 Research article
 Open access
 Published:
Network or regressionbased methods for disease discrimination: a comparison study
BMC Medical Research Methodology volumeÂ 16, ArticleÂ number:Â 100 (2016)
Abstract
Background
In stark contrast to networkcentric view for complex disease, regressionbased methods are preferred in disease prediction, especially for epidemiologists and clinical professionals. It remains a controversy whether the networkbased methods have advantageous performance than regressionbased methods, and to what extent do they outperform.
Methods
Simulations under different scenarios (the input variables are independent or in network relationship) as well as an application were conducted to assess the prediction performance of four typical methods including Bayesian network, neural network, logistic regression and regression splines.
Results
The simulation results reveal that Bayesian network showed a better performance when the variables were in a network relationship or in a chain structure. For the special wheel network structure, logistic regression had a considerable performance compared to others. Further application on GWAS of leprosy show Bayesian network still outperforms other methods.
Conclusion
Although regressionbased methods are still popular and widely used, networkbased approaches should be paid more attention, since they capture the complex relationship between variables.
Background
Recently, an explosion of data has been derived from clinical or epidemiological researches on specific diseases, and the advent of highthroughput technologies also brought an abundance of laboratory data [1â€“4]. The acquired variables may range from subject general characteristics, history, physical examination results, blood, to a particularly large set of genetic markers. It is desirable to develop efficient data mining strategies to extract more information rather than put the data aside. Diagnostic prediction models are widely applied to guide clinical professionals in their decision making by estimating an individualâ€™s probability of having a specific disease [5â€“9]. One common sense is, from a networkcentric perspective, biological phenomena depend on the interplay of different levels of components [10â€“12]. For data on network structure, complex relationships (e.g. high collinearity) inevitably exist in large sets of variables, which pose great challenges on conducting statistical analysis properly. Therefore, it is often hard for clinical researchers to determine whether and when to use which exact model to support their decision making.
Regressionbased methods, although may be unreasonable to some extent under the network framework, is still a priority in disease diagnosis or discrimination problem [6, 13â€“15], which is easier to be accepted by clinical researchers due to the interpretability of model parameters and ease of use. However, for regression model, some assumptions needed to be made may limit the use, such as linearity and additivity [16â€“18]. The performance of the regression model can be affected by the collinearity between the input variables, which is commonly encountered in dataset with complex relationship. Although a logistic regression model can consider the relationship between the covariates by adding interaction terms, the number of possible interactions increases exponentially as the number of input variables increases, resulting in the complex process of specification of interaction and inevitably low power.
To overcome the above problems, numerous machine learning methods have emerged as potential alternatives to logistic regression analysis, such as neural network, random forest, decision trees [5, 19â€“21]. Neural networks, with few assumptions about the data distribution, can reflect the complex nonlinear relationships between the predictor variables and the outcome by the hidden nodes in the hidden layer. This not only greatly simplifies the modeling work compared to logistic regression model but enables us to model complex forms between variables. If the logistic sigmoid activation function is used, the network without a hidden layer is actually identical to a logistic regression model, and neural networks can be thought as a weighted average of logit functions with the weights themselves estimated [22, 23]. Neural networks do not yet jump out from the scope of regression, which can be viewed as a type of nonparametric regression method.
Motivated by the network perspective, a more formal and visualized representation, usually offered by mathematical graph theory, seems to be more appropriate to describe the biological phenomena. Among these, Bayesian networks provide a systematic method for structuring probabilistic information about a network, which have been receiving considerable attention over the last few decades in a number of research fields [24â€“26]. Bayesian networks are easily understood since they represent knowledge through a directed acyclic graph (DAG) with nodes and arrows. The network structure can be either generated from data by structural learning or elicited from experts. It could not only avoid statistical assumptions, but also handle the relationship between a larger numbers of predictors with their interactions.
In stark contrast to commonly accepted networkcentric perspective view for complex disease, regressionbased methods are preferred, especially for epidemiologists and clinical professionals, which usually lead to considerate and easily interpreted results. It remains a controversy whether the networkbased methods have advantageous performance than others in discrimination ability, and to what extent do they outperform. In particular, complex diseases often result from multiple genes or molecules interplays within biological pathways or gene regulatory networks. Under such condition, are regressionbased methods with correlated genetic markers sufficient to reflect biological reality? To the best of our knowledge, few attempts were conducted to determine in which case network or regressionbased methods should be applied. The focus of this paper is, through a series of simulations, to assess how the networkbased methods work compared to regressionbased methods in prediction performance under different scenarios (the input variables are independent or in network relationship). To achieve this goal, we applied logistic regression, neural network, and Bayesian network on the different datasets.
Method
Simulation studies
Simulation studies were conducted to evaluate the performance of the logistic regression, neural network, and Bayesian network. The area under the receiveroperating characteristic curve (AUC) which is normally employed to measure discrimination ability [27],and Brier score was used to compare the accuracy of the three methods. Additional techniques (e.g. crossvalidation (CV), bootstrapping, leverage correction) [28] must be used to alleviate overfitting problem generally encountered in statistics and machine learning. In this paper, the overfitting was corrected using 10fold cross validation (AUCCV) to assess the prediction performance of the above three methods. For each simulation, 100 repeats of 10fold CV were conducted in order to yield sufficient precision.
Under the null hypothesis, the AUC should be around 0.5, meaning that the prediction model is not helpful at all. In order to test whether the prediction methods are stable, we first generated the datasets under the null hypothesis. Network datasets were generated using software Tetrad [29]. For each network, we first generated a directed acyclic graph with a set of binary variables representing the input variables and a binary outcome variable indicating the disease status. Conditional probability table for each variable was defined subsequently. Conditional on the values of its parent variables, there is a defined probability that a variable will take on its possible values. Thus the influence of variables can be reflected by the conditional probability table. Restricting on six nodes including five input variables and one disease outcome, we considered two scenarios of the null hypsthesis: 1) each variable was generated independently; 2) the input variables were network constructed but not associated with the disease. For each scenario, 100,000 individuals were generated to form a hypothetical population from which the samples were randomly selected with different sample sizes (Nâ€‰=â€‰30, 50, 100, 200, 500 or 1000). To examine the stability of the three methods, we randomly sampled N individuals respectively for the calculation of the AUC and the average AUCCV. A total of 1000 simulations were repeated for each sample size.
Under the alternative hypothesis, datasets from different network structures were generated to assess the discriminatory ability as well as the prediction accuracy. We simulated a regular network and two extreme scenarios including chain network and wheel network to evaluate the performance of three different methods. For each data set, similar simulations were accomplished as above to obtain the AUC and the Brier score with different sample size. In particular, more general logistic models were employed to extract the nonlinear effect and interactions between variables for data in regular network. Multivariate regression splines was used to fit the logistic model using earth function in R package earth. We used two strategies to consider the interaction between the input variables: 1) the product term was determined by the network structure (i.e. the product term between two variables was added to the model only if there was an edge between the variables). 2) all the pairwise product terms between the variables were added in the logistic model and selected by stepwise algorithm.
In addition, we might be also interested in how the network methods perform under the special case when the input variables are in completely linear relationship. We generated 100,000 individuals with five independent variables, with each variable following a Binomial distribution. Given the effect of the input variables Î²â€‰=â€‰(1.5,â€‰1.5,â€‰1.5,â€‰1.5,â€‰1.5), the binary response indicating disease status was generated using logistic regression model.
The performances of Bayesian network and neural network were implemented using the R package bnlearn and the R package neuralnet. For Bayesian network, scorebased structure algorithms hill climbing (HC) method (hc function) was employed for structure learning and Bayes method for parameter learning (bn.fit function). The neuralnet function was used to fit the neural network, and the number of hidden nodes in neural network was determined using cross validation.
Application
The Bayesian network, neural network, logistic regression and regression splines were also applied to a real genotype data for predicting leprosy of Han Chinese with a case control design, which contains 706 cases and 514 controls. The genetically unmatched controls were removed to avoid population stratification. Previous genomewide association study (GWAS) of leprosy of Han Chinese [30] has identified significant associations between 16 SNPs in seven genes (CCDC122, C13orf31, NOD2, NFSF15, HLADR, RIPK2and LRRK2). In this paper, we fitted the three models using the identified 16 SNPs respectively to compare their abilities in predicting Leprosy. The 100 repeats of AUC and Brier score with cross validation were calculated for all the methods.
Result
FigureÂ 1 shows the estimated AUC and the average AUCCV of the Bayesian network, neural network and logistic regression under the null hypothesis mentioned above. It reveals that the AUCCV of all the methods are close to 0.5 when the sample size is large (more than 500), illustrating the AUCCV could be a convincing indicator to assess the prediction performance. While AUC is far from 0.5 especially with small sample size and might not be considered in the comparison.
FigureÂ 2a shows a simulated disease network, this network data were generated through software Tetrad [29] under the given conditional probabilities. FigureÂ 2b depicts the average AUCCV slightly increase monotonically by sample size, and they are close to the true value when sample size arrives 1000. The result indicates that Bayesian network outperforms the logistic regression and neural network when such network exists. The logistic regression with interaction terms improved the AUCCV quite slightly, while regression splines improved the discriminatory ability by capturing the nonlinear effect. TableÂ 1 depicts the Brier scores of the methods. The Bayesian network still has the best prediction accuracy, followed by the regression splines. The other four methods have comparably inferior performance.
FigureÂ 2d shows the performance under different sample sizes given the datasets generated from chain network (Fig.Â 2c). It seems that the AUCCV of all methods are not significantly affected by sample size. The Bayesian network has superior performance followed by the neural network, while the regression models work inefficiently that may be partly due to the correlated structure between the input variables. Similar trends can be found for Brier score of the methods.
Given the datasets generated from wheel network shown in Fig.Â 3a, it depicts the discriminatory ability and accuracy of all these methods are comparable, while the regression models have slightly inferior performance with small sample size. FigureÂ 3c demonstrates that the 10fold crossvalidation AUC of these methods slightly increase monotonically by sample size, while the Brier score decrease monotonically by sample size (please see Additional file 1: Table S3). The prediction ability of the methods are quite close when the independent variables satisfied the linearity.
Result of application
TableÂ 2 shows the SNP information and univariate analysis result with Leprosy of the selected 16 SNPs in the model. Seven SNPs entered the multivariate logistic regression model using stepwise approach with results shown in TableÂ 3. Hill climbing method was employed for structure learning and Bayes method for parameter learning using R package bnlearn. Hugin software [31] was used to better visualize the graphical representation of the Bayesian network that is shown in Fig.Â 4. One hidden layer with four units was used in neural network. TableÂ 4 depicts the AUC and Brier score with 100 repeats of 10fold cross validation of all the methods. The results show Bayesian network, though just slightly improved, outperforms other two methods, which indicate the network relationships exist in the 16 SNPs. Neural network has inferior performance than the other methods, which may be due to the fact that it is difficult to determine the optimum value for number of hidden layers and nodes.
Discussion
Several studies demonstrated the importance of investigating a disease from the network perspective. It remains an interesting problem whether the networkbased methods have advantageous performance than others, and to what extent do they outperform. The focus of this paper is to bridge this gap and assess their performance in prediction mainly through a series of simulations, with four methods (Bayesian network, neural network, logistic regression and regression splines). We employed the adjusted AUC and Brier score to assess the prediction performance of all the methods. The adjusted AUC are close to 0.5 under null hypothesis when the sample size is larger than 500. It reveals that the discriminatory ability of all methods varies quite slightly with sample size. Four datasets under different assumptions were designed and Bayesian network showed a better performance when the variables are in a network relationship (Fig.Â 2a) or in a chain structure (Fig.Â 2c). The regression splines improved the model performance a lot by extracting the nonlinear effect, while the interaction model improved slightly. But they are still inferior to Bayesian network, which indicates that it is not straightforward to capture the whole network information using regression method. For the network structure, we partitioned the effects into additive and nonadditive effects to quantify the proportion of the relationships between the input variables and the outcome is nonadditive on the logit scale as one reviewer suggested. We have embedded ordinary regression in a larger model including all twoway interactions and calculated the proportion of likelihood ratio chisquare statistics, it showed that 23Â % of the effects are due to nonadditive effects. The AIC for the additive model and the full model of all the population are 134194.5 and 133034.1 respectively. Particularly, for the special wheel network structure, our simulation results illustrated that the Bayesian network has similar performance of logistic regression model (Fig.Â 3a), which is strongly consistent with the previous findings [31], same phenomenon has also been found in the case when data was generated using a logistic model (Fig.Â 3c). Further application on leprosy GWAS show Bayesian network, though just slightly improved, still outperforms other methods, followed by regression splines and logistic regression, and neural network has the worst performance after cross validation. Considering that it seems to be unreasonable to predict leprosy using the nonrisk SNPs, we thus have chosen the specific 16 risk SNPs which have been identified and validated from the GWAS of leprosy.
Logistic regression models are well suited to be used when some assumptions is satisfied (Fig.Â 3c), while they work inferior when the assumptions are violated and cannot capture the nonlinear and unknown relationships often existed in the variables. It would be of great value to add penalized MLE to the comparators to make the comparison with logistic regression more informative, which remains a goal of our future work. Neural networks can reflect the complex relationships between the predictor variables and the outcome by the hidden nodes in the hidden layer. However, as a weighted average of logit functions with the weights themselves estimated, it does not jump out from the scope of regression yet. Moreover, the network structure must be prespecified and no gold standard can be adopted to determine the optimum value for number of hidden layers and nodes. Bayesian networks capture the complex relationship well between a larger number of predictors with their interactions without statistical assumptions, when the disease is caused through pathways or networks, and the usefulness of Bayesian networks for predicting is clearly recognized through simulation. Even when the dataset were generated from regression model, the Bayesian network techniques had a considerate performance (Fig.Â 3c). Actually, the Bayesian network is confirmed theoretically to be equivalent to a logistic regression problem under a simple graphtheoretic condition (e.g. wheel network in our simulation) [31, 32]. One major drawback of Bayesian network is that its performance can be heavily influenced by the network structure, which sometimes may not capture the real population structure information, though many algorithms have been provided for network structure learning.
These comparisons are dependent on the character of a particular data set, and one cannot conclude whether one method will be superior to the others in a given data set without dissecting the data structure. Overall, regressionbased methods are recommended for welldesigned research projects with a small amount of variables where researchers can understand the potential predictors and possible interactions, since it is easier to be implemented and to be accepted by clinical researchers. For the dataset with complex relationships, especially for commonly accepted networkcentric perspective for complex disease, networkbased methods such as Bayesian network are more appropriate to act as an exploratory tool. These methods can extract the patterns and relationships in data without constraining the predictors, and achieve a high performance in discrimination.
Conclusion
Although regressionbased methods are still popular and widely used, networkbased approaches should be paid more attention, since it captures the complex relationship between variables.
Abbreviations
 AUC:

The area under the receiveroperating characteristic curve
 AUCCV:

The AUC using 10fold cross validation
 BN:

Bayesian network
 CV:

Cross validation
 GWAS:

Genomewide association study
 NN:

Neural network
 RS:

Regression splines
References
Albert R. Network inference, analysis, and modeling in systems biology. Plant Cell. 2007;19(11):3327â€“38.
NguyenNielsen M, Svensson E, Vogel I, Ehrenstein V, Sunde L. Existing data sources for clinical epidemiology: Danish registries for studies of medical genetic diseases. Clin Epidemiol. 2013;5:249.
St Sauver JL, Grossardt BR, Yawn BP, Melton LJ, Pankratz JJ, Brue SM, et al. Data resource profile: the Rochester Epidemiology Project (REP) medical recordslinkage system. Int J Epidemiol. 2012;41(6):1614â€“24.
Mayr M. From data gathering to systems medicine. Cardiovasc Res. 2013;97(4):599â€“600.
Tu JV. Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol. 1996;49(11):1225â€“31.
Yonemori K, Tateishi U, Uno H, Yonemori Y, Tsuta K, Takeuchi M, et al. Development and validation of diagnostic prediction model for solitary pulmonary nodules. Respirology. 2007;12(6):856â€“62.
Hendriksen J, Geersing GJ, Moons K, Groot J. Diagnostic and prognostic prediction models. J Thromb Haemost. 2013;11(s1):129â€“41.
Polit DF, Beck CT. Nursing research: Principles and methods. New York:Â Lippincott Williams & Wilkins; 2004.
Bedogni G. Clinical Prediction Modelsâ€”a Practical Approach to Development, Validation and Updating. J R Stat Soc A Stat Soc. 2009;172(4):944.
Kitano H. Systems biology: a brief overview. Science. 2002;295(5560):1662â€“4.
Beltrao P, Cagney G, Krogan NJ. Quantitative genetic interactions reveal biological modularity. Cell. 2010;141(5):739â€“45.
Aloy P, Russell RB. Structural systems biology: modelling protein interactions. Nat Rev Mol Cell Bio. 2006;7(3):188â€“97.
Visser H, le Cessie S, Vos K, Breedveld FC, Hazes JM. How to diagnose rheumatoid arthritis early: a prediction model for persistent (erosive) arthritis. Arthritis Rheum. 2002;46(2):357â€“65.
Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M, Obuchowski N, et al. Assessing the performance of prediction models: a framework for some traditional and novel measures. Epidemiology. 2010;21(1):128.
Guthrie W, Swineford LB, Nottke C, Wetherby AM. Early diagnosis of autism spectrum disorder: stability and change in clinical diagnosis and symptom presentation. J Child Psychol Psychiatry. 2013;54(5):582â€“90.
Lee S, Abbott P, Johantgen M. Logistic regression and Bayesian networks to study outcomes using large data sets. Nurs Res. 2005;54(2):133â€“8.
Hosmer Jr DW, Lemeshow S. Applied logistic regression. New York: John Wiley & Sons; 2004.
Harrell Jr FE, Lee KL, Matchar DB, Reichert TA. Regression models for prognostic prediction: advantages, problems, and suggested solutions. Cancer Treat Rep. 1985;69(10):1071â€“7.
Khan J, Wei JS, Ringner M, Saal LH, Ladanyi M, Westermann F, et al. Classification and diagnostic prediction of cancers using gene expression profiling and artificial neural networks. Nat Med. 2001;7(6):673â€“9.
Briggs F, Ramsay PP, Madden E, Norris JM, Holers VM, Mikuls TR, et al. Supervised machine learning and logistic regression identifies novel epistatic risk factors with PTPN22 for rheumatoid arthritis. Genes Immun. 2010;11(3):199â€“208.
Dybowski R, Gant V, Weller P, Chang R. Prediction of outcome in critically ill patients using artificial neural network synthesised by genetic algorithm. Lancet. 1996;347(9009):1146â€“50.
Dreiseitl S, OhnoMachado L. Logistic regression and artificial neural network classification models: a methodology review. J Biomed Inform. 2002;35(5):352â€“9.
Hastie T, Tibshirani R, Friedman J, Franklin J. The elements of statistical learning: data mining, inference and prediction. Math Intell. 2005;27(2):83â€“5.
Nielsen TD, Jensen FV. Bayesian networks and decision graphs. New York: Springer; 2009.
Heckerman D. Bayesian networks for data mining. Data Min Knowl Disc. 1997;1(1):79â€“119.
Uusitalo L. Advantages and challenges of Bayesian networks in environmental modelling. Ecol Model. 2007;203(3):312â€“8.
Bradley AP. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recogn. 1997;30(7):1145â€“59.
Faber NM, Rajko R. How to avoid overfitting in multivariate calibrationâ€”The conventional validation approach and an alternative. Anal Chim Acta. 2007;595(1):98â€“106.
Scheines R, Spirtes P, Glymour C, Meek C, Richardson T. The TETRAD project: Constraint based aids to causal model specification. Multivar Behav Res. 1998;33(1):65â€“117.
Zhang F, Huang W, Chen S, Sun L, Liu H, Li Y, et al. Genomewide association study of leprosy. New Engl J Med. 2009;361(27):2609â€“18.
Jensen F, KjÃ¦rulff U, Lang M, et al. HUGIN  The Tool for Bayesian Networks and Influence Diagrams. European Workshop on Probabilistic Graphical MODELS. 2002.
Roos T, Wettig H, GrÃ¼nwald P, MyllymÃ¤ki P, Tirri H. On discriminative Bayesian network classifiers and logistic regression. Mach Learn. 2005;59(3):267â€“96.
Acknowledgements
We are very grateful of research of the Leprosy GWAS and other colleagues for their support.
Funding
This work was jointly supported by grants from National Natural Science Foundation of China [grant numbers 81573259, 81273177, 31200994]. The funding bodies were not involved in the analysis and interpretation of data, or the writing of the manuscript.
Availability of data and materials
Additional data is available on request from the first author.
Authorsâ€™ contributions
XSZ, ZSY and FZX conceptualized the study, XSZ and ZSY analyzed the data and prepared for the manuscript. JDL and HKL contributed on the study design. All authors approved the manuscript.
Competing interests
The authors declare that they have no competing interests.
Consent for publication
Not applicable.
Ethics approval and consent to participate
The data are from published studies [30], in which all the participants were recruited with written informed consent. The study was approved by the institutional IRB committees at the Shandong Provincial Institute of Dermatology and Venereology, Shandong Academy of Medical Science and the Anhui Medical University.
Author information
Authors and Affiliations
Corresponding author
Additional file
Additional file 1:
Relevant tables for the comparison of Brier score. (DOCX 18 kb)
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Cite this article
Zhang, X., Yuan, Z., Ji, J. et al. Network or regressionbased methods for disease discrimination: a comparison study. BMC Med Res Methodol 16, 100 (2016). https://doi.org/10.1186/s1287401602072
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s1287401602072