=Paper=
{{Paper
|id=Vol-1213/paper8
|storemode=property
|title=Medical diagnostics based on combination of sensor and user-provided data
|pdfUrl=https://ceur-ws.org/Vol-1213/paper8.pdf
|volume=Vol-1213
|dblpUrl=https://dblp.org/rec/conf/ecai/SomrakGLMSG14
}}
==Medical diagnostics based on combination of sensor and user-provided data==
https://ceur-ws.org/Vol-1213/paper8.pdf
Medical diagnostics based on combination of
sensor and user-provided data
Maja Somrak1 , Anton Gradišek1 , Mitja Luštrek1 , Ana Mlinar2 ,
Miha Sok3 , and Matjaž Gams1
1 Jožef Stefan Institute, Jamova cesta 39, 1000 Ljubljana, Slovenia
2 MESI d.o.o., Leskoškova cesta 9d, 1000 Ljubljana, Slovenia
3 Faculty of Medicine, University of Ljubljana, Vrazov trg 2, 1000 Ljubljana, Slovenia
{maja.somrak,mitja.lustrek}@ijs.si
Abstract. We present an approach that incorporates multiple machine-
learning and data mining algorithms for prediction of the user’s medical
condition. The decisioning is based on vital signs data and user-provided
input regarding the symptoms expressed. The presented method was
trained and tested on virtual patients, generated using expert medical
knowledge. We discuss future steps in the method development.
Keywords: medical diagnostics, questionnaire, sensor data, vital signs,
user input
1 Introduction
Being able to diagnose common diseases at home can help the patient decide
when to go to the doctor and at the same time reduce the burden on healthcare
system. Here, we present a diagnostic method that incorporates information
about user symptoms and vital signs readings to predict user’s medical condition.
The method was developed as a part of the diagnostic software for HealthStation
HOME, a device competing at the Qualcomm Tricorder XPRIZE $10 million
challenge [1]. We present initial results of experimental tests on virtual patients
and outline possible improvements in future.
1.1 HealthStation HOME
The HealthStation HOME system [2] consists of a set of sensors that measure
vital signs, such as heart rate, breathing rate, blood pressure, body temperature
and blood oxygen saturation. The collected data is communicated to a mobile
device for further diagnostics. The measurements can be interpreted as patholog-
ical symptoms (e.g. high blood pressure) that serve as an input for the diagnostic
application. The user obtains an evaluation of his health condition by running
the diagnostic application on the mobile device by selecting one of the starting
options, I feel pain or I feel unwell (see Fig. 1). The application then guides the
user through intelligently selected questions about the symptoms that are recog-
nized as relevant. The result of the diagnostic method is the initial, home-based,
36
medical condition assessment and a recommendation for further diagnostic test-
ing (additional HealthStation HOME modules, such as blood or urine tests) to
finally confirm or reject the diagnosis.
2 The Method
The HealthStation HOME diagnostic method is designed in form of a question-
naire with multi-modal inputs. The overview of the procedure is shown in Fig.
1. The initial input for the method consists of three types of data: (1) identified
risk factors from the user profile data (e.g. smoking), (2) recognized pathological
symptoms from vital signs measurements or other sensor data (e.g. high blood
pressure) and (3) user selected pain symptoms in the application (e.g. chest
pain). There are 60 predefined symptoms that the method can operate with,
each of which can be treated as unknown or known, where known symptom is ei-
ther present or absent in a patient. The present symptoms from the three inputs
form the initial set of symptoms (4), which serve as the basis for automatically
compiling a list of additional symptoms (5), from which the user is expected to
select those he/she is experiencing. This list is generated to include both the
symptoms that the user most probably experiences at the time and would prob-
ably want to report, and also the most relevant symptoms that would help the
physician or the diagnostic method set a reliable diagnosis.
7
PROFILE 1 SENSOR 2
QUESTION ABOUT
DATA DATA
A NEW SYMPTOM
(identified (recognized
(highest IG)
risk factors) symptoms)
NOT
CONFIDENT
4 5 6
INITIAL ADDITIONAL 8 CONFI-
I FEEL DISEASE CONFI- INITIAL
SYMPTOMS SYMPTOMS DENT
UNWELL PREDICTION DENT DIAGNOSIS
(merged) PREDICTION
(ARs + mRMR)
3
and SELECTION
PAIN (user selected)
I FEEL
SYMPTOMS
PAIN
(user selected)
Fig. 1. Method overview. The numbers in circles indicate step numbers, as explained
in the text.
If there is at least one symptom in the initial symptoms set, the method
aims to find other symptoms that often emerge together with one of these initial
37
symptoms. For this purpose, association rules (ARs) [3–5] of type symptom A Õ
symptom B are searched, where symptom A is any of the initial symptoms and
symptom B is any of the unknown symptoms. Rules with the highest confidence
and minimal support condition satisfied are selected in order to produce a set of
probable additional symptoms.
In addition to the ARs, the minimum-Redundancy-Maximum-Relevance [6]
(mRMR) method is used to identify the most informative, mutually indepen-
dent symptoms that have not been examined yet (unknown symptoms). This
method works even if there are no symptoms in the initial symptom set. The
mRMR resulting attributes subset is the subset of attributes (symptoms) that
a) provide a lot of information about the class (medical condition) and b) are
at the same time mutually uncorrelated. The criterion a) is measured with the
mutual information between each attribute and the class, while b) is measured
with the mutual information between the attributes. The mRMR rule used in
our method is defined upon mutual information difference (MID) criterion [6]
with the following equation
1 X
max [I(i, h) − I(i, j)]. (1)
i ΩS |S|
jS
For selecting each additional symptom, an iteration of mRMR calculation
over all unknown symptoms is repeated to find a symptom for which the value
of the function is maximized (Eq. 1). I(i, h) is mutual information between i,
a symptom from the set of unknown symptoms ΩS , and h, the classification
variable − the medical condition. Likewise, j is a symptom from the set of
known symptoms S, containing |S| symptoms. Once a new symptom i is selected
and added to the additional set of symptoms, it is treated as one of the known
symptoms (the symptom is moved from ΩS to S) in next iteration of mRMR
calculation.
In the following step, the information about present and absent symptoms is
first used for the disease prediction (6). The probabilities for predefined 15 dif-
ferent medical conditions (14 diseases and ’healthy’) are evaluated using a set of
J48 classifiers, one for each condition. There are two probability thresholds that
represent medium and high chance for a certain medical condition, empirically
selected to be 40% and 80%, respectively. If all condition probabilities fall below
the medium threshold (improbable condition) or above the high threshold (very
probable condition), the prediction is considered confident and the diagnostic
procedure terminates, retrieving the diagnosis. However, if one or more medical
condition probabilities lie between the medium and high threshold (neither very
probable nor improbable condition), in the so called gray zone, the disease pre-
diction is not considered confident. In this case, information about at least one
additional symptom is needed to obtain a confident prediction. This is obtained
by asking user a question about a new symptom (7). The additional symptom
is chosen according to the highest information gain (IG), where the values are
recalculated from a reweighted training set, such that the instances with the
conditions from the gray zone are assigned higher weights. This approach, espe-
38
cially when incorporating reweighting, reduces the number of questions required
for the probabilities to emerge out of the gray zone, as opposed to randomly se-
lected questions. In case any condition probabilities still remain in the gray zone
after a maximum number of question has been asked, the procedure terminates,
selecting the medical condition with the highest probability for the diagnosis.
3 Experiments
We utilized expert medical knowledge to obtain the patient data sets. For this
purpose, a table correlating 15 different medical conditions with over 60 symp-
toms was developed by physicians. The table was used for generating the training
set containing 15000 virtual patients. Additionally, a test set of 1500 virtual pa-
tients was generated, where each medical condition was present in 100 patients.
The tests demonstrated high sensitivity and specificity. For example, for otitis
media, 94% of the patients with this medical condition were correctly identi-
fied while 84% of patients, diagnosed with otitis media actually had the disease.
In the case of leukocytosis, the corresponding values were 59% and 60%, re-
spectively [7]. On average for all medical conditions, the obtained values for
sensitivity were 88.4%, for specificity 88.6% and for accuracy 88.3%. Currently,
we are collecting the data of real patients for further testing, an example of a
patient answering to the questions is shown in Fig. 2.
Real disease: Atrial fibrillation 1.) Does your skin and whites of your eyes seem to be
yellow? NO
Initial symptoms: 2.) Have you experienced sudden numbness or weakness
• Rapid breathing on one side of the body? NO
• High heart rate 3.) Do you suffer from blurred vision? NO
• Irregular heart rate 4.) Does your cough produce sputum or mucus? NO
• Chest pain or chest tightness 5.) Do you suffer from vertigo? YES
• Age over 70 years 6.) Have you experienced sudden troubles speaking? NO
7.) Have you been told that you snore? NO
8.) Is your urine cloudy or dark colored? NO
9.) Does your urine have a strong or decaying smell (e.g.
Additional suggested symptoms: smell of fish)? NO
Shortness of breath = YES 10.) Are you able to raise both arms evenly? YES
A sudden start of the disease = YES
Feeling unwell (malaise) = YES
Blood in cough = NO
Swollen glands in the neck = NO Predicted disease: Atrial fibrillation, probability = 99 %
Cough = NO
Fig. 2. Example of testing the diagnostic method on a real patient with atrial fibrilla-
tion.
39
4 Discussion and Conclusion
The initial results, based on both training and testing the method on virtual
patients, show very high sensitivity, specificity, and classification accuracy (all
over 80%). These values are probably too optimistic, according to the opinion of
our medical associates. One of the main reasons for these results is that both the
training and the testing set were generated from the same expert table. Because
virtual patients are not biased when answering the questions, it is even more
necessary to train and test the method on real patients, which we plan to do in
the future. Moreover, the medical conditions were classified only into 15 different
classes, which is far below the number of possible medical conditions in reality.
The predefined medical conditions are also very distinctive in terms of symp-
tom manifestation and it is therefore easier to distinguish between them (higher
classification accuracy). The exceptions here are chronic obstructive pulmonary
disorder, pneumonia, tuberculosis, and sleep apnoea; they are more frequently
misclassified due to the similarity of their symptoms. In the future, we intend to
incorporate hierarchical classification (e.g. additional class ’pulmonary disease’),
when the data is insufficient for reliable differentiation between similar medical
conditions. Current implementation of the method utilizes only the question-
naire for all of the symptom. In the future, we will use actual sensor input to
determine the presence of a few specific symptoms. Additionally, we plan to in-
clude a larger number of medical conditions and implement intelligent methods
for multilabel classification for discovering combinations of conditions.
Acknowledgments. We would like to thank MESI, DLabs, Gigodesign, Fac-
ulty of Medicine, and other partners, collaborating in the MESI Simplifying
diagnostics team for the Qualcomm Tricorder XPRIZE competition.
References
1. Qualcomm Tricorder XPRIZE, http://www.qualcommtricorderxprize.org/
2. MESI, http://www.mesimedical.com/home/
3. McCormick T.H., Rudin C. and Madigan D.: A Hierarchical Model for Association
Rule Mining of Sequential Events: an Approach to Automated Medical Symptom
Prediction. Annals of Applied Statistics. (2012)
4. Soni S. and Vyas O.P.: Fuzzy Weighted Associative Classifier: A Predictive Tech-
nique For Health Care DataMining. International Journal of Computer Science,
Engineering and Information Technology (IJCSEIT), Vol. 2, No.1, (2012)
5. Soni S. and Vyas O.P.: Using Associative Classifiers for Predictive Analysis in Health
Care Data Mining. International Journal of Computer Applications, Vol. 4, No. 5,
(2010)
6. Peng H., Ding C.: Minimum Redundancy Feature Selection from Microarray Gene
Expression Data. Journal of Bioinformatics and Computational Biology, Vol. 3, No.
2, (2005)
7. Somrak M., Luštrek M., Šušterič J., Krivc T., Mlinar A., Travnik T., Stepan L.,
Mavsar M., Gams M.: Tricorder: Consumer Medical Device for Discovering Common
Medical Conditions. Informatica 38, Ljubljana (2014)
40