Patient Trials

Patient Trials

A close collaboration with medical facilities is an important step in creating
a reliable monitoring system. Assuming a lack of knowledge, concerning the
correlation of medical signals, it is reasonable to conduct a long term experimental
procedure. We should run continuous tests in order to collect enough
data from diferent types of patients, with different health conditions and from
different age groups. This will help to take into account every particular detail
and combine a universal set of rules for a reasoning mechanism of the system.
As a first step, we decided to conduct a short term experiments in order
to evaluate systems reliability and discover its weak points to simplify future
development. The following chapter represents a short profile information of
the patients involved in the trials and results for the data analysis described in
Chapter 4.

5.1 Target group

There are two main groups of potential users for the developed system which
were in focus during experimental part. The first group includes a person with
normal health parameters, who has no chronic desease. The data, received from
this target group will help to compare results with the second and more important
one. The second group includes two persons from a medical facility named
Backagården in Örebro, Sweden. These people were chosen by the personnel
and agreed to participate in the experiment after preliminary consultations. A
special step-by-step tutorial was issued and handed out to nursing personnel,
who had consistent access to a patient and was able to provide any kind of
assistance if needed (see Appendix A).
The whole experimental part was conducted in the Backagården medical
center for elderly people (vårdcentral) and involved collaboration between center’s
authorities, nursing personnel and previously chosen patients. The format
of the experiment was discussed and approved on a meeting beforehand. Data
collection was performed with one patient at a time letting him wear a device

for three days without any break and more importantly during the night time.
Keeping experiment in this particular way was supposed to launch a platform
for further testing and create an unique opportunity to monitor and subsequently
register any kind of sleeping disorders. Moreover, a chosen person was
able to maintain his normal level of activity during the day according to a
schedule, which includes taking a long promenade, taking a meal and even cycling.
In case of strong demand for any unacceptable for sensor functionality
activity (e.g. taking a shower, other medical testing), the system can be easily
paused and launched again after necessary action is finished. No data will be
lost in this situation and measurements will be written to the same file.
The following section describes a health profile of the patients involved in
the experiment and data processing results for each of the cases. It is important
to mention, that both patients had different health profiles which gave experimental
part more coverage. All the information is published after preliminary
agreement of the patient.

5.2 Results

Healthy Person
A first stage of the experimental part is processing of the data, received from a
relatively healthy person. An experiment was conducted for two days and two
nights and no significant problems were encountered during this period. No
data was lost or corrupted, a sensor and a phone were maintained and charged
in case of low battery indication. A graphical result is plotted in figures below.
We initially applied a change point detection mechanism (see Figure 5.1,
second plot) to check when the signal dropped or jumped during the monitoring
process. And the second part of analysis performed the anomaly detection,
implemented using the HOT SAX method (see Section 4.2). Anomaly detection
results for a healthy person are presented in the same graph.
We consider it more illustrative to present this data as an activity subplot
(see Figure 5.2) for this particular case and compare pulse variation to a data
received from accelerometer. In many occasions, change in a pulse variation
was caused by a significant change in activity which can be detected from the
graph.We can also relate any event to a particular time, which improves a level
of analysis.

56 CHAPTER 5. PATIENT TRIALS
In this case, and for the rest of the plotting, red sections on the graph (Figure
5.2) are representing areas of interest or situations, which would require
further medical investigation. The depicted plot is a small section cut from the
whole dataset to demonstrate an algorithm performance within several hours
of monitoring.
Patient I
Table 5.1: Personal information (Patient I)
Birth year (Age) 1924 (87 years)
Diagnoses Dementia(non-specified), Hypertension,
Kidney insufficiency, Hypothyreosis,
Depression, Falling tendency
Hospital care Admitted to hospital one time during the last 10 years;
latest March 2010, two days in the Dept. of Surgery
at the University Hospital Örebro
During the next part of the experiment system was launched on the first
patient from Backagården medical center (see Table 5.1). We were expecting
several challenges to appear during this stage, therefore it was considered as
a pilot test and was supposed to uncover all the weak points of the system.
The most important issue refers to the level of collaboration with medical personnel.
None of the patients is capable of dealing with any kind of advanced
technical equipment, which makes it crucial for nursing staff to be around.
In a worst case scenario we have three situations that interrupt a consistently
running monitoring system:
• a sensor runs out of battery
• a smart-phone runs out of batteries
• a patient takes off or displace a sensor thimble
Even though, the first two points are highly related to each other, all three situations
require same measures and approach. If personnel checks a system and
launch a data collection process every time it was interrupted, we will be able
to minimize the loss of the data to nearly zero. Normally, all three situations
are unlikely to emerge more than once or twice a day. So, if a system is checked
within the day several times we avoid a data loss significantly. Another approach
is to set a sensor in a special mode which sends a particular command
to a phone and warns a user about low battery condition. The same procedure
can be done with a smart-phone. In case of sensors thimble displacement, the

operation should be repeated again. If a relatively young person is being monitored,
these issues are unlikely to cause any problems. A user can pause or stop
a system for a short period of time and then launch it again when it is necessary.
In Figure 5.3 and Figure 5.4 we can observe plotted data of the first patient:
Figure 5.3: Patient I result

Birth year (Age) 1937 (74 years)
Diagnoses Dementia (vascular), Diabets mellitus,
Pituitary tumor (prolactinoma), Gait problems; walker,
Polyneuropathy with pain in legs, Depression
Hospital care Admitted to hospital four times during the last 10 years;
latest November 2006, four days in the Dept. of Surgery
at the University Hospital Örebro
Assuming the health profile of the second patient, presented in Table 5.2,
an essential addition to the system could be made. A special blood sugar rate
sensor would help to introduce another vital parameter and improve the level
of analysis. Additionally, a group meeting was held before starting a second
part of the experiment, which helped to improve the monitoring process significantly
and make it more robust. A test was initiated for two day and two
nights and differed from the first one with a low number of interruption. A
system was stopped only for a short period of time (several seconds) in order to
change sensor battery. That allowed us to receive a stable output and simplified
a processing of the data. A pulse and oximetry variation of the second patient
are depicted on the figure below.
A difference in the health profile of the second patient reflected in a final
result which can be observed in Figure 5.5 and Figure 5.6. The pulse value is
varying within a wider range and danger points are detected more frequently.
A next figure is representing results from anomaly detection part where the unusual
sections are again highlighted with red.We use a subplot in order to show
a change point detection performance on the same data sector. This will help
to analyze the performance of both algorithms and compare results.
All the further professional derivation based on the presented graphical results
are to be made by medical specialists. At this point, algorithms applied
for measurements are tend to simplify processing of the data in terms of future
medical investigation. However, we are able to provide a general overview of
the system performance and summarize the main analyzed parameters.

Overview

In order to summarize the performance of the system we combined a table
of the most essential parameters (see Table 5.3) including percent of data loss
during experiment and number of danger points detected for each case.
Moreover, it is possible to make several important observations based on
the data processing results presented above. The subplot of the anomaly and
change point detection can prove the complementary character of both methods

as it was claimed in Chapter 2 and Chapter 4. The same section of the
signal was processed with different approaches and results do not replicate
each other. In other words, the detected anomaly section does not necessarily

1st Patient 2nd Patient Healthy person
duration, hours 63 45 45
data loss,% 9,5 2,2 1,1
danger points 20 23 10
contain a change point and vice versa. On the cutout in Figure 5.7 we consider
the same period of time, processed first by change point detection method and
later by anomaly detection algorithm. In the first case all the detected points are
Figure 5.7: Change point detection vs. anomaly detection cutout
corresponding to an abrupt drop/jump in the signal and should be registered
for further investigation. However, this information is not enough if we want
to see a typical signal behavior for the current patient. Therefore, we should
add an anomaly detection procedure and find the most unusual patterns in the
dataset.
Furthermore, one of the possible SAX algorithm outputs is the exact index
of the registered anomaly. We can use this index to check a corresponding
correlation coefficient. As a result we determine behavior of the measured parameters
within anomaly section which will help to combine a patients health
profile in the future development. The coefficient can vary between 0 and 1
corresponding to the low and high correlation respectively. The bigger the coefficient
the more proportional two signals are. If this number takes a negative
value, proportionality is inverse. A table below provides a summary of the correlation
coefficients corresponding to particular anomaly sections in the signal
represented by an index.

Every index in the table corresponds to a particular time period in the
dataset and therefore, it is possible to determine, which type of correlations
is more typical for each patient and compare this values on different time sections.
Finally, it is important to mention that a variable number of situations might
occur during the monitoring process and data collection. They depend on patients
health condition and more importantly on the number of chronic disease
that he or she has. This fact makes it harder to implement a unique and universal
decision-making algorithm which can cover all possible scenarios. However,
processing algorithms, described above can provide a unique information for
detecting these situations and help to create a combination of rules for sufficient
and reliable decision making algorithm.
if u like the post just say thank u in comment box.