Identification of Knee Osteoarthritis Based on Bayesian Network: Pilot Study

Abstract

Background Early identification of knee osteoarthritis (OA) can improve treatment outcomes and reduce medical costs. However, there are major limitations among existing classification or prediction models, including abstract data processing and complicated dataset attributes, which hinder their applications in clinical practice. Objective The aim of this study was to propose a Bayesian network (BN)–based classification model to classify people with knee OA. The proposed model can be treated as a prescreening tool, which can provide decision support for health professionals. Methods The proposed model’s structure was based on a 3-level BN structure and then retrained by the Bayesian Search (BS) learning algorithm. The model’s parameters were determined by the expectation-maximization algorithm. The used dataset included backgrounds, the target disease, and predictors. The performance of the model was evaluated based on classification accuracy, area under the curve (AUC), specificity, sensitivity, positive predictive value (PPV), and negative predictive value (NPV); it was also compared with other well-known classification models. A test was also performed to explore whether physical fitness tests could improve the performance of the proposed model. Results A total of 249 elderly people between the ages of 60 and 80 years, living in the Kongjiang community (Shanghai), were recruited from April to September 2007. A total of 157 instances were adopted as the dataset after data preprocessing. The experimental results showed that the results of the proposed model were higher than, or equal to, the mean scores of other classification models: .754 for accuracy, .78 for AUC, .78 for specificity, and .73 for sensitivity. The proposed model provided .45 for PPV and .92 for NPV at the prevalence of 20%. The proposed model also showed a significant improvement when compared with the traditional BN model: 6.3% increase in accuracy (from .709 to .754), 4.0% increase in AUC (from .75 to .78), 6.8% increase in specificity (from .73 to .78), 5.8% increase in sensitivity (from .69 to .73), 15.4% increase in PPV (from .39 to .45), and 2.2% increase in NPV (from .90 to .92). Furthermore, the test results showed that the performance of the proposed model could be largely enhanced through physical fitness tests in 3 evaluation indices: 10.6% increase in accuracy (from .682 to .754), 16.4% increase in AUC (from .67 to .78), and 30.0% increase in specificity (from .60 to .78). Conclusions The proposed model presents a promising method to classify people with knee OA when compared with other classification models and the traditional BN model. It could be implemented in clinical practice as a prescreening tool for knee OA, which would not only improve the quality of health care for elderly people but also reduce overall medical costs.

Externally validated models for first diagnosis and risk of progression of knee osteoarthritis

“…The models built using the OAI data perform well when tested on the unseen MOST data and have comparable results with several other models already in existence that use more complex methods. The work using a Bayesian network approach had a model AUC of 0.78, the LogR model used in the paper that also used ANNs had a performance of 0.76 and 0.63 for the testing and validation data respectively [13,59]. The same paper reported the performance of the ANN as slightly higher, 0.81 and 0.67 for the testing and validation data respectively.…”

“…It is worth noting that many modelling techniques have been used when attempting to develop prediction tools for KOA including support vector machines, tree based methods, mixed-effect mixture models and modelling using MRI data [11]. As models for KOA have great potential for use in clinical settings more complex methods are also commonly used [12][13][14]. As with any approach to ML modelling, there are advantages and limitations.…”

“…When considering models for use with diseases such as KOA there are often many issues, spanning from data availability to problems with reported values due to differing interpretation of questions. Often, with disorders such as KOA, when recruiting participants the pool of subjects is limited, resulting in small sample sizes for modelling [13,17]. Another issue is that subjects often do not have a formal diagnosis and therefore the modelling outcome can be selfreported [13].…”

“…Often, with disorders such as KOA, when recruiting participants the pool of subjects is limited, resulting in small sample sizes for modelling [13,17]. Another issue is that subjects often do not have a formal diagnosis and therefore the modelling outcome can be selfreported [13]. A 2017 study developed a risk prediction model to predict knee pain, but stated this could not be extended to the use of KOA [18].…”

Risk factors and a Bayesian network model to predict ischemic stroke in patients with dilated cardiomyopathy

“…Bayesian network (BN) models possess certain advantages in the medical domain, including adaptability and strong robustness against missing values ( Sheng et al, 2019 ). As to adaptability, building the BN model can start with limited domain knowledge, which is then simplified or extended by inputting new knowledge to meet various needs.…”