Modelling the military wearable sensory
ecosystem as a diagnosable system
Master Thesis
Gijsbert de Boer
11
University of Amsterdam, Master Information Studies - track
Business Information Systems, student number: 10280464
August 2017
Abstract
The development of wearable sensors is rapidly increasing. Given that research has shown that in military operations many battlefield fatalities were deemed survivable, there is a strong justification for research into the applicability of wearable sensors in a military context. With the cur-rent pace of developments, this research should focus on the co-operation between sensors and comparing sensors in sensor ecosystems.
This research presents two methods of comparing sensors within sensor configurations. It was found that when comparing sensor configurations, information gain plays a central role and that sensor failure has to be taken into account.
It was also found that in current literature, very little attention is paid to reporting on data values and system models for sensors. This data is required when sensors are expected to co-operate in a sensor ecosystem. In future work, attention should be paid to reporting on data values. If this reporting improves, one can start working on automating the methods presented in this research and testing their practical use.
1
Introduction
Albert Einstein once said: ”I know not with what weapons World War III will be fought, but World War IV will be fought with sticks and stones”. His state-ment is exemplary for the extensive and rapidly increasing use of technology in military forces around the globe, both in harmful and non-harmful ways. This includes the research into and use of all kinds of sensors, ranging from CBRN-sensors1 [8] to wearable shooter localization and weapon classification [36]. Driven by the rapid growth of the consumer market for wearable sensors
[22], using wearable sensors to monitor the physiological status of soldiers for health and tactical purposes has gained renewed attention. Another important driver of this interest can be found in a study by [11], which reveals that between October 2001 and June 2011, 87.3% (n = 4012) of battlefield fatalities during the operations Iraqi Freedom and Enduring Freedom occurred in a pre-MTF2 setting. Of these deaths, 24.3% (n = 976) was deemed survivable. This pro-vides strong justification for research into the potential for wearable sensors in a pre-MTF setting.
[12] have created an extensive overview of all efforts in this field within the US Army and partners. One of their key points is that these sensors will be part of a wearable sensory ecosystem. This ecosystem will be part of a larger tech-nological ecosystem, which also includes items such as night vision goggles and radios. The same trend can also be recognized in the consumer market, where integration between devices and platforms has become key in winning customers over.
Unlike the consumer market, where wearable devices can be easily swapped and are usually by no means mission critical, a military wearable ecosystem poses extra requirements to the devices and technologies being used. It may be clear that a faulty or inaccurate sensor cannot be replaced once a mission has started and that the life of a soldier may depend on the functioning of the technolog-ical ecosystem. In several conversations with military personnel, the following concerns were raised with regards to wearable sensors during missions:
• Communication: data communication is either unavailable, very slow (via satellites), available but restricted (tactical data transmissions should be prioritised) or unwanted (during radio silence). A sensory ecosystem should therefore be able to handle the absence of a communication link while not tearing into its added value to the soldier. A good example of this would be having a sensor-based injury detection system. If a soldier is detected to be injured, the sensory system can provide the soldiers with medical advice based on the sensor data. On top of that, if a communica-tion link is available, it can send this data to a medic or commander, who can then take appropriate action.
• Energy consumption: there is a constant trade-off between the extra en-ergy consumption of an extra sensor versus the gain of having extra data. • Practicality: a sensor should not hinder a soldier in his/her duties, both physically (restricting movements) and practically (having to pay extra attention to the system while under pressure). There is a clear trade-off between having extra data and the extra resources that are required to gain the extra data.
• Rapid development of sensors: given the speed at which new and better wearable sensors are being developed, how should one keep up with these developments?
It is without a doubt that there are many more concerns and trade-offs that have to be taken into account. All these concerns and trade-offs lead to the conclusion that a military wearable sensory ecosystem with the purpose of monitoring the physiological status of soldiers is something that should be carefully planned and monitored.
1.1
The soldier and his/her sensors as a diagnosable
sys-tem
To take all the above mentioned issues into account, a careful way of planning the ecosystem before its deployment has to be constructed. The need for careful planning and continuous analysis and controlling of a sensor configuration was recognized back in 1988 by [2], who studied the subject from the context of visual perception:
”[...] it should be axiomatic that perception is not passive, but active. Perceptual activity is exploratory, probing, searching; percepts do not simply fall onto sensors as rain falls onto ground.”
The very same applies to the use of wearable sensors in a military context. A rapidly changing environment and context leads to continuously changing re-quirements for the technological ecosystem around a soldier. Besides contextual changes, other factors also influence the role and projected capabilities of a sen-sory ecosystem within the military, such as changing global relationships, urgent needs and political preferences. This was also found by [12], who stated:
”The specific objectives of Army physiological monitoring efforts have frequently shifted in reaction to specific urgent needs (e.g., Ranger hypothermia deaths), specific leader interests (e.g., live-dead detection and ballistic impact detection) and novel research oppor-tunities (e.g., ILIR and CEP funding). Transitioning PSM-related S&T products has been challenging due to continuously changing Soldier Modernization Program initiatives.”
In the scope of this thesis, we will assume the ecosystem only contains wearable sensors with the purpose of monitoring the physical health of a soldier.
To be able to analyse and plan such an ecosystem actively, we need to understand this ecosystem and its inner workings. In the most abstract sense, the soldier can be seen as a dynamic, changing system that can incur several conditions. All of these conditions are to some degree ”good” or ”bad”. Automatically de-tecting conditions in dynamic systems (or even specifically in humans) is by no means a new field of research (e.g. [5], [14], [4], [33]). Most literature, however, does not take a rapidly changing environment into account.
In the field of diagnosable systems, the work of [35] is one of the most prominent. It describes the inputs and processes within a diagnosable system by dividing the system into four components: three search spaces and the system model. These components will be reviewed below. In this thesis, we will try to find
out if this model from Stefik can be used to describe the soldier and his sensory ecosystem as a diagnosable system.
Once this ecosystem is understood, a method of comparing various configura-tions has to be found to be able to remain adaptive to various circumstances, as depicted in figure 1.
Figure 1: Multiple diagnosable systems in the scope of this thesis.
1.1.1 Diagnosable systems
A diagnosable system, as described by [35] has three main search spaces: the data space, the hypothesis-space and the repair-space (see figure 2). The data space describes the different types of knowledge we have about probe points. For each probe point, this knowledge can be divided into four types of data-values: possible values, normal values, predicted values and observed values. Possible values are values that are, simply stated, possible for a probe point. The range of possible values for the human heart rate, for example, is limited to numeric values larger than or equal to zero. The normal values, however, for an average human heart are somewhere between 50 and 150 beats per minute. The predicted values are what is expected based on the model and the inputs from other probe points. Lastly, the observed values are what is measured in the physical world.
Using the domain-knowledge that we have about the human body, we can con-struct the possible and normal values for most probe points. The same domain-knowledge might include a system model that allows us to simulate the func-tioning of the human body as a system and predict what values will be observed in the physical world.
The hypothesis space contains all possible hypotheses of what is wrong with the system when the observed values do not match the values that are predicted by the system model. The knowledge in the system model can be represented in the form of rules. These rules describe the condition of the system based on a number of inputs. If our observed values do not match our expected values, we
can use the observed values as input in these rules. These rules then give us a set of hypotheses describing the current condition of the system.
These models can become extremely complex. In a study by [33], a Dead or Alive Determination Model (DADM) was constructed to determine the DAU status of a solier. This model required a total of 4343 rules and 45 variables. The repair space contains the set of possible actions to repair the system if a unwanted hypothesis is confirmed. In the context of this thesis, these possible actions will likely be medical actions.
At the bottom of the figure, the system space is depicted. This space contains all elements that describe the normal functioning of the system, including some elements from its environment. In the context of this thesis, the system model will at least contain a set of physiological rules, which link the observed values to a set of hypotheses.
Given the military context in which the model will be applied, the normal data values within the data space are highly dynamic. Logically, the normal data val-ues for core body temperature are different for a soldier on a training mission in Norway than for a soldier being deployed to Mali. The planning contain-ing the location and tasks a soldier has to conduct thus directly influences the prediction of normal values within the data space. Without contextual infor-mation following from a logistical system, it becomes significantly more difficult to monitor soldiers in their daily duties.
Because the observed values are a separate category of values within the model, it can be derived that we need to take the possibility of faults in the observations into account. Because of this, the sensors/observers automatically become com-ponents of the model themselves. These comcom-ponents can have several states, of which ”normally functioning” is often the most likely. However, as components can break down (including sensors) or function unexpectedly, the likelihood of other states needs to be taken into account as well. Given the circumstances in which a sensory ecosystem will be deployed, the hypothesis that an observation (regardless of whether it is actually true or not) is caused by a faulty sensor always needs to be taken into account.
2
Research questions and -subquestions
The potential capabilities of wearable sensors grow everyday and are likely to keep doing so for the foreseeable future. This inevitably leads to a future where we as humans can no longer grasp the amount of information we can measure and derive from a limited set of sensors. In such a situation, a method enabling the comparison of several sensor configurations is required.
In the previous sections, we have created the context from which we will look at sensors in this thesis. First, we will try to show that the model of diagnosable systems from [35] can also be applied to a soldier and his/her sensory ecosystem. We will do so by showing that all inputs (the four types of values) required by the model can indeed by found in existing literature or are at least known to experts. Besides, we argue that the causal and behavioural models depicted in
Figure 2: Search spaces for diagnosable systems. Image by [35].
the system space are widely known in medical practice, and that therefore a link between the data space, hypothesis space and therapy space is possible. Next, we will compare two methods of finding the most efficient sensor configu-ration, based on the information gain achieved by both. We will also propose a method of including the cost of an observation into the consideration of whether a sensor should be included or not.
This has led to the following research questions:
1. Can a military wearable sensory ecosystem, aimed at monitoring the phys-iological status of a soldier, be modelled and compared as a diagnosable system, as described by [35]?
(a) What potential medical diagnostic capabilities can be discerned from available literature with regards to a system as described above? (b) Can these capabilities be described as diagnosable conditions?
(c) Which sensors are required for each of these capabilities? Can these sensors be modelled in the same way one can model the component of a diagnosable system?
(d) Can a methodology be developed that enables us to compare wear-able sensor configurations based on the amount of information these
configurations generate?
If it turns out the system can indeed be modelled as a diagnosable system, the advantages to the military are multiple. First, it provides a clear oversight of the relations between inputs (the sensors) and the outputs (the diagnosable conditions; the ”data”). Once the desired outputs have been determined, the required inputs can be derived. Research into wearable physiological status monitors can then be steered into the direction of the required inputs. Lastly, it helps the military make use of the capabilities offered by the system as effective as possible.
2.1
Scope of this thesis
It should be noted that the actual modelling of a sensor configuration is not within the scope of this thesis. This is due to a number of reasons:
• The desired result, a methodology that can be used to compare sensor configurations, depends on a number of inputs (medical knowledge, costs of consequences, available sensors). Finding and describing each of these inputs would require a massive amount of work and is therefore out of the scope of this thesis.
• Testing the effectiveness of this method requires that it is fully automated, so that various configurations and scenario’s (inputs) can be compared. Completely automating this method, however, is out of the scope of this thesis.
Because of these reasons, the scope of this thesis will be limited to the theoretical development of the methodology.
3
Methodology
3.1
Identifying potential capabilities
To identify all potential capabilities for a military wearable sensory ecosystem with the purpose of monitoring physiological soldier status, a literature review will be conducted. The main purpose of this literature review is to identify as many capabilities as possible, therefore no strict limitations will be used to what will be included and excluded. We will search for literature using Google Scholar. The results will be checked with military personnel for completeness and accuracy.
Once these capabilities have been found, the first step is to fill the data space. Therefore we need to find all probe points that are required for the capabilities. If no adequate support for the discovered measuring points can be found, the capability will be omitted. For example, if one wants to detect whether a soldier is in need of trauma care, one can follow the following algorithm:
1. Get the observed values by sensors for blood pressure, respiration rate and Glasgow Coma Scale-score.
2. Generate hypotheses based on observed values and known normal data-values such as described in figure 4.
3. Extend hypotheses-set with possibility that one or more of the sensors are faulty. Determine the likeliness of this new hypothesis based on the system model and contextual information.
4. Calculate the most trustworthy hypothesis and take appropriate action. From this, we can derive that in order to detect whether a soldier is in need of trauma care, we need to be able to measure blood pressure, respiration rate and the Glasgow Coma Scale-score.
Normally, the next step would be to determine the possible values and the normal values. In this case, the possible values for blood pressure are found on a numeric scale ranging from upwards from zero and are measured in mmHg. The possible data values for respiration rate are any positive number of zero. The possible data values for Glasgow Coma Scale are any numeric value between 3 and 15. The normal values, as defined in the ”Seventh report of the Joint National Committee on Prevention, Detection, Evaluation and Treatment of High Blood Pressure” can be found in figure 3. This figure also provides us with a first insight into what the rules of the model will look like with regards to blood pressure: which values are out of range and what severity is associated with these values? A more detailed look into which values are out of range when combined with known values for respiration rate and Glasgow Coma Scale-score, can be found in figure 4.
Once the probe points for all identified capabilities have been found, the next step will be calculating the most efficient set of probe points.
Figure 4: The Revised Trauma Score Fast Reference Chart
3.2
Finding literature
The following phrases were used to search Google Scholar for suitable literature: 1. ”wearable sensor applications”
2. ”wearable biosensor systems”
For each of the searches, the option to include patents and citations was disabled. Although this does not prevent that any unwanted results are still in the result set (such as presentations), it reduces the number of unwanted items. For each of the searches, the first two pages of results (10 results per page, 20 results per search in total) were analyzed. This created a result set of 60 publications in total.
Next each of the results was analyzed. The analysis aimed at answering the following question: does the result present a method of extracting one or more medical feature(s) using wearable sensors? If so, the result was included in a subset of the results. If not, the result was omitted. It has to be noted that in the analysis, ”extracting a medical feature” was meant in the broadest sense. For each of the results in the above mentioned subset, the presented use case was described, including the required sensors. The results of this analysis can be found in table 1. The abbreviations for each of the probe points or sensor types can be found in table 2.
3.3
Finding data-values for probe points
Before a soldier can be modelled as a diagnosable system, the capabilities need to be refined into diagnosable conditions. Every capability can extract one or more medical features. For each of the diagnosable conditions, the probe points were identified in the previous step. The next step would be to discern the data-values for each of the probe points. As described earlier, there are four types of data values in the data space, of which three should be known before hand. These are the possible values, normal values and predicted values. The normal values of these probe points can vary: the normal values for heart rate in rest differ from the normal values when exercising. The normal values can also depend on the observed values of other sensors. For example: the normal temperature values for a subject during exercise are much higher than in rest. However, if the observed heart rate and activity are low, the normal values for temperature are much lower, since the subject is likely in rest. Besides varying in range, the type of value can be different. Heart rate, for example, can only be positive and numeric. There are however various other probe points of which the values are not so easily classified. Regarding the use of inputs that are not so easy to classify, clear agreements have to be made. Examples are measuring the probe point ”activity”, for which no clear scale exists. Determining a scale by which ”activity” can be classified is a good example of this.
Once the normal values and predicted values are known, we can compare the observed values with the predicted values and generate our hypotheses based on the difference between them. To be able to do so, a further analysis was made of the results found in table 1. This analysis focused on the extent of reporting on the possible and normal data values, as well as the existence of algorithms or classification systems with which the predicted values can be determined. For each result in table 1, the publication was read again, this time with the aim of noting any mention of data values in any wording. If anything resembling a data value was found, it was noted and reported. The results of this analysis
can be found in table 3.
3.4
Analyzing the operational context
In order to improve our understanding of the operational context in which wear-able sensors are to be deployed, several conversations with military personnel were had. As this context is subject to urgent military needs, specific leader interests and research opportunities [12], these conversations merely served as a beginner’s guide into military operations.
3.5
Finding a method to compare sensor configurations
During the course of this research, various aspects of comparing sensor con-figurations were discussed. There are multiple ways in which sensors can be compared to one another. First of all, one could look at the general properties of a sensor:
• Purchase cost • Operating cost • Weight
• Diagnostic capabilities • Overall added physical strain • Power/battery usage
Of course, there are many more properties for each sensor. It’s safe to say that for each of these properties, the bandwidth of their accepted values depends on their diagnostic capability. If a sensor is capable of detecting more or more serious medical conditions, it is allowed to cost more. Based on discussions with military personnel and Alexander Boer, it was soon decided that, given that the accepted values of properties of all sensors depend on the diagnostic capabilities of the sensor, we should exclude these properties from the comparison and solely focus on comparing diagnostic capabilities of each sensor. We can then define some ”exchange-rate” between added diagnostic capabilities and the ”cost” of a sensor (with cost being used in the broadest sense - it does not necessarily have to be monetary).
However, the diagnostic capability of a single sensor is rarely independent from other sensors. After all, many medical conditions can only be recognized when multiple medical features are combined (see section 3.1 for an example). Having more sensors means more data can be used to confirm or deny more hypotheses. This makes our problem of selecting the best sensor significantly more difficult, as the added value of a sensor now depends on the ecosystem it will be imple-mented in.
Besides being able to confirm or deny more hypotheses, not every hypothesis is worth equally much. Being able to confirm or deny whether a soldier has a
fever might be of less added value then being able to detect if a soldier has a cardiac arrest as a result of a bleeding. There is a ranking in the valuation of hypotheses. We do not know this ranking in advance.
Because of the issues described above, it was decided that our method(s) should be able to calculate the most efficient sensor configuration with a limited, pre-defined set of hypotheses as inputs.
4
Results
4.1
Potential capabilities
As can be seen in table 1, 22 different use cases for sensor or capabilities were identified. The full names for the sensor abbreviations in this table are found in table 2. These capabilities can roughly be divided into three categories:
• Detecting acute conditions
• Detecting health-related patterns (sleep, nutrition, movement, etc) • Detecting either of the above using a custom developed tool or sensor The identified capabilities vary in the extent to which they are generic. Plenty of studies (eg. [27], [34] and [24]) focus on general health monitoring, while others (eg. [6], [13] and [32]) focus on the detection of very specific medical conditions.
Some studies have a common purpose ([27], [1], [30], [9]) but use different ob-servables to reach their goal. This might be of interest, as it suggests there is more than one way of detecting a medical feature.
As can be seen in table 2, 16 distinct medical probe points or sensor types were found, while none of the publications found used more than 6 different types of sensors simultaneously. Of all sensors identified, the ECG-sensor was most commonly used.
Use case Probe point/sensor type Source Medical monitoring in extreme
environments (space & terrestrial) ECG, BP, R, T, SpO2, A [28], [24] Recording of cardiorespiratory
and motion signals during
spontaneous behaviour in daily life
ECG, R, T [28], [10]
General health monitoring ECG, R, T, A [28], [27]
Real-time physiological status
monitoring with wearable sensors ECG, SpO2, A [28], [34]
Detection and prediction of human physiological state (wakefulness, fatigue, stress) during daily activities
ECG, BP, R [28], [23]
Evaluation of the emotional state of an individual in environments where subjects operate at extreme stress conditions
ECG, R, GSR, EMG [28], [18]
Monitoring sleeping conditions ECG, R [6]
Monitoring movement of body parts GYR, ACC [21]
Wearable ubiqutious health monitoring ECG, ACC, SpO2 [7]
Detect if subject is eating a meal ACC, HR, T, GSR [26]
Wireless wearable endoscopy CMOS VGA [13]
Detecting endogenous depression GSR, ACC [32]
Predicting obstructive sleep apnea PPG [19]
Cuffless blood pressure measurements PPG, ECG [16]
Determine physiological strain index T, HR [15]
General health monitoring and detection
of abnormal patters during activity ECG, R, A, T [30]
Measuring heart rate using earphones HR [31]
Unobtrusively measure blood sugar levels GS, pH, T [17]
Predicting obstructive sleep apnea OS [25]
General health monitoring using a
wearable smart shirt ECG, HR, AoI, A, T [9]
General health monitoring using a
wearable smart shirt ECG, BrP, A, T [29]
General health monitoring using a
wearable smart bracelet HR, SpO2, T, A [1]
Abbreviation Probe point/sensor type
ECG Electrocardiography
BP Blood pressure
R Respiration
T Temperature
SpO2 Oxygen saturation
A Activity
GSR Galvanic Skin Response/Electrodermal activity
EMG Electromygraphy
GYR Gyroscope
ACC Accelerometer
CMOS VGA Small camera
pH pH-value
GS Glucose Sensor
OS Oximetry sensor
AoI Angle of Inclination
BrP Breathing Pattern
Table 2: Distinct probe points/sensor types
4.2
Reporting on data-values
Unfortunately, very little publications report on the data values that are used to detect conditions. This is strange, as the rule-set that is used to automatically detect conditions forms an equally important part in a new system as the sensor itself. Of all publications, only one reported the normal data values and an algorithm to derive higher medical features ([1]). Although these values are likely to be known in the medical world, requesting them all is outside the scope of this thesis.
Another element of sensors that is generally not reported on is the types of failures these sensors can encounter. In general, three types of failures can be discern with regards to the data a sensor reports:
• Sensor failure leading to the sensor reporting error messages instead of values,
• Sensor failure leading to the sensor reporting false data (such as a wrongly measured heart rate, caused by sweat or dust) and
• Sensor failure leading to the sensor incorrectly reporting that no observa-tions can be made (such as a sensor falsely reporting that no heart rhythm is present).
Literature does generally not report on the probability of these types of failures happening. This means that, at this moment, we cannot verify whether the values that were reported could be used to model a soldier as a diagnosable system, because we cannot fill the data space.
Use case Data values Medical monitoring in extreme
environments (space & terrestrial) No data values reported Recording of cardio-respiratory
and motion signals during spontaneous behaviour in daily life and in a clinical environment
No data values reported
General health monitoring Reports only on (technically) possible data values Real-time physiological status
monitoring with wearable sensors Reports only on (technically) possible data values Detection and prediction of human
physiological state (wakefulness, fatigue, stress) during daily activities
Reports only on (technically) possible data values Evaluation of the emotional state
of an individual in environments where subjects operate at extreme stress conditions
No data values reported
Monitoring sleeping conditions Reports only on (technically) possible data values Monitoring movement of body
parts Reports only on (technically) possible data values
Wearable ubiqutious health
monitoring Reports only on (technically) possible data values
Detect if subject is eating a meal No data values reported
Wireless wearable endoscopy Reports only on (technically) possible data values. Detecting endogenous depression Reports on (technically) possible data values and
refers to publications presenting normal data values. Predicting obstructive sleep apnea Reports on (technically) possible data values. Cuffless blood pressure
measurements No data values reported
Determine physiological strain
index No data values reported
General health monitoring and detection of abnormal patters during activity
Reports on (technically) possible data values. Measuring heart rate using
earphones Reports only on (technically) possible data values
Unobtrusively measure blood
sugar levels Reports only on (technically) possible data values
Predicting obstructive sleep apnea
Reports on (technically) possible data values and describes an algorythm to detect OSA based on derivation from a baseline. No distinct values are given for this baseline.
General health monitoring using a
wearable smart shirt Reports on (technically) possible data values. General health monitoring using a
wearable smart shirt No data values reported
General health monitoring using a wearable smart bracelet
Reports on normal data values and describes an algorithm to derive higher medical features.
Table 3: Reporting on data-values for each of the use cases described in table 1.
4.3
Comparing sensor configurations
In this section, two methods of comparing sensor configurations are presented. Both methods use conditional entropy to determine the most efficient sensor set. Their main difference is based on prior probabilities. The first method aims at fitting a solution (a hypothesis) that describes the observed values best, whereby a solution is seen as either fitting or unfitting: the solution either matches with the observation or it does not match. The second method employs prior probability to describe the extent to which a set of solutions matches an observed value.
4.3.1 Comparison method 1
The first method of comparing sensor configurations is based on the principle that causality involves necessity [3]. It employs the thought that problem solv-ing, in an abstract sense, can be seen as picking the correct solution from a set of potential solutions [S1...Sn] with a given probability mass function P (S). Let us have the following potential hypotheses (or ”solutions”) for the current state of a ”diagnosable system”, in the context of this thesis:
Hypothesis Description Prior
s1 healthy, no broken sensors 0.9
s2 dead, no broken sensors 0.0001
s3 bleeding, no broken sensors 0.03
s4 healthy & blood pressure sensor failure 0.015 s5 dead & blood pressure sensor failure 0.00005 s6 bleeding & blood pressure sensor failure 0.01479 s7 healthy & respiration sensor failure 0.01 s8 dead & respiration sensor failure 0.00005 s9 bleeding & respiration sensor failure 0.01 s10 healthy & both sensors broken 0.01
s11 dead & both sensors broken 0.00001
s12 bleeding & both sensors broken 0.01
Table 4: Hypothetical solutions, and their prior probability
Note that the priors in table 4 are composed of the prior probability of a soldier being healthy, dead or bleeding, multiplied with the prior probability of none, one or both of the sensors being broken. These composed probabilities are then normalised, given that these components can fail independently. Given these hypotheses, the next step is to decide which observation to make, in order to reduce the uncertainty about the correct hypothesis as much as possible. Suppose we have a set of potential observables [O1...On], each with a potential set of values [O1
1...O1n]. Combined with some basic knowledge about the human body, we can derive which solution(s) fit a given observed value best. Let O1 be blood pressure:
Observation Description o1
1 blood pressure normal
o1
2 blood pressure elevated
o13 blood pressure lowered o14 no blood pressure
Table 5: Blood pressure observable Let O2be respiration: Observation Description o2 1 respiration normal o2 2 respiration elevated o2 3 respiration lowered o2 4 no respiration
Table 6: Respiration observable
Based on the basic knowledge we have of the human body, we can draft a small set of rules connecting the observations to the right solutions. Knowing which symptoms match which solutions allows us to create a table connecting them. We thereby assume that a solution either explains an observation or not: weaker statistical associations are of course possible in real life, but managing the knowledge required for these associations would be enormously complex. In this example, we will adhere to a simple set of rules:
1. If a soldier is healthy, his blood pressure and respiration rate are normal. 2. If a soldier is bleeding, his blood pressure is lowered and his respiration
rate is increased.
3. If a soldier is dead, he has no blood pressure nor respiration.
4. The two employed sensors can only break down in such a way that they can only stop transmitting data: they cannot break down in such a way that they start to generate false data.
Based on the rules described above, we can now make a joint probability table that links the observations to a set of solutions. As the probability P (O) is the sum of the probabilities of the solutions that explain observed value O. The next step would be to calculate the change in entropy from H(S) to H(S)−H(S|O1), as well as the change in entropy from H(S) to H(S) − H(S|O2), and select the observation that gives us the highest information gain.
Solution O1 1 O21 O13 O41 O21 O22 O23 O42 S1 1 0 0 0 1 0 0 0 S2 0 0 0 1 0 0 0 1 S3 0 0 1 0 0 1 0 0 S4 0 0 0 1 1 0 0 0 S5 0 0 0 1 0 0 0 1 S6 0 0 0 1 0 1 0 0 S7 1 0 0 0 0 0 0 1 S8 0 0 0 1 0 0 0 1 S9 0 0 1 0 0 0 0 1 S10 0 0 0 1 0 0 0 1 S11 0 0 0 1 0 0 0 1 S12 0 0 0 1 0 0 0 1
Table 7: Joint probability table for observables O1 and O2
4.3.2 Comparison method 2
The second comparison model expects a symptom to be given. Based on this symptom, a set of hypotheses with an associated probability distribution can be drafted. For this set X, the entropy can be calculated. Next, a hypothetical sensor can be added, yielding an observation Y . This observation, given that it is not completely independent of X, will give us more information about the condition of the soldier. This process will be explained in more detail below. For each of the potential observed values of Y , we will determine the effects of that observation on X. Given each of these observations, a new probability dis-tribution for X is determined. This leads to a change in entropy. The average of the newly determined entropies will be the used to calculate the information gain when that specific sensor is added.
As stated above, if we define the current condition of the soldier as the inde-pendent variable X with probability mass function P (X), we can calculate how informative this distribution is using the formula for Shannon’s Entropy:
H(X) = − n X
i=1
P (xi)logb(P (xi)) (1)
A high entropy implies a high uncertainty about the current condition of the soldier and vica versa: a low entropy implies a low uncertainty about the cur-rent condition of the soldier. We are therefore looking for the lowest possible entropy.
Given that the set of potential hypotheses is described by X and Y is an observa-tion that we make using a sensor, there is a condiobserva-tional probability distribuobserva-tion for X and Y . The formula for the conditional entropy of this distribution is as follows: H(X|Y ) = n X i=1 P (xi)H(Y |X = xi) = H(X, Y ) − H(X) (2)
Unless Y is completely independent from X, learning the value of Y leads to a new probability distribution X with reduced entropy. Given that all circum-stances are equal, the best observation Y is the one that reduces the entropy of X most.
The most important difference between this method and the first one is that this method employs the notion that knowing the value of a first observable changes the probability mass function associated with a potential second ob-servable. Although this is closer to real life, it requires knowing beforehand how each observed value affects the probability mass function of another observable, which makes it (as part of process) very hard to manage.
4.3.3 Advantages and disadvantages of both comparison methods Both methods of comparing sensor configurations have their pros and cons. There is no clear ”best” method; the method that yields the most value is the method that fits the organization that implements it most. However, there are some significant differences between these methods that will be discussed here. First of all, the first method only allows for a hypothesis that it is explained or not explained by a given observed value. Weaker statistical relationships are of course possible in real life and not uncommon in medical science. The second method does allow these weaker statistical relationships as they only need to be applied in a situation where two sensors are compared, in order to find the most efficient next sensor.
Although it is debatable to what extent these relationships can be defined in the first place (the medical processes behind some relationships might not even be fully understood), allowing these relationships does present us with a data-management problem. Having statistical relationships between observed values and hypotheses instead of relationships that simply explain a symptom are exponentially more difficult to manage. Besides, a different issue arises: many of these statistical relationships will not even be researched yet and therefore have to be estimated at best.
Secondly, the first method starts with a set of medical conditions one would want to be able to diagnose. Repetitively, a new sensor is added until the cost of the new sensor does not weigh up to the added value. The second method assumes a symptom is given and starts analyzing from there. Repetitively, two or more new sensors are added and compared in terms of added value. The most cost efficient sensor set is then selected and added to the sensor set, until the added value of the new sensor does not weigh up to its cost. Interestingly, this leads to a situation in which the first method also needs to be able to detect that a soldier is healthy, while the second method only needs to be able to correctly diagnose the given symptom (given that ”being healthy” is not seen as a symptom).
4.3.4 Including cost and value of information
As stated above, the best observation to make is one that reduces uncertainty about the condition of the soldier most. However, doing an observation has a certain cost. If the expected utility of the observation exceeds the cost of the observation, the observation logically should be made. This requires us to be able to calculate two things: the expected utility of an observation and the cost of an observation. Next, an exchange rate has to be determined: how much is one measure of expected utility worth in terms of cost?
The cost of an observation is very hard to determine, since there are so many factors influencing it. A more practical solution to this problem can be found by coupling the expected utility to the cost reductions doing a certain observation can achieve. This approach requires a number of assumptions:
• The soldier is known to be suffering from a condition that requires medical attention. The set of hypotheses about this conditions is described by S. • A division between hypotheses in set S can be made based on the remedial
action that has to be taken, if that hypothesis were true. • Each of these actions has an associated cost.
• A wrong diagnosis ultimately leads to a higher cost, resulting from death or extra remedial actions.
With these assumptions in mind, a sample calculation can be made.
Let us assume that a soldier A on a patrol far away is suffering from an unknown condition that seems to require medical attention. Based on prior data, we know that our set S contains the following hypotheses:
S = {heat stroke, myocardial infarction} (3) The symptoms of these conditions are somewhat similar. Prior data shows that the probability mass function for this specific set is P (X) = 0.5. A heat stroke, however, is relatively easy to treat and does not require the immediate evacu-ation of the soldier. A myocardial infarction, on the other hand, does require immediate attention from medical specialists with specialized equipment. We can therefore split our set S into two subsets, based on their need for immediate evacuation:
S1= {heat stroke} (4)
S2= {myocardial infarction} (5)
If we diagnose the soldier to be suffering from a condition in set S1, the soldier should be treated by a medic on the ground. If we diagnose the soldier to be suffering from a condition in set S2, a medical evacuation helicopter should be sent to the soldier. Each of these remedial actions has an associated cost. However, additional costs are to follow if the diagnosis turned out to be wrong. This leads to the following scenarios, with the costs being defined in a fictional currency:
Scenario Cost Helicopter is sent
soldier has heat stroke 70.000 Helicopter is sent
soldier has myocardial infarction 80.000 Helicopter not sent
soldier has heat stroke 10.000 Helicopter not sent
soldier has myocardial infarction 100.000 Table 8: Costs of sample scenarios
Based on the probability mass function for this set of hypotheses, the prob-ability for a hypothesis to be true is 50%. This implies that in half of the cases, a helicopter will be sent, while in the other half of the cases, no helicopter will be sent. In both cases, the diagnosis and subsequent action could be wrong. We can calculate the average cost of this specific situation:
C = n P i=1 P (i) ∗ CCD + (1 − P (i) ∗ CWD) i (6)
With C being the cost of a given situation with a fixed amount of information, n being the number of subsets derived from the number of unique remedial actions, P (i) being the probability function subset i in S, CCD being the cost of a correct diagnosis and CWD being the cost of a wrong diagnosis. In our case:
C =(.5 ∗ 10.000) + (.5 ∗ 100.000) + (.5 ∗ 70.000) + (.5 ∗ 80.000)
2 = 65.000 (7)
Logically, the next step to take is to find a sensor that reduces the uncertainty about the condition of the soldier most. Let us assume that by observing an ECG of the soldier, the probability mass function for the condition of the soldier changes so that P (send helicopter) = .8 and P (do not send helicopter) = .2. This changes the average cost of the situation:
C =(.2 ∗ 10.000) + (.2 ∗ 100.000) + (.2 ∗ 70.000) + (.8 ∗ 80.000)
2 = 51.500 (8)
If an exchange rate between cost reduction and observation cost can be deter-mined, one could now calculate which observations are economically viable. However, there is a catch: the economic viability of doing an observation is largest when the difference between the cost of a wrong diagnosis and the cost of a correct diagnosis is largest (given that the cost of a wrong diagnosis is higher than the cost of a correct diagnosis). If the cost of a soldier dying from a wrong diagnosis is almost equal to the cost of a soldier being treated after a correct diagnosis, wrongly diagnosing the soldier is hardly felt in terms of
economic consequence. This does mean that when a soldier has either one of two very deadly conditions, the economic viability of an observation would be rather small because wrongly diagnosing the soldier does not cost a lot more.
4.4
Analyzing multiple methods of comparing sensor
con-figurations
4.4.1 Diagnostic quality and issues from a medical perspective When inspecting both methods from a medical point of view, a number of issues arise with both methods. The first method explicitly assumes that an observa-tion either explains a condiobserva-tion or not. This assumpobserva-tion can be challenged for at least two reasons:
• The observation could be explanatory only when observed in conjunction with another observation (or the absence of it). Few medical signs are pathognomonic, although sensors extend the amount of pathognomonic observations quite a bit. This directly relates to the amount of domain-knowledge that we have about the human body and the conditions it can incur: when reasoning from a specific condition, we might be able to require certain observations, of which not all are available.
• The observation is not a 100% medically specific or sensitive test for the hypothesis to be tested.
The second method makes does not suffer from the second objection, as it allows weaker statistical relationships to be used.
Both methods also assume that all potential sensors measure observables of which the normal and expected values can be classified in such a way that they can be combined with information from other sensors. However, many medical tests, such as the wireless wearable endoscopy by [13], produce data that is very difficult to automatically classify. Besides classification issues, many medical diagnostic criteria rely on medical features that can not be abstracted into something measurable by a sensor. For example, one of the diagnostic criteria for sepsis is ”altered mental status” [20]. With the current state of research, such features are not yet detectable.
4.4.2 Diagnostic quality and issues from a diagnosable systems per-spective
Knowing the probability of certain observations to be made is crucial when calculating entropy changes as a result of adding new sensors to the sensory ecosystem. Although both methods use prior probabilities to estimate these probabilities, differences exist between methods in the way these probabilities are used. The first method employs the normalized prior probabilities of all possible observations combined with the probability of one or more sensors be-ing broken. The second method does not use these combined normalized prior probabilities: it sees an actual observation and having a broken sensor as two
different observations. This difference sheds light on a data management issue with method one: the combined probabilities are often difficult to estimate, es-pecially when employing fairly new sensors of which failure rates are unknown. In practice, these failure rates should be provided by literature or manufactur-ers of sensors, however, this is rarely done. When not provided, estimating the probability a sensor is failing in advance is a guess at best, let alone estimating the likelyhood of multiple sensors breaking down simultaneously. The second method suffers less from this problem, as it only uses prior probabilities to esti-mate the probability of observations to be made by the new sensor. It does not make use of combined or normalized probabilities. Using combined probabili-ties, like method one does, also increases the number of hypothetical scenario’s exponentially, given that each potential combination of sensor failures also is seen as a separate scenario.
The first method also suffers from the fact that it employs normalized proba-bilities. When an observable, along with its observations, is added to the set of hypothetical solutions, the normalized probability of all solutions has to be recalculated. With more sensors being included in the potential set of solu-tions, the work involved in recalculating all probabilities grows exponentially with each new sensor added. The second method also suffers from this problem, albeit slightly less. It only includes the potential solutions for a given symptom to be diagnosed, thus effectively enables us to deal with a subset of all normal-ized probabilities. Having this split per symptom enables us to work with prior probabilities on a symptom-basis, which are easier to recalculate when needed. 4.4.3 To what extent adhere both methods to the practical
con-straints as described earlier?
The practical constraints as described by the Dutch Army are as follows: • Sensor should be able to work independently from a
data/communication-link.
• Energy consumption should be as limited as possible.
• Sensor should not hinder a soldier in its duties, both physically and men-tally.
• During the diagnostics process, the system should include the current operational context into its diagnosis.
In neither of the described comparison methods, a data/communication link is strictly required. However, depending on the actual implementation of the con-figuration, such a data/communication link might become a requirement. This depends on whether the system should be capable of detecting conditions au-tonomously or based on a condition, when triggered by another soldier. When automatically detecting conditions, in order to be able to raise an alarm, having a data/communication link is required. When diagnosing on site, triggered by
(for example) a medic, this link is not strictly required. Besides implementa-tion, new sensors might require a data/communication link. The same applies to energy consumption and hindering a soldier during its duties: because the sensors are not known in advance, it is impossible to judge the extent to which these constraints have been met.
4.4.4 What inputs are required for both methods? The first method requires the following inputs:
1. A predefined set of hypotheses one would like to be able to test for, 2. A set of potential sensors to choose from,
3. Data values for each sensors: possible values and normal values, 4. Prior probabilities for each observation,
5. Prior probabilities with regards to sensor breakdowns,
6. A set of rules which can be used to derive the predicted values for each sensor.
Please note that in this list of required inputs contextual information is not directly included, such as outside temperatures and type of activity of the mea-sured person. This contextual information should be included by adjusting prior probabilities and data values. For example: when a soldier is deployed to the Arctic region, prior probability of heat stroke is less then when the same soldier is deployed to the Sahara. One could also include this information by slightly adjusting the rules used to derive the predicted values for a sensor.
The second method requires the following inputs:
1. A set of symptoms one would like the system to be able to diagnose, 2. A set of hypothesis per symptom, that explain said symptom, 3. A set of potential sensors to choose from,
4. Data values for each potential sensor: possible values and normal values, 5. Prior probabilities for each observation,
6. Prior probabilities with regard to sensor breakdowns,
7. A set of rules which can be used to derive the predicted values for each sensor.
Please note that again, in this method, contextual information should be in-cluded by adjusting prior probabilities, normal data values and/or the set of rules being used to predict data values.
4.4.5 Finding the best method
When finding a ”best” method, one would have to keep in mind that the original starting point of this research was to find out how rapid triage could be applied to a soldier using various sensors. Given this starting point, the ”best” method should be the one that enables this rapid triage best. However, enabling rapid triage using sensors is not very straightforward: one can of course automate the process of manual triage, but that would yield little added value over doing it manually, given the weight, physical constraints and power usage of sensors. Being able to triage faster ´and have more accurate results therefore should be the preferred result when employing sensors for this purpose. Besides the main goal of rapid triage, the system should be able to adapt to the operational cir-cumstances, such as the availability of new sensors or different types of missions. According to the above, the criteria for selecting the best method thus are as follows:
• Reliable selection of the correct hypothesis
• Little dependency on prior probabilities, as these might be very difficult to determine or be unknown
• Easily adjustable to operational circumstances
In the next paragraphs, we will analyze the extent to which both methods meet these criteria.
Method one The first method assumes a probability mass function is known for each potential hypothesis explaining the current condition of the soldier. However, as the set of hypotheses has to be determined in advance, the pos-sible conditions range from completely healthy to dead. In medical practice, one would start by analyzing the symptoms of a soldier before drafting a set of potential diagnoses for the condition of the soldier. When the set of hypothe-ses (and the number of observations explaining them) is relatively small, an hypothesis can be confirmed or denied rather accurately. However, as more sen-sors are added, this reliability rapidly declines, given the exponentially growing amount of hypotheses. This method therefore suffices when performing simple acts of triage, but could be suffering when aiming at diagnosing more difficult conditions.
When looking at dependency on prior probabilities, this method suffers when more sensors are added. It makes use of combined hypotheses, including the combined probabilities of one or more sensors failing. As stated earlier, deter-mining the probability of a sensor failing is in most cases a guess at best, let alone when determining the probability of multiple failures happening at the same time.
The extent to which both methods can be adjusted to operational circumstances are difficult to estimate and likely influenced by more factors.
Method two The second method assumes a probability mass function is known for each symptom (or observation). This makes it easier to select the correct hypothesis, as the probabilities are based on a given symptom and not purely on the possibility of a symptom occurring in the first place, such as is the case in method one. It also allows weaker statistical relationships to be used to explain various symptoms, whilst method one only allows that a symptom either explains an hypothesis or not.
This method relies even more heavily on prior probabilities to be known, be-cause symptoms are not just explained or not by an hypothesis such as in the first method. A certain gradation is possible in this case. Besides, where for the first method only the prior probability for a predetermined, fixed number of observations had to be determined, for this method, the prior probabilities of all potential combinations of symptoms and observations have to be known. This makes this method both very dependent on knowing prior probabilities and not very easily adjustable to operational circumstances.
5
Conclusion
In this thesis, we have argued that it is likely that a soldier and the technological ecosystem around him/her can be modelled as a diagnosable system according to the model of [35]. It was found that although plenty of research is being done in the field of wearable sensors, little attention is being paid to the interoperability between sensors and the data they provide. This field of research would hugely benefit from having a standardized set of normal data values and medical rules, which can then be used to create different sensor configurations. It was also found that at this moment, too little is reported in literature about data values and system models to be able to do an analysis of a sensor ecosystem at this time.
We have also presented and discussed two different methods by which sensor configurations can be analyzed and compared and by which the value of a sensor in a sensor configuration can be determined. It was argued that information plays in crucial role in the analysis and comparison of sensor configurations. We have also argued that sensor configurations can be compared by the information gain they provide, by calculating the difference in entropy. For this, we provided two different methods. Both methods have their pros and cons and can be used under different circumstances. In the presented methods of analysis, the possibility of sensor failure was included.
Besides the information gain, a method of calculating the economic viability was proposed. This method calculates the viability by looking at the reduced costs of errors.
All together, future research in this field should focus on reporting data values and rules used to detect medical features. If this data becomes known, modelling a wearable sensory ecosystem is something that is very much doable using the method described in this thesis. However, as the context of this thesis was the military, chances are these data values and rules are already known within the
military. In that case, modelling (and even automating this modelling) of a wearable sensory ecosystem is very much doable.
6
Discussion
6.1
Practical value and compliance with practical
con-straints
Unfortunately, the development of sensors and medical frameworks is not yet mature enough to the extent that we can reuse this data in this thesis. Although the increasing attention into this field is promising, more attention should be paid to a generic method of reporting on data processing and reporting by these sensors. A basic guide into the data processing and reporting by these sensors is included below. When a more generic method of reporting on data processing and reporting is common, meeting these requirements will hopefully become easier.
Medically speaking, very few symptoms are pathognomonic, e.g. only appear in combination with one specific condition. If more than one symptom is ob-served, analyzing these symptoms might help in ruling out incorrect hypotheses and selecting the right one. Besides, not all symptoms are easy to classify using sensors, which limits the diagnostic capabilities of both methods. This is an issue for which no clear fix is available.
An issue that has not been addressed in this thesis is how to determine the right level of granularity of the hypotheses the system has to be able to diagnose. In an ideal world, the system should be able to diagnose the exact condition the sol-dier is suffering from. In reality, however, it might be more practical to limit the granularity to the minimal level that is required to determine the right medical action (for example, whether or not the soldier should be evacuated). Future work should investigate how the right level of granularity can be determined best.
In both methods, combined probabilities of a sensor being broken and a specific hypothesis to be true are used frequently. However, the probability of a sensor being broken is - unless extensively tested - a guess at best. When more sensors are involved, these combined probabilities are then combined with the proba-bility of the next sensor being broken and so on. This results in a combined probability that is at best a vague guess. Therefore, some form of standard-probability for these occasions should be considered.
The methods outlined in this thesis are highly theoretical and need to be further validated in a controlled environment. The purpose of this validation would be twofold: first, it should establish whether the inputs required in the model by Stefik can actually be found. Second, and this is more important, it should be established if some legitimate concerns with both methods of comparing sensor configurations can be eliminated.
6.2
Comparing sensor configurations
These concerns are multiple. First of all, both methods assume an observation can be categorized, as was done in the examples of blood pressure and heart rate. In many cases, such a categorization might not be possible, for example because the values depend on external factors or the values are non-linear. In many cases, in order to be able to make such a categorization, having extra data might be necessary.
Second, both methods assume sensors cannot break down in such a way that they start generating false data. In reality however, smartwatches capable of measuring heart rate are often suffering from unreliable measurements due to sweat, dust, movement or otherwise. In future work, the possibility of this type of errors should be assumed and the role of time-based sensing should be investigated. This type of errors might be detectable when occurring in a time-based series of observations.
Third, the normalized probabilities of the composed hypotheses in table 4 are based on prior probabilities that are at best a vague estimate. This is a problem that is unlikely to be resolved soon. One way of refining these estimates might be through large-scale simulation of sensors.
6.3
Quantifying cost and value of sensors
Several issues were raised with regard to the quantification of cost and value of sensors during the process of this thesis. This thesis employs the thought that the value of a sensor can be derived from the amount of uncertainty it reduces. However, not all uncertainty is worth equally much: reducing uncer-tainty about life-threatening conditions is much more valuable that reducing uncertainty about a common cold. Therefore, a connection was made that ties the reduction of uncertainty to the average cost resulting of all hypotheses. Thereby, if uncertainty about a very ”expensive” scenario could be reduced, the sensor would provide more value for money. However, this assumes that the cost of all possible scenarios are known in advance. Some scenarios, however, might not be quantifiable at all.
Lastly, during the process of this thesis, several methods of including the actual cost of a sensor and the cost of ownership into the model were discussed. All of them lead to the following question: how should the costs of a sensor be mapped to the actual results it produces, in order to find the most cost-efficient sensor set? Several options were discussed, such as mapping the costs over the number of hypotheses that could be confirmed or denied (to a certain extent) and map-ping the costs over the number of medical features it can extract. However, the number of medical features a sensor can extract does not tell us much about the value of these features. On top of that, some features might become much more valuable when combined with other features. Mapping costs over the number of hypotheses, on the other hand, requires some user-set threshold to determine to what extent of certainty an hypothesis should be confirmed or denied. However, this threshold would likely be linked to the severity of the condition, making it
very difficult (if not non-maintainable, from a data-management perspective) to find the right cost of a sensor. The chosen approach, calculating sensor cost and value by mapping them to the cost of consequence of all hypotheses, is not ideal but is in this case the best combination between practicality and truthfulness. In future work, more attention should therefore be paid to the general issue of cost and value of sensors.
6.4
Researching into new sensors: what to publish with
regard to data
When developing a new sort of wearable sensor, one should take into consider-ation that this sensor will one day be part of a larger sensory ecosystem. If this sensor has to function correctly, it is key that data values and business logic are concisely reported.
In general, the following information should always be included about the com-ponent observed by the sensor:
• Data values: possible values • Data values: normal values
• Data values: predicted values and rules (inputs and their relations) with which these values are predicted
• For each of the data values, report the units in which they are measured • If some sort of classification of data values is used, report on this
classifi-cation
• ”Business logic”: which hypotheses are explained to what extent by which observations?
• If any hypotheses can be confirmed or denied in conjunction with other observations, report these observations and their expected values for and hypothesis to be confirmed or denied as well.
• Probability of sensor breakdowns
• Probability of an observation to be made or how this probability can be determined
With the above information reported, it becomes much easier to implement a new sensor into an existing sensory ecosystem.
References
[1] Urs Anliker, Jamie A Ward, Paul Lukowicz, Gerhard Troster, Francois Dolveck, Michel Baer, Fatou Keita, Eran B Schenker, Fabrizio Catarsi, Luca Coluccini, et al. Amon: a wearable multiparameter medical moni-toring and alert system. IEEE Transactions on information technology in Biomedicine, 8(4):415–427, 2004.
[2] Ruzena Bajcsy. Active perception. Proceedings of the IEEE, 76(8):966– 1005, 1988.
[3] Alexander Boer. Information, utility, agents, and good arguments. To be published.
[4] R¨udiger W Brause. Medical analysis and diagnosis by neural networks. In International Symposium on Medical Data Analysis, pages 1–13. Springer, 2001.
[5] Bruce G Buchanan, Edward Hance Shortliffe, et al. Rule-based expert sys-tems, volume 3. Addison-Wesley Reading, MA, 1984.
[6] Samjin Choi and Zhongwei Jiang. A novel wearable sensor device with conductive fabric and pvdf film for monitoring cardiorespiratory signals. Sensors and Actuators A: Physical, 128(2):317–326, 2006.
[7] Wan-Young Chung, Young-Dong Lee, and Sang-Joong Jung. A wireless sensor network compatible wearable u-healthcare monitoring system using integrated ecg, accelerometer and spo 2. In 2008 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pages 1529–1532. IEEE, 2008.
[8] Aitor Coca, Raymond J Roberge, W Jon Williams, Douglas P Landsittel, Jeffrey B Powell, and Andrew Palmiero. Physiological monitoring in fire-fighter ensembles: Wearable plethysmographic sensor vest versus standard equipment. Journal of occupational and environmental hygiene, 7(2):109– 114, 2009.
[9] V´ıctor Custodio, Francisco J Herrera, Gregorio L´opez, and Jos´e Ignacio Moreno. A review on architectures and communications technologies for wearable health-monitoring systems. Sensors, 12(10):13907–13946, 2012. [10] Marco Di Rienzo, Francesco Rizzo, Gianfranco Parati, Gabriella Brambilla,
Maurizio Ferratini, and Paolo Castiglioni. Magic system: A new textile-based wearable device for biological signal monitoring. applicability in daily life and clinical setting. In 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, pages 7167–7169. IEEE, 2005.
[11] Brian J Eastridge, Robert L Mabry, Peter Seguin, Joyce Cantrell, Ter-rill Tops, Paul Uribe, Olga Mallett, Tamara Zubko, Lynne Oetjen-Gerdes,
Todd E Rasmussen, et al. Death on the battlefield (2001–2011): implica-tions for the future of combat casualty care. Journal of trauma and acute care surgery, 73(6):S431–S437, 2012.
[12] Karl E Friedl, Mark J Buller, William J Tharion, Adam W Potter, Glen L Manglapus, and Reed W Hoyt. Real time physiological status monitoring (rt-psm): Accomplishments, requirements, and research roadmap. Techni-cal report, DTIC Document, 2016.
[13] Yuan Gao, Yuanjin Zheng, Shengxi Diao, Wei-Da Toh, Chyuen-Wei Ang, Minkyu Je, and Chun-Huat Heng. Low-power ultrawideband wireless telemetry transceiver for medical sensor applications. IEEE Transactions on Biomedical Engineering, 58(3):768–772, 2011.
[14] Frederick Hayes-Roth. Rule-based systems. Commun. ACM, 28(9):921– 932, September 1985.
[15] Reed W Hoyt, Jaques Reifman, Trinka S Coster, and Mark J Buller. Com-bat medical informatics: present and future. In Proceedings of the AMIA Symposium, page 335. American Medical Informatics Association, 2002. [16] K Hung, YT Zhang, and B Tai. Wearable medical devices for
tele-home healthcare. In Engineering in Medicine and Biology Society, 2004. IEMBS’04. 26th Annual International Conference of the IEEE, volume 2, pages 5384–5387. IEEE, 2004.
[17] Shigehito Iguchi, Hiroyuki Kudo, Takao Saito, Mitsuhiro Ogawa, Hirokazu Saito, Kimio Otsuka, Akio Funakubo, and Kohji Mitsubayashi. A flexible and wearable biosensor for tear glucose measurement. Biomedical microde-vices, 9(4):603–609, 2007.
[18] Christos D Katsis, George Ganiatsas, and Dimitrios I Fotiadis. An inte-grated telemedicine platform for the assessment of affective physiological states. Diagnostic Pathology, 1(1):1, 2006.
[19] Stephen Chiu Kwok Lam, Kai Lap Wong, Kwok On Wong, Wenxiu Wong, and Wai Ho Mow. A smartphone-centric platform for personal health moni-toring using wireless wearable biosensors. In Information, Communications and Signal Processing, 2009. ICICS 2009. 7th International Conference on, pages 1–7. IEEE, 2009.
[20] Mitchell M Levy, Mitchell P Fink, John C Marshall, Edward Abraham, Derek Angus, Deborah Cook, Jonathan Cohen, Steven M Opal, Jean-Louis Vincent, Graham Ramsay, et al. 2001 sccm/esicm/accp/ats/sis interna-tional sepsis definitions conference. Intensive care medicine, 29(4):530–538, 2003.
[21] Tao Liu, Yoshio Inoue, and Kyoko Shibata. Development of a wearable sensor system for quantitative gait analysis. Measurement, 42(7):978–988, 2009.
