help to increases the accuracy of the measurement results. This approach is validated over real data from physical activity monitoring dataset from smartphones 27
Dataset information
This dataset is very closed to data, which we can take from AngelSensor. The experiments have been carried out with a group of 30 volunteers within an age bracket of 19-48 years. Each person performed six activities (WALKING, WALKING_UPSTAIRS, WALKING_DOWNSTAIRS, SITTING, STANDING, LAYING) wearing a smartphone (Samsung Galaxy S II) on the waist. Using its embedded accelerometer and gyroscope, we captured 3-axial linear acceleration and 3-axial angular velocity at a constant rate of 50Hz. The experiments have been video-recorded to label the data manually. The obtained dataset has been randomly partitioned into two sets, where 70% of the volunteers was selected for generating the training data and 30% the test data.
The sensor signals (accelerometer and gyroscope) were pre-processed by applying noise filters and then sampled in fixed-width sliding windows of 2.56 sec and 50% overlap (128 readings/window). The sensor acceleration signal, which has gravitational and body motion components, was separated using a Butterworth low-pass filter into body acceleration and gravity. The gravitational force is assumed to have only low frequency components, therefore a filter with 0.3 Hz cutoff frequency was used. From each window, a vector of features was obtained by calculating variables from the time and frequency domain.
For each record in the dataset it is provided:
- Triaxial acceleration from the accelerometer (total acceleration) and the estimated body acceleration.
- Triaxial Angular velocity from the gyroscope.
- A 561-feature vector with time and frequency domain variables.
- Its activity label.
- An identifier of the subject who carried out the experiment.
Data Exploration
Firstly we investigated data on preprocessing necessities it was found that all signs are within [-1,1] and as a consequence it is not necessary neither normalization, nor scaling.
Then has checked availability of signs with strongly displaced distribution and show three worst variables:
| v389_fBodyAccJerkbandsEnergy57_64 | v479_fBodyGyrobandsEnergy33_40 | v60_tGravityAcciqrX |
|---|---|---|
| 14.70005 | 12.33718 | 12.18477 |
For selecting criterion of quality of mode (Accuracy or Kappa) it is necessary to to check the activity distribution:
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING |
|---|---|---|---|---|---|
| 1226 | 1073 | 986 | 1286 | 1374 | 1407 |
Cases are distributed almost equally, except maybe walking_down. Thus it can be used Accuracy.
Modeling
In this subsection as example of resolution of classification tasks two classical alhorithms will be applied: Random Forest, Support Vectors Machine.
Random Forest
We use Random Forest as basic algorithm, because it has the built-in mechanism of assessment of variable importance. The quality of model is following:
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 481 | 11 | 4 | 0 | 0 | 0 |
| WALKING_UP | 41 | 424 | 6 | 0 | 0 | 0 |
| WALKING_DOWN | 16 | 42 | 362 | 0 | 0 | 0 |
| SITTING | 0 | 0 | 0 | 427 | 64 | 0 |
| STANDING | 0 | 0 | 0 | 51 | 481 | 0 |
| LAYING | 0 | 0 | 0 | 0 | 0 | 537 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.9202579 | 0.9041334 | 0.9098849 | 0.9297870 | 0.1822192 | 0.0000000 | NA |
| NA. | Class: WALKING | Class: WALKING_UP | Class: WALKING_DOWN | Class: SITTING | Class: STANDING | Class: LAYING |
| byClass.Sensitivity | 0.9697581 | 0.9002123 | 0.8619048 | 0.8696538 | 0.9041353 | 1.0000000 |
| byClass.Specificity | 0.9767442 | 0.9785945 | 0.9960427 | 0.9792345 | 0.9734990 | 1.0000000 |
| byClass.Pos.Pred.Value | 0.8940520 | 0.8888889 | 0.9731183 | 0.8933054 | 0.8825688 | 1.0000000 |
| byClass.Neg.Pred.Value | 0.9937733 | 0.9809717 | 0.9774757 | 0.9740786 | 0.9787677 | 1.0000000 |
| byClass.Prevalence | 0.1683068 | 0.1598235 | 0.1425178 | 0.1666101 | 0.1805226 | 0.1822192 |
| byClass.Detection.Rate | 0.1632168 | 0.1438751 | 0.1228368 | 0.1448931 | 0.1632168 | 0.1822192 |
| byClass.Detection.Prevalence | 0.1825585 | 0.1618595 | 0.1262301 | 0.1621988 | 0.1849338 | 0.1822192 |
| byClass.Balanced.Accuracy | 0.9732511 | 0.9394034 | 0.9289738 | 0.9244441 | 0.9388172 | 1.0000000 |
Support Vector Machine
For comparation we build SVM model. The parameters of model is following:
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 482 | 5 | 9 | 0 | 0 | 0 |
| WALKING_UP | 13 | 457 | 1 | 0 | 0 | 0 |
| WALKING_DOWN | 7 | 32 | 381 | 0 | 0 | 0 |
| SITTING | 0 | 1 | 0 | 442 | 46 | 2 |
| STANDING | 0 | 0 | 0 | 26 | 506 | 0 |
| LAYING | 0 | 0 | 0 | 0 | 0 | 537 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.9518154 | 0.9420842 | 0.9434533 | 0.9592650 | 0.1822192 | 0.0000000 | NA |
| NA. | Class: WALKING | Class: WALKING_UP | Class: WALKING_DOWN | Class: SITTING | Class: STANDING | Class: LAYING |
| byClass.Sensitivity | 0.9717742 | 0.9702760 | 0.9071429 | 0.9002037 | 0.9511278 | 1.0000000 |
| byClass.Specificity | 0.9918401 | 0.9846527 | 0.9960427 | 0.9894137 | 0.9809524 | 0.9991701 |
| byClass.Pos.Pred.Value | 0.9601594 | 0.9232323 | 0.9744246 | 0.9444444 | 0.9166667 | 0.9962894 |
| byClass.Neg.Pred.Value | 0.9942740 | 0.9942904 | 0.9847418 | 0.9802340 | 0.9891441 | 1.0000000 |
| byClass.Prevalence | 0.1683068 | 0.1598235 | 0.1425178 | 0.1666101 | 0.1805226 | 0.1822192 |
| byClass.Detection.Rate | 0.1635562 | 0.1550730 | 0.1292840 | 0.1499830 | 0.1717000 | 0.1822192 |
| byClass.Detection.Prevalence | 0.1703427 | 0.1679674 | 0.1326773 | 0.1588056 | 0.1873091 | 0.1828979 |
| byClass.Balanced.Accuracy | 0.9818071 | 0.9774643 | 0.9515928 | 0.9448087 | 0.9660401 | 0.9995851 |
Important variables
Variable extraction - important variables with RF: 
It is seen that two components of signal (Body and Gravity) from accelerometer presents among the most important variables. But there are no data from gyroscope. In fact, that could be suspicious.
Check important variables on SVM
Basic idea is to substitute 10% of the most important variables for the begining, and then to continue increasing by 10% with control of accuracy. Achieving maximum, will reduce step at first to 5%, then to 2.5% and, at last, to one variable. Model with 440 variables:
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 482 | 5 | 9 | 0 | 0 | 0 |
| WALKING_UP | 13 | 457 | 1 | 0 | 0 | 0 |
| WALKING_DOWN | 4 | 33 | 383 | 0 | 0 | 0 |
| SITTING | 0 | 1 | 0 | 436 | 52 | 2 |
| STANDING | 0 | 0 | 0 | 25 | 507 | 0 |
| LAYING | 0 | 0 | 0 | 0 | 0 | 537 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.9507974 | 0.9408597 | 0.9423595 | 0.9583251 | 0.1822192 | 0.0000000 | NA |
Model with 490 variables:
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 483 | 5 | 8 | 0 | 0 | 0 |
| WALKING_UP | 15 | 455 | 1 | 0 | 0 | 0 |
| WALKING_DOWN | 5 | 31 | 384 | 0 | 0 | 0 |
| SITTING | 0 | 1 | 0 | 440 | 48 | 2 |
| STANDING | 0 | 0 | 0 | 25 | 507 | 0 |
| LAYING | 0 | 0 | 0 | 0 | 0 | 537 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.9521547 | 0.9424917 | 0.9438181 | 0.9595781 | 0.1822192 | 0.0000000 | NA |
As result the maximum of accuracy has appeared around 490 variables and has made 0.952 (on 440 variables - 0,950), that is better on 0,1% than previous value from full set of variables, but training time is half less (see table of Time perfomance)
Second approach: important variables with Information gain ratio
For being determined in choice of important variable, we apply Information gain ratio (IGR). Information gain tells us how important a given attribute of the feature vectors is. After calculation of IGR, the list of variables was arranged in decreasing order of value. The first 20 from this list is following:
Maximum IGR value = 0.897, minimum = 0. As it is seen, IGR gives different set of important variables than RF ( only five variables are the same ).The gyroscope again is not present among the most important value. But it is a lot of variables from the accelerometer X components.
The difference of resulatats could be explained by the fact that in dataset some variables give more information than others. It is well visible if to construct diagrams of variables with the different IGR value. For example we have selected the two variables (5 and 500) in two opposite cases; in case of high IGR the points are visually separated over against low IGR case, which is also carry some portion of useful information. 
Finally number of classes also plays some role in it, because reaching the maximum accuracy in the one class, we can lose performance in another. This can be shown on the discrepancy matrices for two different sets of attributes with the same precision: for example, in the first matrix, two first classes have less errors and three others classes more errors than in the seconf one.
Principal component analysis As we have deal with enough high dimensional data it would be really useful to applay Principal component analysis (PCA). The goal of PCA is to explain the most amount of variance with the lowest number of variables. Collinear variables can be combined into underlying factors or principal components that are uncorrelated with one another. Another words, we can reduce dimensions by dropping the components which do not explain much of the variance. So in order to see if PCA could be useful, we applay it in RF and SVM models.
We include all of the variables in the PCA.
And after performing PCA there are 102 variables have been rest.
Below the confusion matrix and statistics for RF and SVM models are shown:
RF with PCA data
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 473 | 5 | 18 | 0 | 0 | 0 |
| WALKING_UP | 40 | 415 | 16 | 0 | 0 | 0 |
| WALKING_DOWN | 72 | 43 | 305 | 0 | 0 | 0 |
| SITTING | 0 | 1 | 0 | 384 | 98 | 8 |
| STANDING | 1 | 0 | 2 | 40 | 488 | 1 |
| LAYING | 2 | 1 | 0 | 3 | 10 | 521 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.8775025 | 0.8526416 | 0.8651198 | 0.8891302 | 0.1822192 | 0.0000000 | NA |
SVM with PCA data
| WALKING | WALKING_UP | WALKING_DOWN | SITTING | STANDING | LAYING | |
|---|---|---|---|---|---|---|
| WALKING | 487 | 1 | 8 | 0 | 0 | 0 |
| WALKING_UP | 41 | 412 | 18 | 0 | 0 | 0 |
| WALKING_DOWN | 2 | 13 | 405 | 0 | 0 | 0 |
| SITTING | 0 | 1 | 2 | 431 | 53 | 4 |
| STANDING | 1 | 0 | 0 | 37 | 494 | 0 |
| LAYING | 0 | 0 | 0 | 0 | 0 | 537 |
| NA. | Accuracy | Kappa | AccuracyLower | AccuracyUpper | AccuracyNull | AccuracyPValue | McnemarPValue |
| overall | 0.9385816 | 0.9261940 | 0.9292992 | 0.9469808 | 0.1822192 | 0.0000000 | NA |
In both cases applying model on the reduced data gives us a worse model. Accuracy for RF model with PCA on 5% lower than for RF model with full features set and for 2% lower in case of SVM. The training time was significant reduced for both models (see table).
| duration | 11.4498 mins | 12.5184 mins | 8.850571 mins | 5.61223 mins | 16.16524 mins | 12.18458 mins | 2.044874 mins | 4.304657 mins |
| model | RF Full model | SVM Full model | SVM Imp Var model 490 | SVM Imp Var model 440 | SVM IGR model 526 | RF IGR model 561 | RF PCA model | SVM PCA model |
Above we demonstrated how successfully perform human activity classification based on a simple feature extraction process of a smartphone time series. Two classifiers (SVM and RF) with varying computational costs and similar profiles of performance were utilised. The lists of important variables by RF and IGR raise the question of whether to use the gyroscope to improve recognition accuracy. But it fact absence of gyroscope data among important variable can be explain. Among accelerometer features using X-axis acceleration showed the next lowest prediction error because X-axis was aligned with the direction of forward movement. And in context of simle activity recognition it is one of the most affecting features. But from physical point of view, due to gyroscopes only capturing dynamic rotational movement free from gravitational bias, using gyroscope features in Y and Z axes have to show lower errors than corresponding acceleration Y and Z axes. For achievement higher accuracy and reliably recognize using stairs and rotations, we need to applay more sophisticated system of activitty recognition, which requied combining accelerometer and gyroscope information for reduceing prediction errors.
Conclusion
- This test demonstrates that in case of simple human activity recognition SVM model shows a little bit better accuracy than RF model;
- The model training time in both cases is approximately the same;
- PCA could be usefull tools - in our test we used only 18% of variables with loss of accuracy of 2-5% and significantly reduced the time on 8-9 minutes.
- For increasing accuracy we have to recognition complex activities which are not as repetitive as simple activities and may involve various hand gestures. It seems to combining accelerometer and gyroscope information helps to reduceing prediction errors but required developing more complex algorithms.
4.1.4 Food recognition
Diet tracking is an important behavior change strategy for controlling obesity. However, manually recording foods eaten is tedious. Many diet tracking approaches have been proposed but have limitations. Recognizing food from photographs fails for complicated foods. Scanning food barcodes works but not all foods have barcodes. (to be continued)
4.2 Possible parameters
The specific parameters to be measured need to be determined during the research process– like genetic, metabolic or social factors implicated in the development of the disease. Obesity factors map These factors have been clustered as in the figure below. This figure is derived from Tackling Obesities: Future Choices – Obesity System Atlas published by the former Department of Innovation Universities and Skills and is used under the Open Government Licence.
Fig
Where from top left: Social psychology (yellow), Individual psychology (orange), Physical activity environment (dark brown), Individual physical activity (light brown), Physiology (blue), Food consumption (light green), Food production (dark green).
Potential parameters include:
- Biological characteristics:
- Age;
- Sex;
- BMI (Body Mass Index) - optimal threshold for defining obesity states;
- BAI (Body Adiposity Index) – is allowing differentiating muscular, skeletal and adipose masses;
- WHR (Waist to Hip Ratio) ;
- Weight gain until the age of 2 years;
- Health Current Parameters (temperature, heart rate, blood pressure, blood oxygen saturation, galvanic skin response, fasting blood glucose (FBG), cholesterol (LDL/HDL ratio, TC), quality of sleep, etc.);
- Hormones;
- Viruses;
- Genetic factors involved in the obesity genesis (endogenous individual factors);
- Metabolic factors implied in the development of the disease: nutrition, as well as predisposition to use glycolytic pathway more than oxidative phosphorylation in order to produce energy, like in the Warburg effect, etc.
- Behavioral characteristics:
- Activity: steps, distance - per day;
- Kind of sport activity per day;
- Energy factors: calories burned, calorie intake (proteins, glucides/carbohydrates, fats) - per day;
- Sedentary period of life – per day;
- Consumption of alcohol;
- Smoking
- etc.
- Social characteristics:
- Size of family;
- Circle of friends;
- etc.
- Environmental characteristics:
- Numbers of accessible green areas;
- Numbers of accessible supermarkets;
- etc.
- Psycho-social characteristics:
- Stigmatization in relation to mental health symptoms, body image;
- Low self-esteem;
- etc.
- Exogenous characteristics:
- Ggeographic location;
- Demographic dynamics;
- etc.
BMI is used as a gauge of obesity. Degree of obesity can be measured by BMI.
LDL/HDL ratio vs subject type (obese or non-obese)
LDL/HDL ratio is also extremely high for obese individuals compared to non-obese individuals.
The density plot of residuals on the figure above shows an approximately normal distribution, which is seems to be positive aspect of the model fit.
The plot of the fitted values against the residuals and the distribution residuals imply that there is no significant relation between the residuals and the fitted values and the residuals seem to have approximately Normal distribution around 0, both of which are crucial model assumptions.
This figure demonstrate the contributions of each phenotype to each ET. Green squares indicate a positive contribution while yellow squares indicate a negative contribution. Gray bars show the percent variance accounted for by each ET. Here, the first eigentrait captures more than 70% of the variance in the three phenotypes. This eigentrait describes the processes by which body weight, glucose levels, and insulin levels all vary together. The second eigentrait captures nearly 20% of the variance in the phenotypes. It captures the processes through which glucose and body weight vary in opposite directions. This eigentrait may be important in distinguishing the genetic discordance between obesity and diabetes.The third eigentrait describes the divergence of blood glucose and insulin levels and may represent a genetic link between glucose and body weight that is non-insulin dependent.
