Добавил:
Sekretar
kiopkiopkiop18@yandex.ru
t.me/Prokururor I Вовсе не секретарь, но почту проверяю
Опубликованный материал нарушает ваши авторские права? Сообщите нам.
Вуз:
Предмет:
Файл:Ординатура / Хирургия / Библиотека им академика М.И. Перельмана / Книга_145_библиотеки_им_акад_М_И_Перельмана
.pdf
A Clinical Decision Support System Disease Prediction using Machine Learning 93
https://t.me/med1917
ing the first stage are effectively used as a significant input in the subsequent age
[3]. A significant fact contributing to the improvement in the prediction of class
outcomes is the use of clustering information during the process of classification
[4]. This is seen as a requisite to the system as asthma is seen to exhibit
polymorphic phenotypes that have their own characterization [5].
Identification of patient populations at high risk is an important intervention in the
early detection and clinical assessment of chronic diseases like asthma and
bronchitis, as it can lead to targeted and personalized therapies. Our research has a
credible potential to be an important intervention through facilitating early risk
assessment and effective prognosis of the disease via risk stratification. In case of
failure to recognize the disease at an early stage, the disease could be progressive
and finally become irreversible, though reversible in nature. This necessitates us
to propose approaches that help us in the prediction of the disease using severity
indicators that can be possibly identified at an early stage, enabling early medical
interventions that alleviate the disease severity, subsequently lowering the
mortality rate. The disease is further heterogeneous, in that it presents itself with
variable overlapping syndromes of other obstructive diseases, which makes it
difficult to delineate the symptoms characterizing asthma.
RELATED WORK
Clinical Decision Support systems incorporating machine learning techniques to
mine the useful data may be adopted to explore asthma data, and to identify latent
patterns that add value, while providing a supplementary source of better
understanding for decision-making with respect to risk factor identification along
with the assessment of the severity of the disease and the extent to which it relates
to other diseases of the lungs such as COPD (Chronic Obstructive Pulmonary
Diseases) [6 - 8].
Bayesian classification approaches implementing supervised learning techniques
have resulted in classes that are principally identified as correct and have a good
amount of face validity along with substantial relationships to asthma, variations
in airway reactivity and lung function, which illustrate content validity with a
good amount of gratification [9, 10]. Predictive analytics have been suitably
deployed to estimate risk factors by exploring a blend of independent variables
from variable sources [11]. Rule-based systems integrating expert-driven feature
selection and rough sets techniques have been identified to describe asthmatic
populations visiting emergency departments [12]. A good amount of precision
was obtained with the task associated with the differentiation of asthma against
other respiratory disorders depending on the nature of the underlying data and the
ensemble learning techniques used [13].
94 Disease Prediction using Machine Learning M.R. Pooja
https://t.me/med1917
MATERIAL AND METHODS
Dataset Description
The response data for the ISAAC core questionnaires gathered under the first
phase, a study that has been used across several pilot studies, was used to validate
the system performance [14]. The questions used in the collection of data
typically signify sensitive and specific indicators of the disease. The data are
cross-sectional in that it involves the analysis of data concerning a specific
population which is widely a significant subset that is utmost representative at a
specific point in time. The data was gathered across two age groups, 13 to 14
years and 6 to 7 years. While in the older group of children, the response was
completed by the children themselves, and the same was completed by the parents
in the other group.
Combatting Class Imbalance
The New Delhi dataset contained the details of 2961 subjects explained with
respect to 47 variables indicative of asthma comorbidities and symptoms. The
disease outcome was reported as unknown for 112 samples. Hence, we excluded
all 112 samples yielding a dataset containing 2849 samples. However, only 110
subjects out of the total population were asthmatics and hence we sought to
combat the imbalance by performing stratified sampling that was able to draw 110
samples from the other category too (non-asthmatics), thus ending up with a
balanced dataset that was fed as the input data. Along similar lines, we worked to
draw 2% amounting to 64 subjects from the non-asthmatics from the total
population to combat the class imbalance that would have resulted in 78
asthmatics in the population, resulting in an input dataset containing 142 subjects.
Feature Clustering
Initially, the predominant features that can be identified as significant
comorbidities of asthma disease are chosen by applying feature clustering. Feature
clustering is performed by employing Modified Fuzzy C means clustering
(MFCM), which uses an objective function based on correlation, unlike the
traditional Fuzzy C Means clustering which uses an objective function based on
distances [10, 11]. We chose to organize the input feature set into four clusters as
per the evaluation results of subtractive clustering. The features that coexist with
asthma attributes are extracted from the cluster containing asthma. This leads to a
reduced feature set in the process.
A correlation-based objective function is used in the fuzzy c means clustering.
The distance metric here incorporates a Pearson’s correlation coefficient that is
A Clinical Decision Support System Disease Prediction using Machine Learning 95
https://t.me/med1917
indicative of the difference between two given features or between a feature and a
cluster. Features may correlate directly or inversely indicating a positive or
negative correlation. Correlated features will be clustered into a common group
irrespective of their expression values. The distance metrics are defined as dij, =
1− ρ2 Xi, C j, where ρXi, C j, the correlation (Pearson measure) between feature xi
and a cluster cj. The correlation would be 0 in case of a high correlation, either
positive or negative.
Subject Clustering
For a disease such as asthma, which is heterogeneous in nature, subject clustering
has been a preferred approach to disambiguate the clinical differences between
subjects [17]. We opted to use K means clustering and the choice of the number
of clusters indicating the parameter “K” was varied from 2 to 8, for which
silhouette scores were obtained. Silhouette score was high for two clusters, with
the naive data and was high for eight clusters with New Delhi data. The
distribution patterns of asthmatics and non-asthmatics in the different clusters are
well illustrated in Fig. (1a and b). The subjects in the two datasets were organized
into two and eight clusters with the reduced feature set as shown in Fig. (2a and
b) respectively. Following this, the prediction of subjects with and without asthma
was done by applying classification and regression trees with the addition of a
feature that indicated cluster evaluation score. The distribution patterns of
asthmatics and non-asthmatics in the different clusters are well illustrated in Fig.
(2).
1.0
2.0
90
30
Frequency
20
10
0
C1 C2 C3 C4 C5 C6 C7 C8
Cluster
30
20
Frequency
10
0
a) New Delhi
Fig. (1). Distribution of subjects within the two groups.
C1 C2
Cluster
(b) Neyveli
1.0
2.0
96 Disease Prediction using Machine Learning M.R. Pooja
https://t.me/med1917
0 0.2 0.4 0.6 0.8
C8 (7) C7 (6) C6 (11) CS (13) C4 (20) C3 (15) C2 (47) C1 (10)
0 0.2 0.4 0.6 0.8
0 0.2 0.4 0.6 0.70.3 0.50.1
C1 (64)C2 (23)
0 0.2 0.4 0.6 0.70.3 0.50.1
(a) New Delhi (b) Neyveli
Fig. (2). Silhouette plots for the two datasets (a) New Delhi, (b) Neyveli.
Both random initialization and K means++ were used to initialize the cluster
centroids for K-means clustering in order to ascertain a strong basis for the choice
of the number of clusters as shown in Tables 1 and 2. It was seen that both the
initialization methods yielded the same result signifying a consensus on the choice
of two and eight clusters in Naively and New Delhi data respectively. A value of
0.325 was obtained as the silhouette score for Naively data via both the methods
and silhouette scores of 0.384 and 0.377 (highest) in the case of random
initialization and K-means++ methods for New Delhi data.
Table 1. Cluster evaluation for rural asthma.
Initialization Method Clusters Silhouette Score
Dataset Naively data
- 2 0.325
3 0.236
4 0.217
Random
initialization
5 0.220
6 0.194
7 0.204
8 0.217
A Clinical Decision Support System Disease Prediction using Machine Learning 97
https://t.me/med1917
(Table 1) co nt.....
Initialization Method Clusters Silhouette Score
2 0.325
3 0.236
4 0.217
K-Means++
Table 2. Cluster evaluation for urban asthma.
Dataset New Delhi data
2 0.384
3 0.376
4 0.409
Random
initialization
K-Means++
5 0.422
6 0.442
7
8
2 0.384
3 0.377
4 0.405
5 0.414
6
7 0.452
8 0.461
5 0.216
6
7 0.219
8 0.232
0.206
0.441
0.464
0.446
Performance Evaluation
The performance of the hybrid decision system evaluated using 10-fold crossvalidation, Stratified random sampling, and Random Sampling is illustrated in
Table 3. A comparative analysis of the performance of other classifiers is also
illustrated in Tables 4 and 5 respectively. It is observed that the performance of
the hybrid decision system outweighs that of the other classifiers.
98 Disease Prediction using Machine Learning M.R. Pooja
https://t.me/med1917
Table 3. Performance evaluation of a hybrid system.
Dataset Method AUC CA F1 Precision Recall
10-fold CV 0.855 0.851 0.851 0.854 0.851
Neyveli
New
Delhi
Table 4. Comparative performance evaluation of traditional classifiers for Neyveli data.
Random Forest 0.903 0.796 0.796 0.799 0.796
Logistic Regression
Random Forest 0.865 0.772 0.773 0.778 0.772
Logistic Regression 0.896 0.793 0.794 0.795 0.793
Random Forest 0.834 0.734 0.735 0.736 0.734
Logistic Regression 0.877 0.755 0.755 0.755 0.755
Stratified RS * 0.844 0.838 0.839 0.840 0.838
RS* 0.845 0.834 0.835 0.835 0.834
10-fold CV
Stratified RS* 0.972 0.945 0.945 0.945 0.945
RS* 0.968 0.936 0.936 0.936 0.936
Classifier AUC
KNN 0.842 0.754 0.750 0.756 0.754
Tree 0.842 0.817 0.817 0.826 0.817
SVM Learner 0.865 0.761 0.745 0.802 0.761
Naive Bayes 0.884
OGB 0.918 0.845 0.845 0.845 0.845
Stratified Random sampling with Shuffled split for 80% train data
KNN 0.830 0.769 0.768 0.769 0.769
Tree
SVM Learner
Naive Bayes 0.856 0.779 0.780 0.786 0.779
Random sampling with Shuffled split for 80% train data
KNN 0.811 0.741 0.739 0.742 0.741
Tree 0.806 0.790 0.789 0.790 0.790
SVM Learner 0.851 0.741 0.721 0.794 0.741
Naive Bayes 0.839 0.752 0.752 0.757 0.752
OGB 0.889 0.790 0.790 0.790 0.790
0.977
10-fold Stratified Cross-validation
0.901
0.784 0.779 0.780 0.783 0.779
0.869
0.953
CA F1 Precision Recall
0.789
0.810
0.766
0.953
0.789
0.810
0.753
0.954
0.797
0.810
0.798
0.789
0.810
0.766
0.953
A Clinical Decision Support System Disease Prediction using Machine Learning 99
https://t.me/med1917
Table 5. Comparative perfromance evaluation of traditional classifiers for New Delhi data.
10-fold Stratified Cross validation
Classifier AUC CA F1 Precision Recall
KNN 0.835 0.814 0.813 0.815 0.814
Tree 0.862 0.845 0.845 0.846 0.845
SVM Learner
Random Forest 0.926 0.855 0.854 0.856 0.855
Naive Bayes 0.909 0.777 0.777 0.778 0.777
Logistic Regression 0.908 0.805 0.804 0.809 0.805
OGB 0.964 0.914 0.914 0.914 0.914
Stratified Random sampling with Shuffled split for 80% train data
KNN 0.798
Tree 0.829 0.798 0.798 0.798 0.798
SVM Learner 0.879 0.768 0.756 0.835 0.768
Random Forest 0.927 0.852 0.852 0.853 0.852
Naive Bayes 0.874
Logistic Regression
OGB 0.953 0.895 0.895 0.896 0.895
Random sampling with Shuffled split for 80% train data
KNN 0.832 0.814 0.814 0.815 0.814
Tree
SVM Learner
Random Forest 0.931 0.866 0.866 0.866 0.866
Naive Bayes 0.877 0.748 0.748 0.750 0.748
Logistic Regression 0.883 0.764 0.763 0.771 0.764
OGB
0.912
0.885
0.808 0.820 0.820 0.821 0.820
0.860
0.959 0.886 0.886 0.888 0.886
0.786
0.786 0.786 0.788 0.786
0.748
0.777
0.764
0.779
0.748
0.777
0.748
0.833
0.748
0.779
0.830
0.786
0.748
0.777
0.764
The classification performance was evaluated by two testing schemes, namely
stratified random sampling and 10-fold cross-validation which are briefly
discussed below.
Cross-validation is a method adopted for the evaluation of prediction models by
separating the original sample into a training set in order to train the model, and a
test set, to evaluate its performance. In 10-fold cross-validation, the given input
sample is comprehensively partitioned into 10 equal-size subsamples. Of the “k”
subsamples, a single sample is reserved as part of the validation set for testing the
100 Disease Prediction using Machine Learning M.R. Pooja
https://t.me/med1917
model, and the residual 9 samples are used for training the model. The process is
iterated ten times. The results from these 10 folds can then be averaged thereby
producing a single estimation. Random sampling with stratification is thus a
sampling option following a probabilistic approach.
The technique chosen is used in place of simple random sampling in scenarios
where, the groupings of the strata are assumed to be too dissimilar and too
imperative to the problem under study, and/or when investigators desire to
oversample a principally minor group of interest. Here, the entire dataset is
separated into a training set and a test set. 80% of the samples are introduced into
the training set while the rest 20% is organized into a test set and the process is
repeated 100 times to validate the results.
CONCLUSION
The hybrid decision support system was quite effective in making accurate
predictions concerning the asthma disease outcome, by achieving sensitivity as
high as 85.1% and 95.3% for the two datasets under consideration in our study.
Further, precision was also satisfactory with 85.4% and
95.4% for the two datasets. It is quite evident from the performance evaluation
that the results of the developed hybrid decision support system are always
promising for suitable deployment in real-time scenario, as it explores the benefits
of both supervised and unsupervised learning techniques. Further, using clustering
information in the form of cluster evaluation scores as an input parameter to the
classifiers can efficiently predict disease outcomes, especially with diseases such
as asthma, as the disease is heterogeneous and exhibits several disease subtypes.
REFERENCES
[1] S. Ram, W. Zhang, M. Williams, and Y. Pengetnze, "Predicting asthma-related emergency department
visits using big data", IEEE J. Biomed. Health Inform., vol. 19, no. 4, pp. 1216-1223, 2015.
[http://dx.doi.org/10.1109/JBHI.2015.2404829] [PMID: 25706935]
[2] C.H. Lee, J.C.Y. Chen, and V.S. Tseng, "A novel data mining mechanism considering bio-signal and
environmental data with applications on asthma monitoring", Comput. Methods Programs Biomed.,
vol. 101, no. 1, pp. 44-61, 2011.
[http://dx.doi.org/10.1016/j.cmpb.2010.04.016] [PMID: 20554074]
[3] M.R. Pooja, and M.P. Pushpalatha, "A hybrid decision support system for the identification of
asthmatic subjects in a cross-sectional study", 2015 International Conference on Emerging Research
in Electronics, Computer Science and Technology (ICERECT)., pp. 288-293, 2015.
[http://dx.doi.org/10.1109/ERECT.2015.7499028]
[4] V. Siroux, X. Basagaña, A. Boudier, I. Pin, J. Garcia-Aymerich, A. Vesin, R. Slama, D. Jarvis, J.M.
Anto, F. Kauffmann, and J. Sunyer, "Identifying adult asthma phenotypes using a clustering
approach", Eur. Respir. J., vol. 38, no. 2, pp. 310-317, 2011.
[http://dx.doi.org/10.1183/09031936.00120810] [PMID: 21233270]
[5] W.C. Moore, D.A. Meyers, S.E. Wenzel, W.G. Teague, H. Li, X. Li, R. D’Agostino Jr, M. Castro, D.
A Clinical Decision Support System Disease Prediction using Machine Learning 101
https://t.me/med1917
Curran-Everett, A.M. Fitzpatrick, B. Gaston, N.N. Jarjour, R. Sorkness, W.J. Calhoun, K.F. Chung,
S.A.A. Comhair, R.A. Dweik, E. Israel, S.P. Peters, W.W. Busse, S.C. Erzurum, and E.R. Bleecker,
"Identification of asthma phenotypes using cluster analysis in the severe asthma research program",
Am. J. Respir. Crit. Care Med., vol. 181, no. 4, pp. 315-323, 2010.
[http://dx.doi.org/10.1164/rccm.200906-0896OC] [PMID: 19892860]
[6] S. Schmidt, G. Li, and Y.-P. P. Chen, "Medical knowledge discovery from a regional asthma dataset",
In: Advanced Intelligent Computing Theories and Applications. With Aspects of Artificial Intelligence.
ICIC vol. 5227. Huang, DS., Wunsch, D.C., Levine, D.S., Jo, KH Springer, Berlin, Heidelberg., 2008.
[http://dx.doi.org/10.1007/978-3-540-85984-0_107]
[7] K. Farion, W. Michalowski, and S. Wilk, Developing a Decision Model for Asthma Exacerbations:
Combining Rough Sets and Expert-Driven Selection of Clinical Attributes. Rough Sets and Current
Trends in Computing, 2006, pp. 428-437.
[http://dx.doi.org/10.1007/11908029_45]
[8] M.R. Pooja, and M.P. Pushpalatha, "Cluster analysis to characterize the patterns of complementary
and alternative medicines usage in asthma controls", Open Public Health J., vol. 13, p. 1, 2020.
[http://dx.doi.org/10.2174/1874944502013010227]
[9] W. Wu, E. Bleecker, W. Moore, W.W. Busse, M. Castro, K.F. Chung, W.J. Calhoun, S. Erzurum, B.
Gaston, E. Israel, D. Curran-Everett, and S.E. Wenzel, "Unsupervised phenotyping of severe asthma
research program participants using expanded lung data", J. Allergy Clin. Immunol., vol. 133, no. 5,
pp. 1280-1288, 2014.
[http://dx.doi.org/10.1016/j.jaci.2013.11.042] [PMID: 24589344]
[10] M. Zedan, G. Attia, M.M. Zedan, A. Osman, N. Abo-Elkheir, N. Maysara, T. Barakat, and N. Gamil,
"Clinical asthma phenotypes and therapeutic responses", ISRN Pediatr., vol. 2013, pp. 1-7, 2013.
[http://dx.doi.org/10.1155/2013/824781] [PMID: 23606983]
[11] M.P. Pushpalatha, and M.R. Pooja, "A predictive model for the effective prognosis of Asthma using
Asthma severity indicators", 2017 International Conference on Computer Communication and
Informatics (ICCCI), 2017pp. 1-6
[http://dx.doi.org/10.1109/ICCCI.2017.8117717]
[12] C. Chakraborty, T. Mitra, A. Mukherjee, and A.K. Ray, "CAIDSA: Computer-aided intelligent
diagnostic system for bronchial asthma", Expert Syst. Appl., vol. 36, no. 3, pp. 4958-4966, 2009.
[http://dx.doi.org/10.1016/j.eswa.2008.06.025]
[13] M.R. Pooja, and M.P. Pushpalatha, "A comparative performance evaluation of hybrid and ensemble
machine learning models for prediction of asthma morbidity", J. Health Med. Inform., vol. 10, no. 330,
2019.
[http://dx.doi.org/10.4172/2157-7420.1000330]
[14] C.K.W. Lai, R. Beasley, J. Crane, S. Foliaki, J. Shah, and S. Weiland, "Global variation in the
prevalence and severity of asthma symptoms: Phase three of the international study of asthma and
allergies in childhood (ISAAC)", Thorax, vol. 64, no. 6, pp. 476-483, 2009.
[http://dx.doi.org/10.1136/thx.2008.106609] [PMID: 19237391]
[15] W.C. Moore, D.A. Meyers, S.E. Wenzel, W.G. Teague, H. Li, X. Li, R. D’Agostino Jr, M. Castro, D.
Curran-Everett, A.M. Fitzpatrick, B. Gaston, N.N. Jarjour, R. Sorkness, W.J. Calhoun, K.F. Chung,
S.A.A. Comhair, R.A. Dweik, E. Israel, S.P. Peters, W.W. Busse, S.C. Erzurum, and E.R. Bleecker,
"Identification of asthma phenotypes using cluster analysis in the severe asthma research program",
Am. J. Respir. Crit. Care Med., vol. 181, no. 4, pp. 315-323, 2010.
[http://dx.doi.org/10.1164/rccm.200906-0896OC] [PMID: 19892860]
[16] M. Zolnoori, M. H. F. Zarandi, and M. Moin, "Application of intelligent systems in asthma disease:
Designing a fuzzy rule-based system for evaluating level of asthma exacerbation", J Med Syst., vol.
36, no. 4, pp. 2071-2083, 2012.
[http://dx.doi.org/10.1007/s10916-011-9671-8]
102 Disease Prediction using Machine Learning M.R. Pooja
https://t.me/med1917
[17] E.M.S. Mäkikyrö, M.S. Jaakkola, and J.J.K. Jaakkola, "Subtypes of asthma based on asthma control
and severity: A latent class analysis", Respir. Res., vol. 18, no. 1, pp. 24-32, 2017.
[http://dx.doi.org/10.1186/s12931-017-0508-y] [PMID: 28114991]
Соседние файлы в папке Библиотека им академика М.И. Перельмана