A probabilistic SVM based decision system for pain diagnosis

Abstract

Low back pain (LBP) affects a large proportion of the population and is the main cause of work disabilities worldwide. The mechanism of LBP remains largely unknown and many existing clinical treatment of LBP may be not effective to individual patients. Thus the diagnosis and treatment evaluation is crucial for LBP patients. Probabilistic support vector machine (PSVM) decision system is proposed in this article to deal with the diagnosis and treatment evaluation of LBP. The decision system consists of qualitative knowledge model and quantitative model. Expert knowledge and clinical experience are integrated into the design. To deal with the uncertainties in patients samples, PSVM is employed to learn the decision rules from data. The proposed decision system is applied to LBP patients and achieves better performance than the original system.

Introduction

Low back pain (LBP) affects a large proportion of the population in the world. Sixty to ninety percent of the adult population is at risk of developing low back pain in their lives (Andersson, 1999, Skovron et al., 1994). In 1997, LBP costs US 171 billion in the industrial setting of the US (Leigh, Markowitz, Fahs, Shin, & Landrigan, 1997), indirect costs related to days lost from work are substantial, with approximately 2% of the US work force compensated for back injuries each year (Andersson, 1999). Researchers and clinicians have developed several treatment regimes for LBP, which includes nonsteroidal anti-inflammatory drugs (NSAIDs), skeletal muscle relaxants, opioid analgesics, benzodiazepines, systemic corticosteroids, antidepressant medications, and antiepileptic drugs (Cherkin et al., 1998, Bernstein et al., 2004, Luo et al., 2004a, Luo et al., 2004b, Di Iorio et al., 2000). However, the efficacy of one treatment regime may vary significantly in populations. In addition, conflicting results were reported in previous researches, for example, sufficient evidence have found to support the usage of NSAIDs and skeletal muscle relaxants for LBP, and limited evidence for other treatments (Deyo, 1996). But another research recommended acetaminophen for mild or moderate pain other than NSAIDs (Chou & Huffman, 2007). Since selecting treatment regime especially pharmacologic therapy is usually associated with the trade-off between benefits and harms, it is critical to diagnosis LBP precisely and evaluates the efficacy of treatment for individual patients.

On the other hand the pathology underlying LBP remains largely unknown. Since most of LBP are happened together with mechanical function alteration, many researchers explore the relationship between LBP and various mechanical properties. Particularly, static or dynamic paraspinal surface electromyography (sEMG) has been observed increasing in chronic LBP patients (Arena et al., 1989, Arena et al., 1990), while some long-term pathology are related with increased stiffness (Comerford & Mottram, 2001). Hu, Wong, Lu, and Kawchuk (2009) found that the correlation of sEMG and muscle stiffness signal may have relation with the treatment efficacy. However, only the general expert knowledge is not sufficient to form clinical diagnosis tools for individual patients since there are various uncertainties in individual samples such as measurement noise, sensor location variance and diversity of patients’ specificities.

In this paper, a decision system is proposed for the diagnosis of LBP. The decision system is the integration of a qualitative knowledge model and a quantitative model. The expert knowledge is integrated into the decision system to formulate the qualitative knowledge model. To reduce the uncertainties in sample patients, probabilistic SVM (PSVM) (Yang & Li, submitted for publication) is employed to formulate the quantitative model. PSVM is an extended modeling method developed from support vector machine (SVM) and has the advantages of SVM, such as the ability to work at high dimensionality, good generalization performance and global optimal solution. Moreover, PSVM could handle uncertainties in data samples, while SVM is sensitive to uncertainties (Matić Guyon and Vapnik, 1996, Zhang, 1999). Similar to SVM, PSVM works by mapping samples in input space to high dimensional feature space, in which those samples could be separated linearly. In PSVM, the uncertain parts are characterized by a probabilistic distribution function (PDF) of the separating margin in feature space. A probabilistic separating margin is created in PSVM instead of deterministic margin in SVM. Through the creation of probabilistic margin, more information is integrated in PSVM, thus PSVM works better than SVM in uncertain circumstances.