Elsevier

Computers in Biology and Medicine

Volume 59, 1 April 2015, Pages 116-124
Computers in Biology and Medicine

An efficient machine learning approach for diagnosis of paraquat-poisoned patients

https://doi.org/10.1016/j.compbiomed.2015.02.003Get rights and content

highlights

  • A novel method is presented for diagnosis of paraquat (PQ) poisoning.

  • The key parameters involved in ELM have been investigated in detail.

  • The effectiveness of the proposed method has been rigorously evaluated.

  • The approach has detected the most influential feature account for PQ poisoning.

Abstract

Numerous people die of paraquat (PQ) poisoning because they were not diagnosed and treated promptly at an early stage. Till now, determination of PQ levels in blood or urine is still the only way to confirm the PQ poisoning. In order to develop a new diagnostic method, the potential of machine learning technique was explored in this study. A newly developed classification technique, extreme learning machine (ELM), was taken to discriminate the PQ-poisoned patients from the healthy controls. 15 PQ-poisoned patients recruited from The First Affiliated Hospital of Wenzhou Medical University who had a history of direct contact with PQ and 16 healthy volunteers were involved in the study. The ELM method is examined based on the metabolites of blood samples determined by gas chromatography coupled with mass spectrometry in terms of classification accuracy, sensitivity, specificity and AUC (area under the receiver operating characteristic (ROC) curve) criterion, respectively. Additionally, the feature selection was also investigated to further boost the performance of ELM and the most influential feature was detected. The experimental results demonstrate that the proposed approach can be regarded as a success with the excellent classification accuracy, AUC, sensitivity and specificity of 91.64%, 0.9156%, 91.33% and 91.78%, respectively. Promisingly, the proposed method might serve as a new candidate of powerful tools for diagnosis of PQ-poisoned patients with excellent performance.

Introduction

Paraquat (1,1′-dimethyl-4,4′-bipyridium dichloride, PQ), the most widely used herbicides in the world, has been deemed as the most highly toxic pesticide for human [1]. Its mortality rate is highly correlated with plasma PQ concentrations [2]. Acute ingesting 7–8 mL PQ can cause serious symptoms such as liver, lung, kidney, and heart failure that directly leads to death if without getting prompt treatment [3]. Although few of PQ poisoning appeared in developed countries, there are thousands of people died of PQ poisoning every year in developing countries [4]. For example, PQ accounts for most fatal poisonings, with 500 or more deaths per year in Korea [5].

PQ intoxication is associated with reactive oxygen species and free radicals that cause early multiorgan failure and late pulmonary fibrosis with respiratory failure [5], [6]. The current treatment strategies of PQ poisoning are increasing the elimination of PQ from the body, administration of antioxidants and the maintenance of vital functions, which is an entirely different from other intervention of intoxication [1], [7]. And earlier the treatment initiated, more effective in reducing mortality, particularly hemoperfusion (HP) within 2–4 h after intoxication [4]. Therefore, diagnosis is very important in treatment of PQ-poisoned patients.

Till now, the diagnosis of PQ poisoning mainly is according to the PQ concentration in blood. However, PQ is absorbed poorly from the stomach and small intestine (< 5%) and distributed into all organs in the body within 5 h, which means it is difficult to detect PQ in blood after poisoned 5 h [1], [8]. And in some cases, the patients cannot provide a clear contact history of PQ poisoning, such as disturbance of consciousness or language. This poses a more serious problem for diagnosis. How to develop a new diagnosis method of PQ poisoning is becoming an important topic in medicine.

In this study, patients with acute PQ intoxication were involved and determined by gas chromatography coupled with mass spectrometry (GC–MS). According to their plasma metabonomics, a rapid diagnosis method was developed based on the extreme learning machine (ELM) technique [9], a new learning algorithm for a single hidden layer feed-forward neural network (SLFNs). Different from the common parameter tuning strategy of neural network, ELM tries to choose input weights and hidden biases randomly, and the output weights are analytically determined by using Moore–Penrose (MP) generalized inverse. It not only learns much faster with higher generalization performance, but also keeping very few parameters for tuning. Thanks to its good properties, ELM has found its applications in a wide range of fields such as cancer diagnosis [10], image quality assessment [11], face recognition [12], land cover classification [13] and hyperspectral images classification [14]. To the best of our knowledge, there is no research work dealing with the problem of PQ poisoning from the machine learning perspective. Therefore, an attempt was made in this study to explore the potential of the performance of ELM in discriminating the PQ-poisoned patients from the healthy controls. For comparison purpose, the support vector machines [15] (SVM) was also taken for diagnosis of PQ poisoning. In addition, the effectiveness of feature selection was investigated as well. The efficient and commonly used feature selection method, maximum relevance minimum redundancy (mRMR) [16], was employed for pre-processing before the classification models were constructed. mRMR is a filter type feature selection method that seeks to choose features which are relevant to the target class (maximum relevance) and come up with the feature subset containing as non-redundant features as possible (minimum redundancy). The effectiveness of the proposed approach is examined in terms of classification accuracy, AUC, sensitivity and specificity respectively on the diagnosis of PQ-poisoned cases whose samples were collected from The First Affiliated Hospital of Wenzhou Medical University. Promisingly, the developed ELM based approach has achieved high diagnosis accuracy, AUC, sensitivity and specificity of 91.64%, 0.9156%, 91.33% and 91.78%, respectively.

The remainder of this paper is organized as follows. Section 2 offers brief background knowledge on ELM. The detail of the ELM method is described in Section 3. Section 4 presents the detailed experimental designs. The experimental results and discussion of the proposed method are presented in Section 5. Finally, conclusions and recommendations for future work are summarized in Section 6.

Section snippets

Extreme learning machine (ELM)

A brief description of ELM is given in this section, for more details, one can refer to [9], [17]. Given a training set ={(xi,ti)|xiRn,tiRm,i=1,2,,N}, where xi is the n×1 input feature vector and ti is a m×1 target vector. The standard SLFNs with an activation function g(x) and N˜ hidden neurons can be mathematically modeled as follows [9]:i=1N˜βig(wixj+bi)=oj,j=1,2,,Nwhere wi is the weight vector between the ith neuron in the hidden layer and the input layer, bi means the bias of the ith

Proposed ELM model

In this section, we briefly described the proposed ELM method for distinguishing PQ-poisoned patients from the healthy controls. As depicted in Fig. 1, the diagnosis model is created by ELM. From the figure, the input data consist of the weights of metabolites in blood. The optimal hidden nodes in the hidden layer of ELM will be determined by the ten repetitions of 3-fold cross-validation method. When the trained model was performed for prediction, the output of ELM has two statuses, ‘1’ means

Data description

This study was approved by the Medical Ethics Committee of The First Affiliated Hospital of Wenzhou Medical University and conducted in accordance with the Declaration of Helsinki. All individual information of PQ-poisoned patients was securely protected and only available to the investigators. All data were analyzed anonymously.

There were 15 patients (aged from 18–64 years, 9 male/6 female) who had a history of direct contact with PQ poisoning were involved in this study. Their timeframe of PQ

Experimental results and discussions

Different types of activation functions may have different impact on the performance of ELM. Therefore, the first attempt was made to evaluate the influence of different activation functions on the performance of the ELM model. Five different types of activation function including Sigmoid function (sig), Sine function (sin), Hard-limit function (hardlim), Triangular basis function (tribas) and Radial basis function (radbas) were evaluated in the experiment. The relationship between the

Conclusions and future work

This paper presents a novel method for diagnosis of PQ poisoning from the machine learning perspective. The empirical experiments on the data, collected from The First Affiliated Hospital of Wenzhou Medical University, have demonstrated the excellent superiority of the proposed ELM based approach in terms of classification accuracy, AUC, sensitivity and specificity. With the aid of feature selection, we have identified the most crucial feature for PQ poisoning diagnosis. Based on these

Acknowledgments

This research is supported by the National Natural Science Foundation of China (NSFC) (61303113, 81401558 and 61402337). This work is also supported by Science and Technology Committee of Shanghai Municipality of China (KF1405), Zhejiang Provincial Natural Science Foundation of China (LY14H230001, LQ13G010007, LQ13F020011, LY14F020035), and the key construction academic subject (medical innovation) of Zhejiang Province (11-CX26).

References (28)

  • C.W. Hsu

    Early hemoperfusion may improve survival of severely Paraquat-poisoned patients

    PLoS One

    (2012)
  • W.P. Wu

    Addition of immunosuppressive treatment to hemoperfusion is associated with improved survival after paraquat poisoning: a nationwide study

    PLoS One

    (2014)
  • P. Houze

    Toxicokinetics of paraquat in humans

    Hum. Exp. Toxicol.

    (1990)
  • R. Zhang

    Multicategory classification using an extreme learning machine for microarray gene expression cancer diagnosis

    IEEE/ACM Trans. Comput. Biol. Bioinf.

    (2007)
  • Cited by (133)

    • Aluminum oxide nano porous: Synthesis, properties, and applications

      2023, Case Studies in Chemical and Environmental Engineering
    • Nanomaterials as transmitters of non-viral gene vectors: A review

      2023, Case Studies in Chemical and Environmental Engineering
    • Nanowires Properties and Applications: A Review Study

      2023, South African Journal of Chemical Engineering
    • An evolutionary machine learning for multiple myeloma using Runge Kutta Optimizer from multi characteristic indexes

      2022, Computers in Biology and Medicine
      Citation Excerpt :

      Li et al. [25] came up with an idea for an innovative intelligent system that could efficiently diagnose tuberculous pleural effusion. An effective machine learning method was proposed by Hu et al. [26] for the diagnosis of patients who had been poisoned by paraquat. This method had excellent classification accuracy, the area under the curve (AUC), sensitivity, and specificity, with respective values of 91.64%, 91.56%, 91.33%, and 91.78%.

    • Implementation of Modified Multi-Objective Particle Swarm Optimization to multi-machine power system stability

      2022, Journal of Cleaner Production
      Citation Excerpt :

      Abd-Elazim and Ali (2013) have proposed an algorithm that combines PSO and BFOA methods to tune the PSS variables in a system (Abd-Elazim et al., 2013). Modernization of electric power systems and interconnection of renewable plants, that is, wind and solar cells, lead to a new concept (Hu et al., 2015). Shan et al. (2021) implemented with success double adaptive weight mechanism into the MFO algorithm, named WEMFO.

    View all citing articles on Scopus
    1

    Contributed equally to this work.

    View full text