Personalised Context Aware Content Relevant

Disease prediction and Diet Recommendation System

T Ramakrishnudu Tangediaplli Balakrishna
Dept. of Computer Science and Engineering M.Tech(CSE)
National Institute of Technology National Institute of Technology
Warangal, Telangana, India Warangal, Telangana, India
[email protected] [email protected]

Abstract— the problem of predicting possible diseases in the The newest report released by the Ministry of Health
future reduces some of the diseases in Health recommendation and Welfare shows us that the major part of the top 10 causes
system by providing suitable diet and exercise recommendations of death are chronic diseases. According to the report, in 2015,
To address the challenges due to depending on individual users
more than half of the top 10 causes of death are some kinds of
are resolved by the context logs of different existing users so we
have used this concept in proposing a similarity based approach
chronic diseases, such as the heart diseases, hypertensions,
Disease Prediction and Diet Recommendation systemdiabetes, chronic liver diseases or kidney diseases. In this
(DPDR) where normalization is done to get most possible high-tension modern life, the result seems reasonable, but that
diseases. The performance study shows that the proposed method it should not be acceptable, because there are many self-health
is the first approach for predicting diseases based on the contextmanagement ways to make people stay away from the chronic
logs. Experimental results have shown better results than the diseases. Unlike cancers, most chronic diseases can be well
existing system in terms of efficiency and complexity for the controlled, even can be prevented, if people can live with a
proposed similarity function. healthy lifestyle, such as the regular sleep, balanced diet and
proper exercise.
Keywords – Context logs, Similarity Function, Context aware,
normalization.
The res the rest of this paper is organized as follows.
SectionII briefly presents the relevant concepts and
definitions. In Section III,IV Existing recommendation system
I. INTRODUCTION
and challenges. The proposed algorithm is presented in
A recommendation system is a software program which Section V. Section VI, illustrate the experimental results. The
attempts to narrow down selections for users based on their concluding remarks are finally made in Section VII.
expressed preferences, past behavior, or other data. Different
recommendation systems have been extensively studied and II CONCEPTS AND DEFINITIONS
first introduced in [3,4,13]and several types of recommendation
systems in [12,13] and several context aware recommendation
system like for restaurant, tourist based and location based Similarity Function:
studied in [ 5,6,10].Even though there are different
recommendation systems exists there is no proper Similarity is for determining the degree match between the
recommendation system for disease prediction and diet query and item. It is one of the fundamental concepts in
recommendation system. recommendation system and measured by how much closer
two items are related. Similarity score is a normalized value
Jerry C.C. Tseng, Kuen-Rong Lo[7] analyzes the result ranges from 0 to 1.There are several functions existing for
of regular physical examination to evaluate the health risk and finding out the similarity value
provide personalized healthcare services for users in terms of
diet and exercise guideline recommendations. They developed To obtain the similarity between two numeric
some interactive ways for users to easily feedback their vital contextual preference cases, we used the formula as following
signs to the system and quickly get the suggestions for health
management from the system S(C n , C )
Importance of disease Prediction recommendation ie = {1 − p if 0 ≤ p<=1
system: Nowadays people can easily get the information of 0 p > 1}
their glucose, blood lipids, blood pressure and waist
circumference to be aware of their own health status. Self- Similarity (Cx, Cy) = 1 -- |C l−C e|
Max−Min
management is the key factor in preventing chronic illness, but
only few people can keep the healthy living style with enough
Total Similarity(Cx,Cy) = Similarity(Cx, Cy) * Weight
exercise and good diet habits mostly because of the very busy
and high pressure modern life consequently, the populations of
chronic diseases are getting more and younger than ever.

1|Page
Max is the maximum value of the ith features of all cases in only consider those contexts that are relevant to content usage
the case base and Min is the minimum value of the ith for reducing the computation complexity, but it takes large
feature for all cases in the case base amount of time and memory which is not efficient when there
Similarity is for retrieving the most similar is lot of missing information.
contextual situations and preferences. This is an important Jerry C.C. Tseng, Kuen-Rong Lo[7] analyzes the
phase of the disease prediction process. To determine the
result of regular physical examination to evaluate the health
expected preferences of the user in the current contextual
risk and provide personalized healthcare services for users in
situation, it requires a method or algorithm for determining the
degree of similarity between contextual situations. terms of diet and exercise guideline recommendations. They
developed some interactive ways for users to easily feedback
their vital signs to the system and quickly get the suggestions
Assigning Weight to Context Variables: for health management from the system
When comparing two contexts, the different context variables Jong-Hun Kim, Jung-Hyun Lee[8] proposes
composing them have to be compared. In principle the entire personalization diet recommended service for the users who
context variable can be considered as equally important; so, we require the prevention and management for coronary heart
could assume that each context variable has a weight that is disease. This service consists of a collection module that draws
obtained by dividing one between the numbers of context nutrients, which are to be adopted by users based on the
variables. However, there are circumstances where some collection of some constraints in users, a configuration module
variables are more important than others, which should be that determines the preference of foods through the input of the
considered during the recommendation process diet of users, and a scoring module that makes a score for the
extractable diet.
Weights wi ∈ [0,1] is assigned to the context variables,
Higher weights to the context variables that are more relevant Lin, Y.-F., Chu, C.-H [9] presented a social
to the domain. The user could consider some context variables media-based recommender system which makes
as irrelevant, assigning them a weight of zero. Alternatively, recommendations by considering a user's own health
he/she could highlight the importance of some context concerns, the trustworthiness of the information providers,
variables over others, assigning them higher weights. the similarity between the user and the information provider,
and the test item's general acceptance in the social media. In
Data Normalization
particular, they proposed a semantics-enhanced fuzzy-based
Data Normalization is done in order to find the measure of model to facilitate recommendation. The model consists of
exactness it determines the fraction of relevant items retrieved three important factors affecting recommendation in health
out of all items retrieved. We can understand very easily that if care social networking environments: trust, similarity and
the sum of contexts total similarity sum is greater than review. Fuzzy logic is used in the model because it is
threshold then it first computes the occurrences of each context tolerant of imprecisely defined data and can model non-linear
variables and divides it with the sum of all the context variables functions of arbitrary complexity. Most importantly, fuzzy
to obtain the precision of each individual context variable even logic can accommodate vagueness, intuitive and experiences
after the normalization if more number of the diseases remains in modeling recommendation in a healthcare social network,
then there is a threshold value even for normalization. If the because human observation forms the basis of
precision of each disease is more than that then only it adds the recommendation assessments. Semantics-based profile
diseases to the list of predicted diseases.
similarity metric is adopted to measure the similarity
tp proposed Personalized Healthcare Recommender Based on
Precision = tp+ fp Social Media in which item ratings are used in the Similarity
function to give the suggestions to the users interactively.
.
IV CONTEXT AWARE RECOMMENDATION SYSTEM
III RELATED WORK
The basic idea is first mining common context-aware
preferences from the context logs of many users and then
Hengshu Zhu and Enhong Chen[1] illustrated how to representing each user’s context-aware preference by a
extract personal context-aware preferences from the context-
distribution of common context-aware preferences. The mined
rich device logs, or context logs for short, and exploit these
personal context-aware preferences can enable the
identified preferences for building personalized context aware
recommender systems design three effective methods for development of personalized context-aware recommender
mining common context-aware preferences based on two systems and other related services, such as mobile online
different assumptions about context data dependency. If advertising.
context data are assumed to be conditionally independent, they Second, we design three effective methods for
proposed algorithm to mine common context aware preferences mining common context-aware preferences based on two
through topic models. Otherwise, if context data are assumed to different assumptions about context data dependency. If
be dependent, exploit the constraint-based Matrix Factorization context data are assumed to be conditionally independent, then
techniques for mining common context-aware preferences and mining of common context aware preferences is done through

2|Page
topic models[14]. Otherwise, if context data are assumed to be 7. Occurrence of each disease after passing threshold test is
dependent, then we have to exploit the constraint-based Matrix counted and normalization of each disease is done by dividing
Factorization techniques for mining common context-aware with sum of all other diseases contribution. Normalization
preferences and only consider those contexts that threshold also has to be set to eliminate less probable diseases.
are relevant to content usage for reducing the computation 8. After disease prediction diet can be recommended as per the
complexity. disease as we have already record of recommended diet from
So here in this paper we are going to use the concept of the authorized doctor.
Using other users Context Logs which will be helpful in 9. Exercise guidelines also can be given to the user if we can
overcoming the data scarcity of traditional recommendation predict the disease in the beginning stage of it.
system .
Algorithm :

C
Input: n ,W
V PROPOSED METHOD
Output: Diseases(D)
1.A registered user enter his credentials and login into the 1: Begin
system and system validates and shows the user interface to 2: TS0
enter his personal information about his health such as blood
pressure, BMI, temperature, TSH etc all the symptoms and 3: ST0.6
also the most important field of diseases already he has
suffered in the past as shown in fig1. Ce
i
4: Retrieve where i=1 to n
5. Max (CV) and Min (CV)
6. for each CV for each L

(C n !=C e )do
7. If

S(C n , C )
ei
8:
9: TS=TS+S*W
n
∑ TsS s
s=1
10: TZ =
Fig1. GUI to enter personal Information of users 11: If (TZ≥ threshold)
2. It uses proposed similarity function to compute the 12: for Each User Log L
similarity score with respect to each user context log in the
data base. Ce
i
3. Context variables such as temperature, blood pressure, X= Normalization (L) for all
cold such parameters are given with the weights as we 13:
described in the fundamental concepts.
4. Total similarity score of each user Log is computed by 14: Else
multiplying with weights.
5. After Obtaining Total Similarity Sum of the weights for 15: Remove L;
each of the users is computed as disease doesn’t depend on the
single weight of the context variable. 16: End If
6. If the total sum for each context log is greater than the
17: D=Max(X)/*if more than one value satisfies
Threshold value then it will be considered for the further
consider all nearby values*/
process. Otherwise that user Log will be dropped as it is not
contributing much for the disease prediction 18: End for

3|Page
 19: Output (D) Proposed method is shown as flow chart in fig2 which describes
the step by step events occurs during prediction of diseases and
diet recommendation

VI EXPERIMENTAL RESULTS

Challenges of developing and evaluating context-aware

disease recommendations
New User
Ultimately, real users under normal and real-world conditions
should evaluate systems meant for human use. However, as
we know, this is impractical most of the time and thus
simulations are generally used instead. Sometimes simulations
Retrieve Past Assign can even outperform the efficacy of real experiments as they
Set
contextual weights allow testing the system under different conditions, which
Threshold
variable data might occur rarely in real-life. Eventually, the ideal solution is
to use both simulation and real experimentation. Accordingly,
it becomes much more difficult to model the expected
behavior of the system. We have thus decided to implement a
prototypical instantiation of the proposed conceptual
Find Max Min from framework together as a case-study application, notably, a
existing data context-aware personalized disease prediction where we have
used synthetic data filtered from large set of real data.

For experimenting we have collected the input data

Compute Similarity from the source http://dm.ustc.edu.cn/datalist.html which is a
with each of data store repository containing 4500 user profiles. It is a
existing Context medical data, containing several details of the patients. For
variable making the experiment simple we have used the synthetic data
after processing unwanted data for our experiment.

Compute
total
similarity

Yes
TZ>=t No
Normalization Ignore
of each context the CV
variable

Disease Prediction

Diet Exercise
Guidelines

Fig2. Flow chart for DPDR

4|Page
Fig.3. Disease Prediction Accuracy of our algorithm w.r.to context Log of the existing users in the data base.It is compared
each disease with the existing Case based methodology and content based
recommendation system in the following graphs.Our algorithm
gave 76% more accurate results than existing systems in case
of disease prediction system in execution with RMSD

In Case of disease prediction shown in fig3 with respect to

each disease it gave better results when we have the data about
the particular disease for example precision of thyroid is
around 0.78 because we have enough context log data related
to thyroid so it gives precise prediction
Whereas leg pain prediction accuracy is 0.54 which
is less because of less content and based on threshold and
weight of the context variables also disease prediction
accuracy varies so these two has to be taken very precisely to
obtain exact results. Based on the disease even proper diet can
be recommended and other related services like exercise
guidelines also can be given to the users.

Fig5. Disease prediction accuracy of DPDR comparison with

case based and collaborative recommendation systems

VII CONCLUSION AND FUTURE WORK

In this paper, we have designed a new similarity function

which compare with the other users context variable Logs with
the new user. With the proposed method we got better results
than the existing where normalization is done to get the most
possible diseases and based on the diseases predicted diet and
exercise guidelines can be provided.
As a future work if it can be implemented in a dynamic
environment it will be more useful for the user where data is to
be updated dynamically.

Fig.4. Our similarity function vs. Pearson correlation REFERENCES

coefficient
[1] Hengshu zhu and Enhong Chen.”Mining Mobile User
Preferences for Personalized Context-Aware
An experiment is done by using Pearson correlation Recommendation”, ACM Transactions on Intelligent Systems
coefficient function in place of our similarity function. A and Technology, New York, Dec 2014.Vol. 5, No. 4, Article
graph is obtained as shown in fig4 where it always gave 58.
negative values between 0 to -1 indicating negative correlation
which means data set is always in downward graph which is [2] J.H. Roh and S. Jin.Personalized advertisement
showing not suitable for numeric type of data where as our recommendation system based on user profile in the smart
similarity function gave very close values to the real time phone. In Advanced Communication Technology (ICACT),
14th International Conference, Beijing Feb 2012, pp.1300-
accuracy. Even in terms of complexity Pearson has O( N 2 ) 1303.
whereas our similarity function is in terms of O (N).
[3]A.M. Otebolaku and M.T. Andrade. Context
A bar graph is drawn as shown in fig5. by taking the several Representation for Context-Aware Mobile Multimedia
diseases on X-axis and its respective accuracy on Y-axis.It Recommendation. In Proceedings of the 15th IASTED
gave better result when the disease data is availble in the

5|Page
International Conference on Internet and Multimedia Systems [12]Tengfei Bao, Happia Cao, Enhong Chen, Jilei Tian, and
and Applications, Washington, Aug,2011.pp 248-262. Hui Xiong. 2010. An unsupervised approach to modeling
[4]Vincent Wenchen Zheng, Bin Cao, Yu Zheng, Xing Xie, personalized contexts of mobile users. In Proceedings of the
and Qiang Yang: Collaborative filtering meets mobile
IEEE International Conference on Data Mining (ICDM’10).
recommendation: A user-centered approach. In Proceedings of
the 24th AAAI Conference on Artificial Intelligence IEEE Computer Society, Washington, DC, 2010, 38–47.
(AAAI’10). Chile,2012.pp.236–241. [13]. Gediminas Adomavicius, and Alexander Tuzhilin
[5] ]Kyoung Jae Kim, Hyunchul Ahn, and Sangwon Jeong. Towards the Next Generation of Recommender Systems: A
Context-aware recommender systems using data mining
Survey of the State-of-the-Art and Possible Extensions. IEEE
techniques. Journal of World Academy of Science,
Engineering and Technology,London,2010,64 ,pp.357–362. Transactions on knowledge and data Engineering, London 6,
JUNE 2005,VOL. 17.
[6]Yu Zheng, Lizhu Zhang, Xing Xie, and Wei-Ying Ma:
Mining interesting locations and travel sequences from GPS [14]. Kuifei Yu, Baoxian Zhang, Hengshu Zhu, Huanhuan
trajectories. In Proceedings of the 18th International Cao, and Jilei Tian Towards personalized contextaware
Conference on World Wide Web (WWW’09)ACM, New
York,2009. recommendation by mining context logs through topic models.
In Proceedings of the 16th PacificAsia conference on
[7] Jerry C.C. Tseng, Kuen-Rong Lo: An Interactive
Healthcare System with Personalized Diet and Exercise Advances in Knowledge Discovery and Data Mining -
Guideline Recommendation. International Journal of Volume Part I (PAKDD’12).Springer-Verlag, Berlin,2012
Environmental Research and Public Health (IJERPH),
2014,11(4), pp.4262-4279. pp.431–443.
15. Yu Zheng, Lizhu Zhang, Xing Xie, and Wei-Ying Ma.
[8]Wu, Chun-Hui, Fang, Kwot-Ting, and Chen, Ta-Cheng:
Mining interesting locations and travel sequences
Applying data mining for prostate cancer. In: the International
from GPS trajectories. In Proceedings of the 18th International
Conference on New Trends in Information and Service
Conference on World Wide Web (WWW’09).
Science, Beijing,2009. pp. 1063-1065.
ACM, New York, 2009.pp 791–800
[9]Lin, Y.-F., Chu, C.-H., Lin, B.-H, Yang, Y.-C., and Tseng,
Vincent A Framework for Personalized Diet and Exercise
Guideline Recommendation. The Conference on Technologies
and Applications of Artificial Intelligence (TAAI), 2014, pp.
274-283.
[10]Moon-Hee Park, Jin-Hyuk Hong, and Sung-Bae Cho.
Location-based recommendation system using bayesian user s
preference model in mobile devices. In Ubiquitous
Intelligence and Computing, 4th International Conference, UIC
Hong Kong, China, July 11-13, 2007, Vol. 4611. Springer,
1130–1139.
[11]. Alexandros Karatzoglou, Xavier Amatriain, Linas
Baltrunas, and Nuria Oliver. 2010. Multiverse
recommendation: N-dimensional tensor factorization for
context-aware collaborative filtering. In Proceedings of the 4th
ACM Conference on Recommender Systems (RecSys’10).
ACM, New York, NY, 79–86.

6|Page