Multi-objective Hierarchical Classification using Wearable Sensors in a Health Application

Abstract:

This paper introduces a novel multi-classification technique which improves two conflicting main objectives of classification problems i.e. classification accuracy and worst-case sensitivity. Global performance measures such as overall accuracy might not be enough to evaluate classifiers and alternative measurements are essentially required. This paper addresses a new model selection problem to construct a tree-based hierarchical classification model based on ensemble of six different classifiers. In our proposed approach, the model selection is tackled as a multi-objective optimization which not only considers the accuracy of the classification, but also tries to maximize the worst-case sensitivity of the multi-class problem. The proposed technique is applied on nine different classes corresponding to various breathing disorders for designing a wearable remote monitoring system. This model correctly classified the respiratory patterns of 10 subjects with accuracy of 99.25% and sensitivity of 97.78% with detecting the changes in the anterior–posterior diameter of the chest wall during breathing function by means of two accelerometer sensors worn on subject’s rib cage and abdomen. The effects of the number of sensors, sensor placement, as well as feature selection on the classification performance are also discussed.

 

Existing System:

 

The advent of mobile health technology and maturity of pervasive sensing, wireless technology as well as data processing techniques enables us to provide an effective solution for remote detection of breathing problems and promote individual’s health. Mobile health or m-health market was estimated to be valued at USD 1,950 million in 2012, with an estimated Compound Annual Growth Rate (CAGR) of 47.6% from 2014 to 2020 [2]. The previous available techniques and devices, albeit their accuracy, are expensive and could not be integrated in m-health applications. This unmet need for unobtrusive monitoring of respiration signal for the goal of respiration illnesses diagnosis has triggered research in introducing the use of wearable sensors. Therefore, such a system can help reducing the use of emergency department and hospital services resulted in increased health care team productivity.

The problem of remote diagnosis is modeled as a supervised classification problem where a respiration disorder is assigned into a predefined class based on a number of observed attributes. Once the classifier has been developed, it is used to anticipate the breathing problems that correspond to unseen samples, remotely.

 

Proposed System:

 

There exist different types of technological or hardware anomalies which may happen to occur on electronic devices. Due to decalibration or battery failures, wearable inertial sensors are subject to changes in the offset, scale factors, non-linearity or electronic noise a others [24]. Calibration, as a means of mapping raw sensor readings into the corrected values, can be used to compensate the systematic offset and gain. Note that when the gain and offset are both constant values and independent from the sensor measurements, then the calibration is translated into a linear curve-fitting function [25]. Generally, different special tools with specialists’ experience are required for sensors calibration; however, a straightforward method to calibrate an accelerometer is performed at 6 stationary positions. We need to collect a few seconds of accelerometer raw data at each position. The misalignment of the sensor in these stationary positions will influence the calibration procedure. Therefore, to minimize the impact of misalignment, two boxes with a goniometer are used to help fix the module in different positions to obtain stable acceleration measures. In our setting, we put the boxes on a flat surface and the module was placed between two boxes where it faced one box, and the other box was used to stop the module from gliding. The two boxes keep the module in a stationary position validated via the goniometer for at least 10 seconds. Then the least square method is applied to obtain 12 calibration parameters. The sensor quality and criticality of the application determine the calibration frequency that can be manually performed.

 

 

 

Conclusion:

With the integration of emerging sensor technology and analysis methods, the low-level sensor data can be translated into rich contextual information in a real-life application. In this paper, we exploited recent advances in wearable sensing and machine learning principles to provide innovative decision making capabilities for subjects’ breathing characteristics and to discern valuable information. A new multi-objective approach was developed to optimize the accuracy as well as the worst-case sensitivity in our multi-class classification problem. An evolutionary algorithm was applied which attempts to intelligently get closer and closer to the best hierarchical model. The effectiveness of the proposed technique is compared with the previously published machine learning algorithms which are stand-alone multi-class classifiers. The results showed an overall classification accuracy of 95.05%, 98.55%, and 99.25% when using K-fold with Acc1, Acc2 and both sensors data, respectively. It is concluded that the proposed method improved the worst-case sensitivity by about 9.48%, 5.09% and 2.8% compared to the best classifier in a single objective problem for the three sensor placements. However, due to large variance in per-user accuracy, the accuracy and worst-case sensitivity in average decreased 5.92% and 6.55% with subject-independent cross-validation, correspondingly. The results indicated the ability of the proposed method in designing a more robust and responsive machine learning model in the wearables. Moreover, the runtime complexity (testing step) of the proposed model was improved because there exists a binary classifier at each tree node and the number of labels to be recognized decreases as a new instance moves down in the tree.

 

Reference:

 

[1] Noam Gavriely, “Breath Sounds Methodology,” CRC Press: Florida, 1995.

 

[2] “mHealth Market Analysis And Segment Forecasts To 2020,” Grand View Research, ISBN Code: 978-1-68038-076-7, Feb. 2014.

 

[3] J. C. Fernandez, F. J. Martinez, C. Hervas, P. A. Gutierrez, “Sensitivity Versus Accuracy in Multiclass Problems Using Memetic Pareto Evolutionary Neural Networks,” Neural Networks, IEEE Transactions on, vol.21, no.5, pp.750-770, May 2010.

 

 

[4] P. De Chazal, C. Heneghan, E. Sheridan, R. Reilly, P. Nolan, and M. O’Malley, “Automated processing of the single lead electrocardiogram for the detection of obstructive sleep apnoea,” IEEE Trans. Biomed. Eng., vol.50, no.6, pp.686-696, Jun 2003.

 

[5] F. Roche, V. Pichot, E. Sforza, I. Court-Fortune, D. Duverney, F. Costes, M. Garet, and J.-C. Barth´el´emy, “Predicting sleep apnoea syndrome from heart period: A time–frequency wavelet analysis,” Eur. Respir. J., vol.22, pp.937-942, 2003.

 

[6] N. F. Garcia, P. Gomis, A. La Cruz, G. Passeriello, and F. Mora, “Bayesian hierarchical model with wavelet transform coefficients of the ECG in obstructive sleep apnea screening,” Comput. Cardiol. vol.27, pp.275-278, 2000.

 

[7] A. H. Khandoker, M. Palaniswami, C. K. Karmakar, “Support vector machines for automated recognition of obstructive sleep apnea syndrome from ECG recordings,” IEEE Trans Inf Tech Biomedicine, vol.13, pp.37-48, 2009.

 

[8] R. Lin, R. Lee, C. Tseng, H. Zhou, C. Chao, J. Jiang, “A New Approach for Identifying Sleep Apnea Syndrome Using Wavelet Transform and Neural Networks,” Biomed. Eng. Appl. Basis Commun. 2006, 18:138–143.

 

[9] B. Koley and D. Dey, “Adaptive classification system for real-time detection of apnea and hypopnea events,” in Proc. IEEE Point-of-Care Healthcare Technol., pp. 42-45, 2013.

 

[10] S. Jafari, H. Arabalibeik, K. Agin, “Classification of normal and abnormal respiration patterns using flow volume curve and neural network,” Health Informatics and Bioinformatics (HIBIT), International Symposium on, pp.110-113, April 2010.