 Research
 Open Access
 Published:
Classification of breast cancer recurrence based on imputed data: a simulation study
BioData Mining volume 15, Article number: 30 (2022)
Abstract
Several studies have been conducted to classify various real life events but few are in medical fields; particularly about breast recurrence under statistical techniques. To our knowledge, there is no reported comparison of statistical classification accuracy and classifiers’ discriminative ability on breast cancer recurrence in presence of imputed missing data. Therefore, this article aims to fill this analysis gap by comparing the performance of binary classifiers (logistic regression, linear and quadratic discriminant analysis) using several datasets resulted from imputation process using various simulation conditions. Our study aids the knowledge about how classifiers’ accuracy and discriminative ability in classifying a binary outcome variable are affected by the presence of imputed numerical missing data. We simulated incomplete datasets with 15, 30, 45 and 60% of missingness under Missing At Random (MAR) and Missing Completely At Random (MCAR) mechanisms. Mean imputation, hot deck, knearest neighbour, multiple imputations via chained equation, expectedmaximisation, and predictive mean matching were used to impute incomplete datasets. For each classifier, correct classification accuracy and area under the Receiver Operating Characteristic (ROC) curves under MAR and MCAR mechanisms were compared. The linear discriminant classifier attained the highest classification accuracy (73.9%) based on meanimputed data at 45% of missing data under MCAR mechanism. As a classifier, the logistic regression based on predictive mean matching imputeddata yields the greatest areas under ROC curves (0.6418) at 30% missingness while knearest neighbour tops the value (0.6428) at 60% of missing data under MCAR mechanism.
Introduction
Background
Missing data are frequently encountered in clinical research and they can negatively impact the research findings if not properly handled prior to data analysis due to their likelihood of biasing the results [1]. In clinical context, missing data may arise because of random errors caused by a measuring equipment, attrition due to social or natural processes as for instance death, nonresponse to some sensitive or unclear questions in a survey, and patients failing to report to a routine clinic [2]. Many researchers often discard incomplete cases during analysis stage, a technique known as complete case analysis (CCA). The CCA excludes any patients with at least one missing data value in statistical analysis, and thus reducing statistical power and introducing bias to the results [3]. The proper way to handle problems caused by missing data is to use imputation techniques [4] that estimate and replace missing values toyield complete dataset [5]. Imputation is an active area of research and several techniques including mean imputation, hot deck, knearest neighbour, various forms of multiple imputations by chained equations, predictive mean matching imputations, expectedmaximization via bootstrapping imputations among others are used to handle with missing data under different conditions; however, there is no unique conclusive answer to which technique is best under a set of several conditions such as missing probabilities, missingness mechanisms, and patterns of missing data. An appropriate method to check the suitability and usefulness of imputation method is the use of simulation studies.
Prior to imputing missing data, it is important to understand the mechanism that generated the missing data. Every data value in a dataset has probability to be missing. Missing probabilities are governed by the process known as missing data mechanisms; they are divided into: Missing Completely At Random (MCAR), Missing At Random (MAR), and Missing Not At Random (MNAR) [6]. In MCAR, missing values neither depend on observed nor on unobserved values [3, 7] An example of MCAR data is when a test tube with patient’s urine sample is accidentally dropped and broken, yielding missing data associated with that laboratory testing. Data is said to be MAR when missing values depend only on observed values [3, 8, 9]. The reasons behind the missing data are associated with patient’s characteristics that are known [3]. An example of MAR data is when a patient deliberatively decide not to answer a certain question, especially if it is about his or her privacy [7]. The MNAR data occurs if the distribution of dataset containing missing values depends on missing values [9]. This implies that missing probability is associated to characteristics that the researcher cannot know about [8]. An example of MNAR data is when patients with low level of education tend to avoid questions concerning their educational status.
Related work and analysis gap
Various studies have been conducted on distinct clinical datasets to classify breast recurrence under statistical and machine learning techniques. In some situations, simulation has been used along with classification process. Researchers use simulations to get answer to their specific questions about data analysis, and to assess models performance under specified desired conditions [10] Asimulation study on numerical data resulted from imputation techniques used to access the accuracy of prognostic models. Models’ discriminative ability to separate breast cancer patients with and without recurrence was determined by areas under ROC curves [11]. Another simulation study compared the performance of fully parametric imputations when imputation model was correctly specified and when miss specified using predictive mean matching and local residual draws imputation techniques [12].
The other study [13] evaluated the performance of different statistical and machine learning approaches in predicting recurrence of breast cancer from patients’ large real breast cancer dataset containing missing values. The study used discrimination and calibration measures to assess usefulness of the prognostic model. On the other hand, the comparison of imputation algorithms for building sensor data across several percentage of missing data was conducted by comparing the differences between real and imputed values through the use of Root Mean Squared Error and Mean Absolute Error estimates; its conclusion emphasized the necessity of identifying percentage of missingness prior to selecting proper imputation technique so as to reach plausible results [14]. Moreover, imputation techniques were used to evaluate the performance of model via discrimination, calibration, and effectiveness of classifiers in relation to time used to build a model in estimating the risk of unprovoked venous thromboembolism recurrence [15]. The focus was on diagnostic model development via multivariable logistic or prognostic model via survival regression analysis.
With due respect to mentioned related work, we aim to fill an analysis gap about performance of statistical classification techniques based on several datasets resulted from imputation techniques via simulation settings under various conditions. Our study observes the behaviour of three binary classifiers (logistic regression, linear and quadratic discriminant analysis) on numerical missing data based on simulation of real breast cancer dataset. Six imputed datasets across four different percentages of missing data and two missing data mechanisms were used for classification purpose. The goal was to assess how classifiers’ accuracy and discriminative ability (in classifying a binary outcome: breast cancer recurrence) are affected by the presence of imputed missing data values under various simulation conditions. We report the percentage accuracy and the areas under the ROC curves resulted from each classifier for 15, 30, 45 and 60% of missing data under MAR and MCAR mechanisms.
Materials and methods
Data descriprion
The study uses the simulated breast cancer datasets, each having 693 observations (with varying percentages of missing data contained in independent variables only). The variables used in this article were extracted from several previous breast cancerrelated studies [10, 13, 16, 17]. The dependent or outcome variable namely ‘cancer recurrence’ has two response categories; ‘yes’ and ‘no’. The response ‘yes’ means a breast cancer comes back after recommended therapy, ‘no’ indicates that the cancer does not come back after respective therapy. The independent or predictor variables are: the age, heart rate, respiratory rate, body mass index, body surface area.
Classification techniques
Classification techniques also known as classifiers are used to predict a categorical response for a case by assigning that case to a certain category. In this article, a binary logistic regression, linear, and quadratic discriminant classifiers [18, 19] were used to predict a group membership for breast cancer cases and classify to either response ‘recurrence’ or ‘nonrecurrence’ based on patients’ demographic and clinical factors; ‘age, body mass index, body surface area, heart rate, and respiratory rate’.
Binary logistic regression
Binary logistic regression (BLR) model considers the probability that a response variable Y belongs to a particular group [18, 20]. In this article, the BLR model intended to predict and classify breast cancer cases to either recurrence or nonrecurrence events given predictors. The model uses a logit or logistic function (a sigmoid function applied in binary classification). A sigmoid function takes predictors’ data (real numbers) and maps them to certain probability value (ρ_{i}) between 0 and 1.
for i = 1, 2, 3, …, 693, β_{0} is a constant term of the model, β_{j} for j = 1, 2, …, 5 are the parameters of the model, ρ_{i} is the probability that i^{th} patient has cancer recurrence, X_{p} for p = 1, 2, …, 5 are the predictors in the model (age, heart rate, respiratory rate, body mass index, body surface area).
The patients’ probabilities (ρ_{i}) of having breast cancer recurrence are estimated by the fitted logistic regression model. The (ρ_{i}) are then used for classification of patients into either of the two categories of breast cancer response variable (recurrence or nonrecurrence) according to the cutoff point of 0.5 for classification; a value greater than 0.5 is classified as ‘recurrence’ event, otherwise classified as ‘nonrecurrence’ event for a patient i (i= 1,2, …, 693). Mathematically, the predicted probability for any patient i is given by ρ_{i}(x):
Linear discriminant analysis
Linear discriminant analysis (LDA) is used to separate different sets cases and allocate new cases to previously defined groups [21]. It assumes that each group are drawn from Multivariate Normal distribution with group definite mean vector (μ) and covariance matrix (∑) [19]. As a classifier, LDA allocates cases x to group k provided that x is in falls inside the group or region k [22]. In this study we applied LDA to discriminate or separate the cases with breast cancer recurrence from those without recurrence by using the discrimination function, δ_{k}(x). An assumption about equal covariance matrix, ∑ among the k = 2 groups was made to make discrimination function taking a linear form in x. A Bayes’ theorem with grouping prior probabilities (π) is used to allocate x in group k for which a function.
\({\delta}_k\left(\boldsymbol{x}\right)={x}^T{\sum}^{1}{\mu}_k\frac{1}{2}{\mu}_k^T{\sum}^{1}{\mu}_k+\log {\pi}_k\Big)\) is largest [19].
Quadratic discriminant analysis (QDA)
As classifier, QDA is like LDA, it makes assumption about Multivariate Normality for observations within each group; however, it doesn’t assume common covariance matrix, i.e., each group assumed to have its own matrix. Bayes’ theorem is utilized to make prediction and allocate cases into respective groups [21, 23]. Nevertheless, . Based on this assumption, a case x assigned to a group k for which the function.
\({\delta}_k\left(\boldsymbol{x}\right)=\frac{1}{2}{\left(x{\mu}_k\right)}^T{\sum}_k^{1}\left(x{\mu}_k\right)\frac{1}{2}\log \mid {\sum}_k\mid +\log {\pi}_k\) is largest [19].
Since LDA and QDA the training and validation datasets, in this article we used 70 and 30% of the sample in each dataset as training and validation data respectively LDA is likely to perform better than QDA when the training cases are relatively few and hence reducing variance is vital. Meanwhile, QDA tends work better when number of training cases is very large, as a result the classifier’s variance does not harm, also it is a recommended classifier if covariance matrix is clearly common among the k groups [19]. .
Simulation settings
The simulation experiment was conducted to compare the performance of imputation methods and classifiers in presence of imputed missing values across different simulation conditions. The simulation of data was based on real breast cancer dataset of size 693 containing both observed and missing values for variables; X_{1}, X_{2}, …, X_{6} respectively, denoting the age, heart rate, respiratory rate, body mass index, body surface area, and recurrence of breast cancer. Data were sampled with fixed mean vector and covariance matrix of all variables while varying the percentages of missing data (i = 1 : 4), changing missingness mechanisms (j = 1 : 2), and imputation methods (k = 1 : 6) producing a total of 4 × 2 × 6 = 48 simulation conditions. The experimental design was built in four steps: (1) Generating complete data set, (2) Amputation (making incomplete data sets from the completed one), (3) Imputation (to fillin missing data values), and (4) Evaluate the performance of each imputation techniques.
In step (1) we generated a complete dataset with size N = 693 observations from multivariate normal distribution [24, 25] with vector of means (μ), and a positive definite covariance matrix (∑) given below; archived by the application of ‘mvrnorm’ function in the package ‘MASS’ [26] of R statistical program.
Step (2) performed the ‘amputation’ (i.e., generating missing data values from the complete data set, in step 1). This involved creating incomplete data sets with varying percentages of missing data (15, 30, 45, and 60%) and two missing data mechanisms (MAR and MCAR). The R function, ‘amput’ [27] was applied to meet this purpose. The third step was to impute the created missing data sets using the following distinct imputation techniques:
Mean imputation
The idea based on this approach is to replace missing data values by the variable’s average score from observed data [3, 13, 28]. Theoretically, the mean imputation is more appropriate when the amount of missing data is small whilst the sample size is large [7]. Mean imputation also known as ‘series mean’ (SMEAN) is calculated as \(\sum_{\boldsymbol{i}=\textbf{1}}^{\boldsymbol{n}}{\boldsymbol{x}}_{\boldsymbol{i}}/\boldsymbol{n}\) where x_{i} is a numerical variable and i = 1, 2, …, n; number of patients with observed covariate’s data values.
Hot deck imputation
This technique replaces each missing data in a variable by the observed data from a patient with ‘identical response’or data in the same variable [28]. The patient with missing data (unrespondent) is known as ‘recipient’ while the responded one with observed data is called the ‘donor’ [7, 29]. This method works better with MCAR or MAR data mechanisms [7]. Consider the values x_{i} = (x_{i1}, …, x_{ip} ) for subject i of p covariates. For a matching recipient i and a donor j, the proximity of potential candidate donors to recipients is defined by maximum deviation given by D_{(i, j)} = max_{k}x_{ik} − x_{jk}, for nicely scaled x_{k} so that the comparability of difference) can be made [29]. In this study the hot deck imputation was implemented by using function ‘hot deck’ from the ‘VIM’ (Visualization and Imputation of Missing Values) package [30] in R statistical software (version 3.6.3).
Knearest neighbour (KNN)
A nonparametric approach used to impute missing data by averaging its neighbouring observed data [14]. The approach is donorbased in which imputed values are either measured as a single records in the dataset (1NN) or as an average value obtained from k records (kNN) [31]. The distance between two observations that is used to define the nearest neighbours is defined as\({D}_{ij}=\frac{\sum_{k=1}^P{w}_k{\tau}_{i,j,k}}{\sum_{k=1}^P{w}_k}\), where w_{k} is the weight and τ_{i, j, k} is the contribution of k^{th} variable. The ratio of absolute distance to range is used for τ_{i, j, k} of continuous variables; \({\tau}_{i,j,k}=\frac{\mid {x}_{i,k}{x}_{j,k}\mid }{r_k}\), whereas x_{i, k} is a value of k^{th} variable of i^{th} observation and r_{k} is the range of k^{th} variable (30). In this study, the hot deck imputation method was performed by using function ‘hot deck’ from the ‘VIM’ (Visualization and Imputation of Missing Values) package [30] in R statistical software (version 3.6.3).
Multiple imputation by chained equations (MICE)
MICE Replaces each missing data with set of P acceptable values [32]. The method works with MCAR or MAR missing data mechanisms. The technique helps to remove potential selection bias which would result if observations with missing values were deleted from the dataset. Moreover, the chance of getting biased standard errors is also reduced [7, 8]. The procedure for the multiple imputations involves the following three steps: ‘imputing data m times, analysis of m imputed datasets, and pooling of results’ [6]. We applied R statistical software to perform these three steps in MICE by storing the results from each in the three special classes; mids (multiple imputed data sets), mira (multiple imputed repeated analysis), and mipo (multiple imputed pooled results) [33].
Predictive mean matching (PMM)
The PMM utilizes both parametric and nonparametric approaches to impute missing data. The parametric aspect, establishes a predictive mean value corresponding to each observation in data. These predictive means are then used to match complete and incomplete observations during imputation process. The nonparametric stage applies the method of Nearest Neighbour Donor to produce original data value from nonmissing observation having nearest predictive mean distance close to missing one so as to impute a missing data value [34, 35]. The function and package ‘mice’ in R statistical software [33] was used to perform the PMM imputation five times, storing results from five complete datasets, and combining the results from five analysed datasets.
Expected maximization via bootstrapping (EMB)
The EMB is a bootstrapbased algorithm used to impute missing data multiple times. It utilizes existing sample data of size n to make new M samples of size n with replacement [36]. Since EMB is a multiple imputationsbased procedure, it performs better under MAR assumption. It is based on EM algorithm, summarized as follows: Firstly, the EM (ExpectationMaximisation) algorithm make use of certain distribution and propose starting/initial values for mean, μ and covariance matrix, ∑ that are then used to calculate an expected value of model’s likelihood. This likelihood is maximised and parameters of the model are estimated and updated. The steps of expectation and maximisation are repeated until convergence of the values is reached [36, 37]. The implementation of EMB in done in Amelia II program starts with bootstrapping an incomplete dataset to generate several bootstrapped datasets, followed by EM process of these data to imputed datasets and then the imputed datasets are analysed separately by standard statistical method, and the results are combine to provide single final results.
Evaluation of imputation methods
The evaluation of applied imputation techniques was through comparison of Root Mean Squared Errors (RMSE) obtained from each imputation techniques under various simulations conditions. For each imputation method, we also report classification accuracy resulted from three classification methods via estimated area under the Receiver Operating Characteristics (ROC) curves resulted from logistic regression using imputed data sets at 15, 30, 45, and 60% missing data under MAR and MCAR mechanisms.
Results
The average values of RMSE for each imputation method (Table 1) tend to increase as percentages of missing data are increased. This reflects the reduction of efficiency of imputations as missing percentages increase in datasets. Under MAR, the EMB technique attained the least RMSE values at lowest percentage (15%) of missing data followed by PMM compared to other methods. The SMEAN acquired lowest values of RMSE followed by PMM at highest percentage of missing data (60%), this indicates that the PMM is more efficient low and high percentages of missing data for MAR and MCAR mechanisms.
The distributions of average RMSE (Fig. 1) for imputation methods under MCAR are much close to symmetric shape as compared those under MAR mechanism, meaning that the MCAR mechanism yields imputation results whose central tendency measures are close to each other, implying a little difference among them under the current study and scenario. All areas under the ROC curves (Table 2) from binary logistic regression classifier using each imputed dataset under four percentages of missing data and two missing data mechanisms are above 0.5 implies that all methods can discriminate recurrence from nonrecurrence cases.
Across all percentages and missing mechanisms, the maximum area under ROC curve is 0.6428 from KNN at 60% missing data with MCAR, followed by PMM at 30% under MAR. Under MAR, the PMM attains the highest values of areas under at 15, 30, and 45% of missing data while the HD appeared top at 60% of missing data; again, the PMM yields the greatest areas under ROC curves at 15 and 45%, while KNN top the value at 60% and MICE at 30% under MCAR assumption. This information and Fig. 2 suggests that the PMM and KNN are more plausible based on discriminating the recurrence and nonrecurrence cases.
We observe (Table 3) that by, using a MAR mechanism for the 15% missing data, the series mean imputation (SMEAN) yields better classification accuracy via both classifiers; binary logistic regression (BLR) got the top result (70.42%) amongst the three classifiers. For the MCAR mechanism at 15%, the expectedmaximisation via bootstrapping (EMB) imputation method yields best accuracy of 71.60% using linear discriminant analysis (LDA) classifier. At 30% of missing data and MCAR mechanism, the linear discriminant analysis (LDA) via SMEAN achieved best classification accuracy (72.9%) while the BLR produced highest accuracy (71.4%) via SMEAN at 30% with MAR mechanism. Using 45% of missing data, the LDA classifier indicates best classification accuracy (73.9%) under MCAR mechanism, meanwhile, the 60% missing rate reveals that BLR via SMEAN produces best accuracy (72.3%) under MAR mechanism.
Discussion
The article intended to assess classifiers’ accuracy and discriminative ability in classifying breast cancer recurrence using simulation approach based on imputation methods. Before assessing the accuracy and discrimination of classifiers, we used the average RMSE to assess the performance of imputation methods used to fillin missing data generated through simulation of breast cancer datasets under various percentages (15, 30, 45 and 60%) of missing data and missing mechanisms (MAR and MCAR). It was observed that the performance of imputation approaches were decreased with increased missing percentages since the values of RMSE are direct proportion to percentages of missing data under both missing data mechanisms; MAR and MCAR. In addition, the KNN and PMM methods performed better at low and high percentages of missing data for both MAR and MCAR mechanisms. This finding is in line with Javadh et al. [38] and Kleinke at al [39].
It was observed that the performance of imputation approaches were decreased with increased missing percentages since the values of RMSE are direct proportion to percentages of missing data under both missing data mechanisms; MAR and MCAR. In addition, the PMM method performed better at low and high percentages of missing data for both MAR and MCAR mechanisms.
The assessment of performance of classification of three classifiers; binary logistic regression (BLR), linear discriminant analysis (LDA), and quadratic discriminant analysis (QDA) across 15, 30, 45 and 60% of missing data and missing mechanisms (MAR and MCAR) from each applied imputation method (Table 3). The highest classification accuracy (73.9%) was achieved via LDA based on meanimputed data (at 45% missingness with MCAR mechanism), and the minimum was 64.8% from BLR (at 30% missingness with MAR mechanism). This finding implies that the best classifier is LDA under MCAR data. This observation agree the information revealed in Fig. 1 where it was worth noting that the MCAR mechanism provided much alike imputed values than MAR and hence making sense when the mechanism lead to slight better classification accuracy via LDA. However, Ghorbani and Desmarais [40] claim that the performance of imputation techniques considerably changes between various classifiers under different percentages of missing data while EMB gave best classification accuracy across all percentages of missing data.
The binary logistic regression was applied to compare the discriminative ability (how observations from ‘recurrence group’ are separated from ‘nonrecurrence’ group) through the magnitude of area under the ROC across 15, 30, 45 and 60% of missing data and missing mechanisms, MAR and MCAR from each applied imputation method (Table 2). The model revealed that knearest neighbour (KNN) and predictive mean matching (PMM) imputations provide close maximum area under ROC curves, 0.6428 (at 60% missingness with MAR mechanism) and 0.6418 (at 30% of missing data with MCAR mechanism) respectively. A study [41] demonstrated the ability to predict or discriminate the recurrence from nonrecurrence cases of breast cancer prior to ‘neoadjuvant chemotherapy’ treatment yield area under ROC curve of 75%; a difference of about 11% compared to ours. This implies that the imputed data via KNN and PMM are most plausible to enable classifiers to separate breast cancer recurrence cases for at least 64%.
Conclusions
Based on simulation study under various percentages and mechanisms of missing numerical data, the article’s conclusions are focused on three parts. First, a predictive mean matching is most plausible imputation methods for numerical variables; second, the linear discriminant analysis provides better classification accuracy over binary logistic regression and quadratic discriminant analysis for dichotomous response variable; third, logistic regression model is able to separate or discriminate breast cancer recurrence from nonrecurrence cases for at least 64%. The future breast cancer classificationrelated studies may focus on the of imputing longitudinal survival truncated data prior to classification of breast cancer recurrence.
Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Abbreviations
 BSA:

Body Surface Area
 BMI:

Body Mass Index
 EMB:

ExpectationMaximization via Bootstrap
 MAR:

Missing At Random
 MICE:

Multiple Imputations by Chained Equations
 MCAR:

Missing Completely At Random
 NMAR:

Not Missing At Random
 ROC:

Receiver Operating Characteristics
 KNN:

KNearest Neighbour
 VIM:

Visualization and Imputation of Missing Values
 PMM:

Predictive Mean Matching
 R:

R statistical software
References
Nekouie A, Moattar MH. Missing Value Imputation for Breast Cancer Diagnosis Data Using Tensor Factorization Improved by Enhanced Reduced Adaptive Particle Swarm Optimization Atefeh Nekouie Cancer refers to a disease in which a group of cells show uncontrolled growth , invasion . J King Saud Univ  Comput Inf Sci [Internet]. 2018; Available from: https://doi.org/10.1016/j.jksuci.2018.01.006.
Humphries M. Missing Data & How to Deal: an overview of missing data. Popul Res Cent [Internet] 2013;45. Available from: http://www.texaslonghornsl.com/cola/centers/prc/_files/cs/MissingData.pdf
de Goeij MC, van Diepen M, Jager KJ, Tripepi G, Zoccali C, Dekker FW. Multiple imputation: dealing with missing data. Nephrol Dial Trans. 2013;28(10):2415–20.
Zhang Z. Missing data imputation: focusing on single imputation. Ann Transl Med. 2016;4(1). https://doi.org/10.3978/j.issn.23055839.2015.12.38.
Iren M, Tokle R. Comparison of Missing data imputation methods for improving detection of obstructive sleep apnea; 2017.
Little RJ, Rubin DB. Statistical Analysis with Missing data: Willey; 1987.
Curley C, Krause RM, Feiock R, Hawkins CV. Dealing with missing data: A comparative exploration of approaches using the integrated city sustainability database. Urb Aff Rev. 2019;55(2):591–615.
Alruhaymi AZ, Kim CJ. Study on the Missing Data Mechanisms and Imputation Methods. Open J of Stat. 2021;11(4):477–92.
Luengo J, García S, Herrera F. On the choice of the best imputation methods for missing values considering three groups of classification methods. Know and Inform Sys. 2012;32(1):77–108.
Jerez JM, Molina I, Subirats JL, Franco L. Missing data imputation in breast cancer prognosis. Survival. 2006;8(9):1.
Hallgren KA. Conducting simulation studies in the R programming environment. Tutor In Quan Meth For Psychol. 2013;9(2):43.
Morris TP, White IR, Royston P. Tuning multiple imputation by predictive mean matching and local residual draws. BMC Med Res Methodol. 2014;14(1):1–3.
Jerez JM, Molina I, GarcíaLaencina PJ, Alba E, Ribelles N, Martín M, et al. Missing data imputation using statistical and machine learning methods in a real breast cancer problem. Arti Intell In Med. 2010;50(2):105–15.
Pazhoohesh M, Pourmirza Z, Walker S. A comparison of methods for missing data treatment in building sensor data. In: In2019 IEEE 7th International Conference on Smart Energy Grid Engineering (SEGE), vol. 12. Oshawa, ON, Canada: IEEE; 2019. p. 255–9.
Hendriksen J, Geersing G, Moons KG, H GA. Diagnostic and prognostic prediction models. J of Throm and Haemos. 2013;11:129–41.
Burson et al. NIH public access. Bone [Internet] 2014;23(1):1–7. Available from: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3624763/pdf/nihms412728.pdf.
Song WJ, Kim KI, Park SH, Kwon MS, Lee TH, Park HK, et al. The risk factors influencing between the early and late recurrence in systemic recurrent breast cancer. J of Br Can. 2012;15(2):218–23.
James G, Witten D, Hastie T, Tibshirani R. An introduction to statistical learning with applications in R [internet], vol. 102: Design. Springer; 2014. p. 618. http://books.google.com/books?id=9tv0taI8l6YC
Casella G, Fienberg S, Olkin I. An introduction to statistical learning with applications in R: Springer Texts in Statistics; 2014.
Agrest A. Categorical data Analysis. Second Edi: Willey; 2002.
Johnson R, Wichern D. In: Recter P, Hoag C, Ryan D, editors. Applied multivariate statistical Analysis. 6th ed. New Jersey: Pearson Education, Inc.; 2007.
Xiaozhou Y. Linear Discriminant Analysis, Explained : Towards Data Science [Internet]. 2020 [cited 2021 Aug 24]. Available from: https://towardsdatascience.com/lineardiscriminantanalysisexplainedf88be6c1e00b
James G, Witten D, Hastie T, Tibshirani R. In: Casella G, Fienberg S, Olkin I, editors. An introduction to statistical learning with applications in R: Springer Texts in Statistics; 2014.
Roussas G. Some Generalizations to k Random Variables, and Three Multivariate Distributions. Academic Press. 2014;179–199. https://doi.org/10.1016/B9780128000410.000092.
Tacq J. Multivariate normal distribution. International Encyclopedia of Education. 2010;332–8. https://doi.org/10.1016/B9780080448947.013518.
Ripley B, Venables B, Bates DM, Firth D, Hornik K, Gebhardt A. Support Functions and Datasets for Venables and Ripley’s MASS. 2018 [cited 2022 Jan 17];169. Available from: http://www.stats.ox.ac.uk/pub/MASS4/
Schouten RM, Lugtig P, Vink G. Generating missing values for simulation purposes: a multivariate amputation procedure. J of Stat Com and Sim. 2018;88(15):2909–30. https://doi.org/10.1080/00949655.2018.1491577.
Glas CA. Imputation methods. Int Encycl Educ 2010;(Third Edition).
Andridge RR, Little RJ. A review of hot deck imputation for survey nonresponse. Int Stat Rev. 2011;78(1):40–64. https://doi.org/10.1111/j.17515823.2010.00103.x.
Kowarik A, Templ M. Imputation with the R Package VIM. J of Stat Soft. 2016;20(74):1–6.
Beretta L, Santaniello A. Nearest neighbor imputation algorithms : a critical evaluation. BMC Med Inform Decis Mak [Internet]. 2016;16(Suppl 3). https://doi.org/10.1186/s129110160318z.
Van Buuren S, Oudshoorn K. Flexible multivariate imputation by MICE. Leiden: TNO; 1999.
Van Buuren S, GroothuisOudshoorn K. Mice: Multivariate imputation by chained equations in R. J of Stat Soft. 2011;12(45):1–67.
Akmam EF, Siswantining T, Soemartojo SM, Sarwinda D. Multiple Imputation with Predictive Mean Matching Method for Numerical Missing Data. In: In2019 3rd International Conference on Informatics and Computational Sciences (ICICoS), vol. 29. Semarang, Indonesia: IEEE; 2019. p. 1–6.
Bailey BE, Andridge R, Shoben AB. Multiple imputation by predictive mean matching in clusterrandomized trials. BMC Med Res Methodol. 2020;20(1):1–16.
Takahashi M. Multiple ratio imputation by the EMB algorithm: Theory and simulation. J of Mod App Stat Method. 2017;16(1):34.
Do CB, Batzoglou S. What is the expectation maximization algorithm? Nat Biotech. 2008;26(8):897–9.
Javadi S, Bahrampour A, Saber MM, Garrusi B, Baneshi MR. Evaluation of four multiple imputation methods for handling missing binary outcome data in the presence of an interaction between a dummy and a continuous variable. J of Prob and Stat. 2021;2021:6668822. https://doi.org/10.1155/2021/6668822.
Kleinke K. Multiple imputation under violated distributional assumptions: a systematic evaluation of the assumed robustness of predictive mean matching. J Educ Behav Stat. 2017;42(4):371–404.
Ghorbani S, Desmarais MC. Performance comparison of recent imputation methods for classification tasks over binary data. Appl Arti Int. 2017;31(1):1–22 https://www.tandfonline.com/action/journalInformation?journalCode=uaai20.
RabinoviciCohen S, Fernández XM, Grandal Rejo B, Hexter E, Hijano Cubelos O, Pajula J, et al. Multimodal prediction of fiveyear breast Cancer recurrence in women who receive Neoadjuvant chemotherapy. Cancers (Basel). 2022;14(16):3848.
Acknowledgements
This article is a part of the requirement for the PhD program at the University of Dar es Salaam.
Funding
No funding was received.
Author information
Authors and Affiliations
Contributions
R.A.A, A.S.M: study concept and design^{.} R.A.A: data collection, analysis and manuscript development. A.S.M: active manuscript revision and editing. All authors read and approved the final manuscript.
Corresponding author
Ethics declarations
Consent of publication
Not applicable
Ethics approval and consent to participate
University of Dar es Salaam Research Ethics Committee (UDSMREC) issued the ethical approval for the study. The need for informed consent was waived by the Institutional review board of Muhimbili National hospital and Ocean Road Cancer Institute because the data were extracted from the Patient files.
Competing interests
The authors declare that they have no competing interests.
Additional information
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. 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 in a credit line to the data.
About this article
Cite this article
Abassi, R.A., Msengwa, A.S. Classification of breast cancer recurrence based on imputed data: a simulation study. BioData Mining 15, 30 (2022). https://doi.org/10.1186/s13040022003168
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13040022003168
Keywords
 Classification accuracy
 Imputed data
 Missing data mechanisms
 Missingness percentages
 Simulation