Comparative Analysis of Pure and Hybrid Machine Learning Algorithms for Risk Prediction of Diabetes Mellitus

Author Name:
Author Email:

Abstract

Diabetes, a chronic disease, occurs due to abnormal levels of glucose and insulin in our bodies. The major factors leading to the disease are lifestyle factors such as diet, insufficient physical activity, increased stress levels and obesity. One of the major issue with diabetes is the mildness of its symptoms which delays its diagnosis until the disease has significantly progressed leading to other medical complications like blindness and nerve and kidney damage. This paper compares the efficiency of various machine learning algorithms for predicting the onset of diabetes. The performance analysis is demonstrated on Support Vector Machines (SVM), Deep Neural Networks and Hybrid Deep Learning. The study shows that the classification accuracy is significantly higher for hybrid deep neural networks indicating its better performance. The results also justify that hybrid deep learning has better processing capability and produces accurate results, thus helping in the prediction of many diseases.

Introduction

Deep learning is an unsupervised learning algorithm which mimics the human brain. The deep network has the capacity to process large datasets, implement complex functions as well as it can work with unlabeled data with small human inputs (training set) and generates the convoluted representation of unprocessed data. For enhanced performance, it has been suggested to combine supervised learning algorithm (SVM) with the deep neural networks. The unsupervised nature combined with complex learning algorithms ensures fast processing and accurate predictions for multi-dimensional data.

 

Diabetes is a chronic disease which in general does not have any cure except for very few rare cases. According to World Health Organization (WHO), the number of people suffering from diabetes is very high and is expected to increase due to the high prevalence of obesity and sedentary lifestyle. Diabetes leads to quite grave medical complications and sometimes death also. Diabetes occurs due to inefficient use or improper production of insulin by the pancreas. There are majorly three types of diabetes. Type-1 or juvenile-onset diabetes occurs due to insulin deficiency which is caused when the immune system attacks cells responsible for the production of insulin. Type-II or adult-onset diabetes is usually characterized by resistance to insulin. Gestational diabetes occurs during pregnancy due to high insulin levels which may or may not improve after delivery.

Diagnosis of diabetes and its type based on various symptoms is often time-consuming. Hence, there is a need for some automated process which will provide help in detecting the presence and progression of diabetes. And moreover predicting the onset of this disease by studying and analyzing the huge amounts of medical data available by applying various machine learning algorithms will immensely help the medical society and also assist in eradicating challenges faced by medical researchers. In this paper we use a hybrid deep learning model based on deep learning and support vector machine that predicts the presence Diabetes Mellitus. The model is trained using SVM followed by Deep Learning. We are effectively replacing the first layer of the deep learning architecture with a linear SVM layer. Further, this article also discusses the progression of the disease based on the standards set by the American Diabetes Association.

Conclusion

It can be concluded from the study that hybrid deep learning provides the most satisfactory results for prediction of diabetes. Least error rate and highest area under the ROC curve, accuracy and precision values provide evidence of better performance as compared to pure SVM and pure Deep Learning models. But even with extensive training, there is a chance of misdiagnosis as the model is not 100% accurate. For future work, the performance of the decision support system can be optimized by training with more extensive data. Newer data cleansing techniques like using Regression Substitution for handling missing values can be used for better accuracy. Furthermore, the current data only focuses on the female population and we intend to explore the trends in the male population of the society as well.

References

  1. Chen, X. W., & Lin, X.: Big data deep learning: challenges and perspectives, pp. 514— 525. IEEE Access (2015)
  2. Wang, L., & Sng, D.: Deep Learning Algorithms with Applications to Video Analytics for A Smart City: A Survey (2015)
  3. Du, T., & Shanker, V. K.: Deep Learning for Natural Language Processing
  4. Wang, Z.: The Applications of Deep Learning on Traffic Identification. BlackHat USA (2015)
  5. Han, Jianchao, Juan C. Rodriguez, and Mohsen Beheshti.: Diabetes data analysis and prediction model discovery using rapidminer. In: Second International Conference on Future Generation Communication and Networking, FGCN’08. vol. 3. IEEE(2008)
  6. Songthung, P., & Sripanidkulchai, K.: Improving type 2 diabetes mellitus risk prediction using classification. In: 13th International Joint Conference, pp. 1—6. Computer Science and Software Engineering (JCSSE), IEEE(2016)
  7. Kamble, M. T. P., & Patil, S. T.:Diabetes Detection using Deep Learning Approach. IJIRST, vol. 2, Issue 12 (2016)
  8. UCI Machine Learning Repository, https://archive.ics.uci.edu
  9. Gordon C. Weir & Susan Bonner-Weir: Five Stages of Evolving Beta-Cell Dysfunction During Progression to Diabetes (2004).

DOWNLOAD THE COMPLETE RESEARCH PAPER PDF

30 total views, 0 views today

About the author: tej