Elsevier

Computers in Industry

Volume 97, May 2018, Pages 54-66
Computers in Industry

Data accuracy aware mobile healthcare applications

https://doi.org/10.1016/j.compind.2018.01.020Get rights and content

Highlights

  • Medical data accuracy challenges for mobile healthcare applications are studied.

  • An anomaly detection and isolation approach is proposed.

  • Multivariate anomaly detection is preceded by principal component analysis based dimension reduction step.

  • Univariate anomaly isolation is provided to distinguish between inaccurate data and patient health degradation.

  • The proposed approach is evaluated on real medical dataset and compared with existing solutions.

Abstract

This paper proposes a new approach for online detection and isolation of inaccurate vital sign measurements in mobile healthcare applications. Our primary objective is to distinguish between inaccurate measurement and patient health degradation to reduce false medical alarms. The proposed approach couples dimensionality reduction with inaccurate data detection and isolation. On the one hand, dimension reduction is based on robust incremental Principal Component Analysis. On the other hand, multivariate anomaly detection relies on squared prediction errors and anomalous vital sign isolation is based on contribution plots. We apply our proposed approach on real medical dataset. Our simulation results prove the effectiveness of our approach in achieving good recall with a low false alarm rate compared with existing solutions. The benefit gained by our approach in terms of time and space complexities make it useful and efficient for real time mobile healthcare applications.

Introduction

Mobile health [[1], [2]] can be broadly defined as the use of information and communication technologies that are accessible to people or healthcare providers through mobile devices such as mobile phones, patient monitoring devices other wireless devices to support medical and public healthcare practices. mHealth plays a significant role in delivering remote healthcare services while empowering people to better manage their own health state and improve the ability to track diseases. It aims at reducing socio-economic costs [3] and even surmount organizational barriers [2].

Mobile healthcare systems are rapidly growing and evolving. The architecture of such system contains three domains. In the device domain, a patient is equipped with a set of wearable and/or implantable sensors that constantly sense vital signs. Sensors are interconnected via a medical body area network. The central point of the device domain is the application hosting device. It aggregates vital sign measurements and transmits them via the network domain to the application domain. Mobile health applications cover several areas to include disease management (such as chronic disease management [4], women's health and pregnancy [5] and elderly people monitoring [6]), as well as, wellness management [7].

Particularly, mobile health applications highly depend on continuous sensor readings to provide high-quality healthcare services and improve the quality of patient supervision [8]. However, sensor readings introduce critical accuracy challenges [9] due to several internal and external factors. Internal factors are related to the performance variation of sensors involved in the application such as energy depletion, hardware faults, calibration, etc., and other external environmental-related ones.

To achieve its full potential, a mHealth application needs to deal with medical data accuracy challenges.

In the last few years, there has been a growing interest in proposing approaches dealing with anomaly detection and isolation for mhealth applications. The proposed approaches are divided into two main classes including supervised [[10], [11], [12]] and unsupervised approaches [[13], [14], [15], [16], [17]].

Despite the tremendous research effort around approaches for anomaly detection and isolation in mHealth applications, there are still several issues preventing their deployment in real time healthcare systems. Classification-based supervised approaches [[10], [12]] rely on trustable training dataset. However, this assumption may not be conventional in real time mHealth applications, where anomalies are not known in advance. Furthermore, training data often contains abnormal measurements which lead to overfitted classification or regression models. A majority of existing unsupervised approaches [[14], [13], [16]] cannot be applicable in high-dimensional data and can generate additional complexity in terms of time and space. Indeed, in real time mHealth applications, wearable devices such as smart vests [18] and other implantable sensors like Implantable Cardioverter Defibrillator [19] increase the amount of multi-dimensional medical data by being able to send vital signs at any time and any place.

In this paper, we address these issues by proposing an unsupervised anomaly detection and isolation approach for mHealth applications. The proposed approach combines two key characteristics. First, it couples a dimension reduction algorithm based on the spatial and temporal correlation among monitored vital signs with anomaly detection and isolation methods. Second, it uses lightweight and robust unsupervised techniques in both steps. Dimension reduction is based on robust incremental principal component analysis which is not sensitive to anomalies and precludes the necessity of training using a trustable dataset, a strong advantage from the operational point of view. Multivariate anomaly detection is based on squared prediction errors and adequate threshold. Once multivariate anomaly is noticed, an isolation step is carried out using contribution plots [20] to identify suspicious vital signs.

Simulations in real medical record are firstly conducted to choose the suitable principal component analysis for dimensionality reduction. Second, our simulation results prove that our proposed approach achieves good detection performances and demonstrate the importance of dimension reduction in reducing computational time and memory. Finally, they validate that our approach outperforms other related approaches [[11], [13]].

The rest of the paper is organized into four sections: Section 2 outlines related work. The proposed approach is discussed in Section 3. Section 4 shows the experimental results of this work. Finally, Section 5 includes a summary and gives some directions for the future.

Section snippets

Literature review

Various mobile healthcare applications for periodic vital signs monitoring have been proposed, such as CodeBlue [21], HealthGear [22], SmartVest [18], MEDiSN [23], PIaIMoS [24], and so on. Their main features are summarized in Table 1.

Recent surveys of mobile healthcare applications are available in [[24], [25]].

Unfortunately, real-time vital sign measurements often contain inconsistencies and errors due to diverse internal and external factors. The internal factors are, especially, related to

Proposed approach

In real-time mobile health scenario, a patient performs his own vital sign measurements (respiration rate, pulse rate, blood pressure, SpO2, etc.) using M implantable and wearable medical sensors such as Pulse oximeter (as depicted in Fig. 1). Each sensor is used to sense one or many vital signs and transmit data to an Application Hosting Device (e.g. Smartphone) via Bluetooth technology. The application hosting device allows monitoring medical data from sensors and transferring it via internet

Simulations

In order to evaluate the proposed approach, we achieve simulations. In the first part, we begin by demonstrating the importance of robustness on PCA during the offline regime. Then, we perform a comparative study between two online PCAs to choice the suitable algorithm for updating model in the online regime. The second part provides an explanation of how our monitor module succeeds in detecting multivariate anomalies and identifying adequate symptoms. We denote the third part to compare the

Conclusion and future research

This paper proposes a lightweight anomaly detection and isolation approach for mobile healthcare applications, where anomalous vital signs measurements have a significant impact on the effectiveness of such applications and consequently affects the patient's life. The proposed approach is based on Principal Component Analysis for dimension reduction, squared prediction errors for multivariate anomaly detection and contribution plots for anomaly isolation. It is suitable for real time mobile

References (44)

  • J. McNaull et al.

    Data and information quality issues in ambient assisted living systems

    J. Data Inform. Qual.

    (2012)
  • H. Aloulou et al.

    Deployment of assistive living technology in a nursing home environment: methods and lessons learned

    BMC Med. Inform. Decis. Mak.

    (2013)
  • O. Salem et al.

    Anomaly detection in medical wireless sensor networks using SVM and linear regression models

    IJEHMC

    (2014)
  • H.S. Ahsanul et al.

    Sensor anomaly detection in wireless sensor networks for healthcare

    Sensors

    (2015)
  • O. Salem et al.

    Sensor fault and patient anomaly detection and classification in medical wireless sensor networks

    IEEE International Conference on Communications

    (2013)
  • O. Salem et al.

    Anomaly detection in medical wireless sensor networks

    JCSE

    (2013)
  • A. Naseem et al.

    Reliable vital sign collection in medical wireless sensor networks

    15th IEEE International Conference on e-Health Networking, Applications and Services

    (2013)
  • O. Salem et al.

    Online anomaly detection in wireless body area networks for reliable healthcare monitoring

    IEEE J. Biomed. Health Inform.

    (2014)
  • O. Salem et al.

    Anomaly detection in medical WSNS using enclosing ellipse and chi-square distance

    IEEE International Conference on Communications

    (2014)
  • V. Deneshkumar et al.

    Identification of outliers in medical diagnostic system using data mining techniques

    Int. J. Stat. Appl.

    (2014)
  • X. Fu et al.

    A wireless implantable sensor network system for in vivo monitoring of physiological signals

    IEEE Trans. Inform. Technol. Biomed.

    (2011)
  • P.V. den Kerkhof et al.

    Contribution plots for statistical process control: analysis of the smearing-out effect

    2013 European Control Conference

    (2013)
  • Cited by (13)

    View all citing articles on Scopus
    View full text