Adaptive boosting model for breast cancer prediction

Abstract

The aggressive form of cancer commonly in breast cells is breast cancer. The highly aggressive form of cancer is frequently created in breast cells. The need for the predictive model to accurately measure the prognosis prediction of breast cancer in the earlier stage is highly recommended. This development of methods for protecting people from fatal diseases by the researchers from the different disciplines who are all working altogether. An accurate breast cancer prognosis prediction is made by using a good predictive model to assist Medical Internet of Things (mIoT). Various advantages such as cancer detection in an earlier stage, medical expenses related to treatment, and having unwanted treatment gives the accurate prediction attains spare patients. Existing models lie on the uni-modal data such as chosen gene expression to predict the model’s design. Few learning-based predictive models are used in the proposed method to improve breast cancer prognosis prediction from the current data sets. Most of the peculiar benefits of the suggested method rely on the model’s architecture. Here, a novel adaptive boosting model (a-BM) is used to measure the loss function of every individual and intends to reduce the error rate. Various performances metrics are used to evaluate the predictive performance, which provides the model gives a good outcome rather than the previous techniques.

Keywords

Machine learning breast cancer prediction rate loss function error rate

1 Introduction

Uncontrolled development of cells which are occurred in the breast causes Breast Cancer. The development forms the sheets of cells, termed tumour or a mass. The tumours that are caused cancer are classified into two categories. They are (i) benign and (ii) malignant [1]. Benign tumours can develop, but they cannot be spread over other body parts. On the other hand, malignant tumours can spread and grow. There are two classifications of breast cancer. They are (i) non-invasive and (ii) invasive. Non-invasive breast cancer [1] is restricted to breast’s lobules or milk ducts. At the same time, invasive breast cancer [2] is distributed over nearby tissues in the body. Lobular carcinoma or ductal carcinoma that develops at the lobes or in the ducts is concerned like an incredibly aggressive type of cancer that turns into considerable health issues in females. It is one of the essential generators of cancer-oriented mortality globally [3]. Cancer.Net is a website under ASCO that provides the proposed data that 2,66,120 women have invasive breast cancer. On the other hand, 63,960 women in the US have non-invasive breast cancer.

In 2018, because breast cancer has been computed in the USA, about 41,000 deaths were recorded with 1.18% men and 98.82% women. At the non-invasive stage, about 62% of breast cancer cases are identified, and 99% of the cases with certain survival rate. When the breast cancer is distributed to the local lymph nodes, the survival rate is 85% in the case of invasive breast cancer. When cancer is distributed to the body parts, it is 27%. Because of the complications and the many clinical results [4], doctors’ prediction and treatment of breast cancer cases are challenging. Predicting survival chances gives a more accurate predictive model that assists patients affected by breast cancer. It can help doctors consider relevant patients’ therapy decisions [5].

The expectancy of life of the cancer patients is categorized as short-time survivors, i.e. such with lower than 5-years survival and the long-time survivors with greater than it [6]. Doctors need to advise patients to have the personal treatments for cancer and spare them from unwanted suffering and unwanted extra therapy from the toxic side effects using predictive models if the patients are expected to be the short time survivor [7]. Multiple sources have information on breast cancer like the clinical data like early mensuration, age, pregnancy timing, late menopause, factors of lifestyle, and so on and generic information like the gene expression and the variation data [8]. The integrated multi-modal data can enhance models’ predictive power [9] –[10]. However, the existing methods failed to effectively measure the model’s loss function. This work concentrates on modelling prognostic prediction model. It assists in measuring the survival rate by examining the disease progression of breast cancer. The model intends to model a robust and excellent prediction model for real-world conditions. The anticipated model is an empowered boosting model using the adaptive approach. Decision tree plays a base classifier where the proposed models prediction is made of training set and the input weight is revised based on the error rate. Then, the classification is trained with the updated weight and makes prediction with the training set. Thus, the output gives the remarkable reduction in error. To be specific, the machine learning algorithm is anticipated and tuned for effectual and quick learning process. This process is performed for authorizing the model using the adaptive boosting model. One of the finest ability of the anticipated model is that it can attain better results in shorter period of time without the use of GPU. The proposed model establishes a relationship among the input and the output classes of the dataset. The following are the significant research contributions:

An online available breast cancer dataset is taken for evaluation purposes. Some preliminary pre-processing steps are performed to eliminate the outliers over the dataset;

The features of the provided input dataset are analyzed with the feature learning process where the most dominant features are considered for successive analysis;

Finally, a novel approximated boosting model (a - BM) is proposed to examine the loss rate and enhance prediction accuracy. The simulation is done in the MATLAB 2020a environment, and various metrics are evaluated and compared with the anticipated model. The proposed model is intended to assist mIoT.

The work is structured as follows: Section 2 provides a comprehensive analysis of various prevailing approaches; Section 3 gives the methodological analysis for measuring the survival rate of breast cancer patients. The numerical outcomes are acquired from the anticipated model and compared with existing approaches. The summary is provided in Section 5 with future research improvements.

2 Related works

Some gene expression of patterns is suggested in the previous research on breast cancer for understanding the breast cancer molecular signature. The predictive model is indicated with the help of gene expression profiles [11]. There are 98 primary patients with gene expression having breast cancer, and the 70 gene signatures are chosen using the supervised classification techniques. The validation of results has the unaware data of nineteen young patients with breast cancers. The hybrid signature is used with three gene markers. There are two clinical markers to be created from clinical features. The genetic signatures as 70 concerned the multi-modality by Sun et al. to predict breast cancer prognosis [12]. Al et al. [13] propose the combination of genomic data and clinical data for the prognosis of lymph-node negative breast cancer based on Bayesian Network (BN). These studies provide that many existing systems are inferred that the profile of gene expression is conductive. Yet, some research is first related to the assumption that the specific patient has various genes that are not concerned with one another.

The fact is that various genes of the specific patient have many relationships among them. Different tasks such as the practical selection of feature technique are suggested by Porter et al. [14] depending on the support vector machine (SVM), and the Yoon et al. [15] indicate the selection of feature joined with the random forest (RF) which performs well than the existing 70 gene signatures concerning the breast cancer prognosis prediction. Hazra et al. [16] suggest a probabilistic graphical model (PGM) that integrates two microarray data’s independent models like the higher dimensional data has around 25 000 genes per patient. Moreover, the principal component analysis (PCA) reduces the dimensionality of the data and establishes the deep network for extracting the features that represent the data. The structure learning algorithm is utilized for medical information.

Better performing models have resulted from all these approaches. Yet, these models are combined directly from various data modalities such as clinical and gene expression without considering that various modalities can have multiple representations of the feature. Recently, the enhancement of deep learning methods provides the model with different input data sources of modalities that outperforms the model with one input source data. Few studies have the fact to be validated regarding the diagnosis and prognosis of breast cancer that depends on the multi-modal data [17, 18]. In the same way, the theoretical MKL (Multiple Kernel Learning) methods called HI-MKL [19] are presented. by integrating the multi-omics data and the histopathological image from the prognosis prediction of Glioblastoma Multiforme (GBM). The research need for survival analysis is considered and the histopathological images are available and termed as CapSurv [20] is designed. The technique depends on the function of new loss is called the loss of survival, which is mainly developed for the analysis of survival of the cancer patients. The superiority is proved using multi-modal deep learning in the areas of bioinformatics [21], multimedia analysis [22], computer visual recognition [23], sentiment analysis [24], and speech recognition [25]. Some models depend on the one source input data that affects the non-universal death, noisy and specific data. On the other hand, the process of combining the needed information from different sources overcomes all the restrictions using the multi-modal models [26] and [27 –30].

3 Methodology

The data materials have 5246 patients and the records are normalized to original records in the MP4Ei framework depending on the medical expert’s suggestions. The stratified random sampling approach primarily divides the information of 5246 patients into test data as 1050 cases and train data as 4196 cases. Twenty-three major features are chosen from the original 89 patient features using the training data in the stratified feature selection process having the two stages. Statistical Feature Selection (SFS) is the first stage. Eighty-nine patient features are chosen from the data materials needed to test statistically, depending on every feature type. Ensemble Feature Selection (EFS) is the second stage. Moreover, there are 51 features or attributes are chosen. These attributes are considered statistically and affect the a-BM ending point. Twenty-three essential features with important scores are higher than the 0.016 thresholds created in the selection from 51 features in the stratified feature selection second stage to increase the application of the model and improve the MP4Ei’s stability. The features’ importance score is computed by taking the average of importance scores which are output using the XGBoost and gradient boosting decision tree algorithm after considering the input as fifty-one features. The K-fold cross-validation is repeated many times on the training data. Then, again 23 attributes are provided to the XGBoost algorithm having the tuning parameter of Bayesian and building the 5-year a-BM classifier using the K-fold cross-validation. Lastly, a powerful and excellent model is built for predicting the a-BM for breast cancer patients at an earlier stage. At the same time, 1050 cases are considered the testing data for evaluating the a-BM model after considering the 23 features from the 89 original features. These attributes are similar to the selected one in the procedure of training data.

3.1 Pre-processing

Due to the vast amount of abnormal, missing, inconsistent, and duplicate data, many obstacles in the medical breast cancers are run-up in the natural environment. Moreover, the relevant metrics are taken to confirm the data for minimizing the data issues. The telephone revisiting and rechecking of patients are used to make the corrections with obvious outliers, data inconsistencies have happened in the raw data, and the essential data is missing. At the same time, the multiple records of the patient are merged for making the patient related to the final description due to every patient having different therapeutic, diagnostic, or pathological records. The features of the patient and the related integration rules are used forward in the process using the breast cancer professional medical group.

Moreover, the database system performs the particular establishment for data consistency. The outcomes are analyzed and determined for these issues in the cleaning rules and the consolidated data. These iterative processes ensure data quality. The numeric coding is used to obtain the data to model and encode the categorical variables in the patient’s features. On the other hand, numerical coding encodes some continuous variables classified into the discretization intervals depending on the professional medical group recommendations. There is no history of breastfeeding for the female patients who are not married and encode the variable using the maximum classification values such as nine or ten correspondingly when there are data gaps or a few particular variables are not involved.

3.2 Feature learning

A set of 89 variables have the diagnosis, demographics, therapy, and pathology included in the patient’s description. This model predicts the a-BM ending point variable if this occurs within five years. A stratified feature selection technique is utilized for choosing the 23 essential features that are integrated to follow the work for building the a-BM model for making the model more practicable and more flexible. The stratified feature selection happened on the trained data. The statistical technique is performed on the patients having or not having the a-BM to evaluate the features that are not dependent on the a-BM target variable to improve the model’s stability and reduce the irrelevant features impacts. However, repeated 10-fold cross-validation is performed five times to accurately score on averaging for every essential feature called XGBoost algorithm having the optimized hyper-parameter in the trained data. Moreover, the cut-off value is found by using the subsets of features for evaluating the feature score, which is essential or not needed using step by step backward selection. Suppose every feature is a single factor that has a considerable effect on the result of a variable by using the SFS as the statistical method. Generally, Kolmogorov-Smirnov (KS) statistical test [19] is used for the interval scale feature, which needs to obey the normal distribution. Moreover, the independent sample T-test is used for the feature having the considered distribution on the 5-year a-BM separated from others.

A Wilcoxon Mann-Whitney statistical test is utilized for testing the feature simultaneously, having the ordinal scale, yet the distribution is not distributed normally [20]. The Chi-square test is used for analyzing the statistical significance of the Nominal scale feature. Fifty-one features are chosen from eighty-nine original variables after the SFS to select features in EFS. The stable variable score is obtained by repeatedly running the XGBoost Classifier to get the average score for determining every feature’s importance in Function xgb-scores. Three methods, gain, cover, and weight, are utilized to calculate the feature’s importance score. XGBoost, which has the average gain of splits and the weight is the count of features that appear in the tree, and the number of samples suffered using the splits [16] is defined as the coverage. It is used for the importance scores of output for every feature using the 10-fold cross-validation is performed five times even though there is a critical reference factor called the gain methodology for the splits. The average of 5 scores is used to calculate the final importance score of the feature. The sequential backward selection technique evaluates the vital factor, the 5-year a-BM, after obtaining the final importance score of 51 features when the importance score is higher than 0.016. The process of getting the importance threshold is 0.016. The author in [21] explained the recursive feature elimination algorithm. The features are considered to model in the proposed system having the predictability and applicability that helps for impelling in the proposed method for choosing the target features for the model having (i) compared with the traditional outstanding models, there is an acceptable loss, (ii) count of final features are considered as much as there is a possibility. Moreover, the stability is assured by suggesting the EFS evaluate the essential attributes and reduce the complexity of the tuning parameters. The features are removed having the less average importance score after selection. In addition, if a feature presents a similar meaning to others precisely, only one feature is retained. Consider that mlctp1 and mlctp2 follow different standards, yet all the features give molecular typing. It is avoided since mlctp2 has the lower average importance score. Lastly, for the earlier stage of breast cancer, 23 essential features are chosen from the 5-year a-BM model.

Pearson correlation coefficient is used to measure the correlation among two features. The overall features list is provided in Table 1 based on the variables category. Features are selected before having Nominal (N) scale variables with 75% after having Nominal (N) scale variables of 48%. Scale variables of 21.3%, O scale variables of 43.5%, Interval scale variables of 3.4%, and I scale variables of 8.7%. At the same time, there are demographic variables of 12.4% and demo variables of 21.7%, Pathology variables of 13.5% and path variables of 17.4%, therapy variables of 26.9% and the Variables of (21.7%), Diagnosis variables of 47.2% and diag variables of 39.2%. Existing breast cancer risk prediction methods facilitates targeting and identification of women at higher risk by diminishing the interventions at lower risk. Various breast cancer risk prediction approaches are employed in clinical practices which however show lesser prediction accuracy on a range of 55% to 65%. To resolve this issue, machine learning provides a solution with standard prediction approaches which acts as a clinical decision support system during the time of emergency by addressing the present constraints and enhances the prediction accuracy with the available tools. The motivation of this research is to provide a better CDSS for enhancing the discriminatory prediction accuracy with the approximation model and to help the experts during the time of complexity. A stratified feature selection finds and eliminates the unwanted and redundant features which are not helped for improving a-BM models’ performance, and the model’s accuracy is reduced. Notably, the complexity is reduced by fewer features, making it more accessible for the model to be interpreted and understood [22]. However, some existing methods like Support Vector Machine (SVM) gives 75% prediction accuracy, stacked SVM gives 76% accuracy, hybrid SVM gives 77%, Random forest (RF) gives 76%, stacked RF gives 85%, XGboosting gives 85%, adaboosting gives 83% and ensemble learning gives 83% accuracy which is substantially lower for the provided dataset. However, this research concentrates on modeling an efficient approximation based boosting model for breast cancer prediction.

Table 1
Feature analysis

Features (from the provided dataset)

Demographics Diagnosis Pathology Theraphy %

Original Features (dataset based features) Nominal Sex addr nat edu mar medi urb meno menoR bearF feed ocs hrt adt adtT fhBC fhGC fhC hisF cmbdt cmUN cmCC cmCH cmMS cmBT cmNE cmRS cmDS cmBS cmCS ER PR HER2 Ki67 CK56 hrmRP mlctp1 mlctp2 sur surT chem chemS chemB chemA radN1 radN2 rad radT surCT druCT edcN edc edcBS smBS aiBS edcAS edcST edcS1 edcS2 traN tra traS 75.3

Ordinal Age M bmi mediR ageMR ageMU age MP preg bearN pregB ageB1 ageB2 ageBM ageME T N as ps Mstp who – 21.3

Interval Aged – ND1 ND2 – 3.4

% 12 47.2 13.5 26.9 –

Selected features (chosen features from the dataset for validation) Nominal Edu medi mediR – Ki67 CK56 mlctp1 chemB chemA radT edcST edcS1 47.8

Ordinal BMI bearN ageB1 ageMP ageBM ageME T N as ps – – 43.5

Interval Aged – ND1 – 8.7

% 21.7 39.2 17.4 21.7 –

Features (from the provided dataset)
		Demographics	Diagnosis	Pathology	Theraphy	%
Original Features (dataset based features)	Nominal	Sex addr nat edu mar medi urb	meno menoR bearF feed ocs hrt adt adtT fhBC fhGC fhC hisF cmbdt cmUN cmCC cmCH cmMS cmBT cmNE cmRS cmDS cmBS cmCS	ER PR HER2 Ki67 CK56 hrmRP mlctp1 mlctp2	sur surT chem chemS chemB chemA radN1 radN2 rad radT surCT druCT edcN edc edcBS smBS aiBS edcAS edcST edcS1 edcS2 traN tra traS	75.3
	Ordinal	Age M bmi mediR	ageMR ageMU age MP preg bearN pregB ageB1 ageB2 ageBM ageME T N as ps	Mstp who	–	21.3
	Interval	Aged	–	ND1 ND2	–	3.4
%	12	47.2	13.5	26.9	–
Selected features (chosen features from the dataset for validation)	Nominal	Edu medi mediR	–	Ki67 CK56 mlctp1	chemB chemA radT edcST edcS1	47.8
	Ordinal	BMI	bearN ageB1 ageMP ageBM ageME T N as ps	–	–	43.5
	Interval	Aged	–	ND1	–	8.7
%	21.7	39.2	17.4	21.7	–

3.3 Deciding random survival

The Deciding Random Survival (DRS) is an improved version of the existing Random Forest (RF), which adopts a log-rank for partitioning the trees (survival) and evaluates the function related to the tree’s terminal node. It performs cumulative functioning associated with the terminal nodes attained from the estimator. For every individual, the DRS evaluate the cumulative functioning by averaging the overall nodes statistics of the tree.

3.4 XGBoost

XGBoost utilizes a precise learning objective approximation and includes regularization terms for eliminating the over-fitting issues effectively. It anticipates a novel approach (See Figs. 1 2) for tree node partitioning and node evaluation. The regularized objective for t–the iteration is provided in Equation (1):

Fig. 1

a-BM model architecture.

Fig. 2

Flow diagram.

$L^{(t)} = \sum_{i = 1}^{n} l (y_{i}, {\hat{y}}_{i}^{(t - 1)} + F_{t} (x_{i})) + Ω (F_{t})$ (1)

The second-order approximation is utilized to optimize the objective function quickly and it is expressed in Equation (2) where $g_{i} = \frac{l (y_{i}, {\hat{y}}^{(t - 1)}}{\partial {\hat{y}}^{(t - 1)}}$ and $h_{i} = \frac{\partial^{2} (y_{i}, {\hat{y}}^{(t - 1)}}{\partial {\hat{y}}^{(t - 1)}}$ are gradient statistics (1st and 2nd). The provided boosting model predicts the node partitioning that reduces the loss function and computes the terminal node estimation while fitting the tree at every iteration (M specifies total iterations). The final node includes the evaluated quantities of every regression tree over the terminal node to identify the log ratio with Equation (2):

${\hat{y}}_{i} = f (x_{i}) = \sum_{t = 1}^{M} F_{t} (x_{i})$ (2)

The existing gradient boosting machine (GBM) performs a first-order approximation for loss function and computes the negative gradient (optimization). XGBoost utilizes this precise approximation to attain diverse optimization techniques. The regularization process helps in the effective elimination of over-fitting. It is based on ratio assumption and represents the non-linear functions (complex).

3.5 Proposed approximated boosting model (a - BM)

The proposed a-BM model considers a more precise likelihood function (partial) approximation as a target learning objective. It extracts the gradient expression for the newly attained learning objective. It optimize the existing XGBoost model for various analyses. The expression is customized with loss function based on analysis data and attains the loss function. Here, the likelihood function (partial) is determined as the customized loss functions and derives gradient expression. Similarly, a theorem for gradient derivation is proposed to specify the core optimization.

3.5.1 Customized approximation

When the numbers of related events of any point are relatively more significant, the given Efron approximation can provide a suitable approximation related to the threshold value. The loss function with the customized approximation is expressed in Equation (3): $L = \prod_{t \in D} \frac{e^{\sum_{j \in q (t)} {\hat{y}}_{j}}}{\prod_{l = 1}^{C_{t}} [\sum_{j \in R (t)} e^{{\hat{y}}_{j} - \frac{l - 1}{C_{t}} \sum_{j \in q (t)} e^{{\hat{y}}_{j}}}]}$ (3)

Here, $SR (t) = \sum_{j \in R (t)} e^{{\hat{y}}_{j}}$ represents the sum of the evaluated hazard ratio considered to be at risk at time t, and $SD (t) = \sum_{k \in q (t)} e^{{\hat{y}}_{j}}$ specifies the sum of the evaluated hazard ratio of every individuals’ whose events are traced at t time. Consider negative logarithms on both sides of the Equation (3) and acquire the loss function. $L_{E} = \sum_{t \in D} {\sum_{j \in q (t)} ln (SR (t) - w_{j} * SD) t) - {\hat{y}}_{j}]}}$ (4)

Here, $W_{j} = \frac{l - 1}{C_{t}}, l = 1, \dots, C_{t}$ specifies the allocated unique weight for every individual at q(t). Two lemma’s are derived from measuring the first- and second-order gradients L_E.

Lemma 1: Assume α (t) , β (t) as two functions at time ‘t’ as expressed in Equations (6): $α (t) = \sum_{j \in q (t)} 1 / [SR (t) - W_{j} * SD (t)]$ (5) $β (t) = \sum_{j \in q (t)} W_{j} / [SR (t) - W_{j} * SD (t)]$ (6)

Explain Lemma 2 to derive the second-order gradient form: Assume φ (t) , ω (t) specifies two functions based on ‘t’ time. $φ (t) = \sum_{j \in q (t)} {W_{j} / [SR (t) - W_{j} * SD (t)]}^{2}$ (7) $ω (t) = \sum_{j \in q (t)} [1 - {(1 - W_{j})}^{2}] / {[SR (t) - W_{j} * SD (t)]}^{2}$ (8)

Based on Definition 1, Theorem 1 can be derived.

Theorem 1: For every individual i, the sensed indicator variables δ_i, then the gradients (first order) of L_E based on ${\hat{y}}_{i}$ are expressed as in Equation (9): $g_{i} = {\begin{matrix} e^{\hat{y_{i}}} {[\sum_{t ⩽ T_{i}} α (t)] - β (T_{i})} & if δ_{i} = 1 \\ e^{{\hat{y}}_{i}} \sum_{t ⩽ T_{i}} α (t) & if δ_{i} = 0 \end{matrix}$ (9)

Proof: Based on Equation (6), for every individual i, we derive the gradient-based on individual prediction ${\hat{y}}_{i}$ . However, ${\hat{y}}_{i}$ acquires from SR (t) when the individual i is at risk with t (t ∈ D) time point. Thus, we must consider the L_E where t ⩽ T_i while deriving the individual gradient i . Also, ${\hat{y}}_{i}$ occurs in SD (t) when the individual i is represented as δ_i = 1 and T_i = t . Thus, two different cases are discussed:

1) For every individual i, when the event occurs, i.e. δ_i = 1; L_E is considered two different parts. With the chain rule, the first-order gradient is expressed in Equation (10): $g_{i} | t = T_{i}, δ_{i} = 1 = e^{{\hat{y}}_{i}} \sum_{j \in q (T_{i})} \frac{1 - W_{j}}{SR (T_{i}) - W_{j} * SD (T_{i})} - 1$ (10) $g_{i} | t < T_{i}, δ_{i} = 1 = e^{{\hat{y}}_{i}} [\sum_{j \in q (t)} \frac{1}{SR (t) - W_{j} * SD (t)}]$ (11)

2) For every individual i, when an event is not observed, then δ_i = 0 . The prediction $\hat{y_{i}}$ does not occur at SD (T_i), then the first-order gradient is alike of g_i|t < T_i, δ_i = 1 and expressed in Equation (12): $g_{i} | δ_{i} = 0 = e^{{\hat{y}}_{i}} \sum_{t ⩽ T_{i}} [\sum_{j \in q (t)} \frac{1}{SR (t) - W_{j} * SD (t)}]$ (12)

Based on definition 1, the first-order gradient L_E with ${\hat{y}}_{i}$ is provided in Equation (13): $\frac{\partial L_{E}}{\partial {\hat{y}}_{i}} = g_{i} = {\begin{matrix} e^{{\hat{y}}_{i}} {[\sum_{t ⩽ T_{i}} α (t)] - β (T_{i})} & if δ_{i} = 1 \\ e^{{\hat{y}}_{i}} \sum_{t ⩽ T_{i}} α (t) & if δ_{i} = 0 \end{matrix}$ (13)

Then, based on definition 2 and g_i the second-order gradient are derived in Theorem 2.

Theorem 2: For every individual i, observed indicator variable δ_i, then the gradient of L_E (second-order) with $\hat{y_{i}}$ is expressed as in Equation (14): $h_{i} = {\begin{matrix} g_{i} | δ_{i} = 1 - {(e^{{\hat{y}}_{i}})}^{2} {\sum_{t < T_{i}} (t) - ω (T_{i})} & if δ_{i} = 1 \\ g_{i} | δ_{i} = 0 - {(e^{{\hat{y}}_{i}})}^{2} \sum_{t ⩽ T_{i}} φ (t), & if δ_{i} = 0 \end{matrix}$ (14)

Proof: Based on the above theorem, the first-order gradient of L_E. The second-order gradient of g_i is derived. The below-given cases are discussed for gradient analysis:

1) If δ_i = 1, the second-order gradient of L_E with ${\hat{y}}_{i}$ using the chain rule is given as:

$\begin{matrix} - h_{i} | δ_{i} = 1 = g_{i} | δ_{i} = 1 + 1 - {(e^{\hat{y_{i}}})}^{2} \sum_{j \in q (T_{i})} \\ \frac{(1 - w_{j})}{SR (T_{i}) - W_{j} * SD (T_{i})} - {(e^{\hat{y_{i}}})}^{2} \\ \sum_{t < T_{i} s} [\sum_{j \in q (t)} \frac{1}{{SR (t) - W_{j} * SD (t)}^{2}}] \end{matrix}$ (15)

2) When δ_i = 0, we can acquire,

$\begin{matrix} - h_{i} | δ_{i} = 0 = g_{i} | δ_{i} = 0 - {(e^{{\hat{y}}_{i}})}^{2} \sum_{t ⩽ T_{i}} \\ [\sum_{j \in q (t)} \frac{1}{{[SR (t) - W_{j} * SD (t)]}^{2}}] \end{matrix}$ (16)

The second-order gradient with $\hat{y_{i}}$ is provided more straightforwardly from the above definition. $\frac{\partial^{2} LE}{\partial {\hat{y}}_{i}} = h_{i} = {\begin{matrix} g_{i} | δ_{i} = 1 - {(e^{{\hat{y}}_{i}})}^{2} {[\sum_{t ⩽ T_{i}} - ω (T_{i})] + 1} & if δ_{i} = 1 \\ g_{i} | δ_{i} = 0 - {(e^{{\hat{y}}_{i}})}^{2} \sum_{t ⩽ T_{i}} φ (t) & if δ_{i} = 0 \end{matrix}$ (17)

3.5.2 a - BM execution

The pseudo-code for executing the loss function is given below. For every ‘t’ time, the event occurrence point (t ∈ D) , the value of SD (t) , SR (t), and C_t is evaluated. Then, the loss value is included for every individual with the occurred events in ‘t’ time. At last, the loss function for every model prediction is attained. Based on Theorem 1 & 2, the algorithm for gradient loss function evaluation is provided. Initially, the array A_t is sorted with T time during the survival data and eliminates the duplications. For every t time, the values of SR (t) , SD (t) , φ (t) , β_t, α_t and ω_t for the gradient evaluation need to be evaluated. The gradients are assessed based on the prediction formula provided in the Theorems 1 & 2.

Algorithm 1. Loss function evaluation
Input: Individual prediction (ratio scale); survival data;
Output: loss
1. Loss = 0
2. for t → D
3. Predict $S R_{t} \to \sum_{k \in R (t)} e^{{\hat{y}}_{k}};$
4. Predict $S D_{t} = \sum_{k \in q (t)} e^{{\hat{y}}_{k}};$
5. Predict C_t = \|q (t) \| ;
6. Set l → 1
7. for j → q (t)
8. set W_j = (l - 1)/C_t ;
9. evaluate l = l+ 1 ;
10. compute $loss = loss + ln (S R_{t} - W_{j} * S D_{t}) - {\hat{y}}_{j};$
11. end
12. end
13. return loss

Algorithm 2. Gradient loss function computation
Input: Individual prediction (ratio scale); survival data;
Output: g, h
1. Set A_t→ sort (T) ;
2. Set C₁ = 0 and C₂ = 0;
3. for t of every individual
4. Set $S R_{t} = \sum_{k \in R (t)} e^{{\hat{y}}_{k}}$
5. Set $S D_{t} = \sum_{k \in q (t)} e^{{\hat{y}}_{k}}$
6. Evaluate first-order gradient
7. Evaluate the values of ∝_t, β_t, C₁ ;
8. Patients observed in ‘t’ time;
9. for j → N_t
10. if δ_i = 0
11. Evaluate g_j and h_j ;
12. else if //δ_i = 1
13. Compute g_j and h_j ;
14. end
15. end
16. end
17. return h, g

3.5.3 a - BM performance evaluation

To compute the model evaluation, it is essential to handle the likelihood function approximately. The anticipated model provided the possible way to determine the probable orders of events that occurred for every individual. The above-given expression Q₁ specifies the set of C_t ! For successive C_t events at t time. Similarly, P ={ p₁, p₂, …, p_{c_t} } is an element of Q_t . It consumes huge time when there is a larger number of ties, i.e. C_t is larger, and it is provided in Equation (18): $L = \prod_{t \in D} \frac{e^{\sum_{j \in q (t)} {\hat{y}}_{j}}}{\sum_{P \in Q_{t}} \prod_{t = 1}^{C_{t}} \sum_{j \in R (t) - {P_{1}, \dots, P_{l . . t}}} e^{{\hat{y}}_{j}}}$ (18)

The threshold approximation does not consider the occurrence of every individual with survival time; however, it directly takes the sum of the evaluated hazard ratio as the denominator. It allocates weight to the evaluated hazard ratio of every individual. The approximation with every individual based on the approximation is more significant than several ties. It is outcomes as the worst model. When the survival data has no relationship, the likelihood expressions are reduced.

4 Experimentation

Because of the restricted dataset’s size, the ten-fold CV is performed in the suggested system for the variance problem to be overcome, which is raised for determining the recommended method. In the proposed system, there are 1980 patients classified into ten subsets randomly, and nine subsets are combined and concerned as the training set. On the other hand, the remaining subset is involved with the testing set as one after the other. Further, 20% validation and 80% training sets are classified from the training set. Primarily, the model is trained for every data modality, and the optimal parameters are fixed using the AUC value as the criteria. Secondary, the outputs are extracted to form the feature set. Finally, the features are transferred to the approximated model, the machine learning technique like Random Forest, Support Vector Machine, Linear Regression, and Naïve Bayes. The ROC (Receiver Operating Characteristics) curve is used in the proposed system for evaluating the performance that is plotted against the false positive rate (Specificity) and true positive rate (Sensitivity) using the threshold variation. This proposed system calculates the AUC value using the ROC curve as the model’s efficiency metric. There are evaluation measures like the accuracy (ACC), sensitivity (Sn), specificity (Sp), precision (Pre) and Matthew’s correlation coefficient (MCC) utilized to evaluate the performance that is depicted in the below equation. $Sn = \frac{TP}{Total samples} * 100$ (19) $Sp = \frac{TN}{TN + FP}$ (20) $Pre = \frac{TP}{TP + FP}$ (21) $Acc = \frac{TP + TN}{TP + TN + FP + FN}$ (22) $MCC = \frac{TP * TN - FP * FN}{\sqrt{(TP + FN) * (TP + FP) * (TN + FN) * (TN + FP)}}$ (23)

Here FP, TP, FN and TN stand for the false positive, true positive, false negative and true negative in a confusion matrix.

4.1 Analysis

There are 5246 cases in which earlier phase breast cancer patients are examined and researched. Twenty-three features, the binary outcome variable, and a-BM represent the patient. The performance evaluation of the built model is performed using the Holdout method [23] with the help of XGBoost. 5248 cases are divided into the testing set having 1050 cases and the training set having 4196 cases using the Stratified random sampling method. The considerable difference between the testing set and the training set is not there as p > 0.05. The ROC curve (AUC) to assess the performance of the CRCB dataset. Since there is a balance between sensitivity and specificity of a model, the efficient evaluating index is the AUC. Moreover, the AUC index can reflect the better performance of the model for the prediction at the time of having the data with unbalanced positive and negative samples [24]. It gives very good, good and excellent diagnosis accuracies that are (0.8–0.9), (0.7–0.8), and (0.9–1.0) accordingly [25]. The training set evaluates the performance using 10-fold cross-validation five times. It provides a set of model parameters that can attain a good model compared with existing suggested techniques that make the model have good stability and predictability [27]. The AUC values are tested at various boosting iterations to verify whether the over-fitting is available in the previous system. The complexity of the model is increased with the increasing number of iterations. The training data is fit in the model, having a small error. Yet, the model’s generalizability is reduced by the overall complex model having a more significant error in the testing data. The AUC values increase in both testing and training processes (initial stage) with the increased number of iterations. AUC values decrease during the several iterations tested in the proposed system. The boosting iteration has the optimal number of 110 as n_estimators = 110. On the other hand, the boosting iterations have the best number, 107, in the proposed system. The model’s performance on the prediction is lowered because of the over-fitting issue after the 107 iterations. Hence, this is nearer to the optimal parameters, which are chosen using the optimization algorithm using the early stopping criteria, i.e. 10.

Other machine learning approaches like Adaboost, Random Forest (RF), and Support Vector Machine (SVM) are used to build models compared with the proposed model. Various algorithms are getting trained on a similar training data set and getting tested in similar testing data set. scikit-learn machine learning repository [29] provides the Random Forest, Support Vector Machine, and Adaboost in the experiments. The K-fold cross-validation technique is needed to obtain the optimal parameters, and the Bayesian parameter optimization is utilized as the parameter tuning. Then, SVM model has the parameters are kernel=’RBF’, C = 0.3, class_weight={1 : 2.0, 0 : 1}, gamma = 0.005, and probability = True. The RF model has the parameters which are min_samples_split = 15, max_depth = 14, min_samples_leaf = 2, n_estimators = 120. Then, Adaboost model has the parameters which are n_estimators = 150, and learning_rate = 1.0. The proposed system has 0.8451 for AUC and 0.605 for F-measure, as shown in Table 2. The performance differences between the RF, SVM, proposed system, and the Adaboost are measured as presented in Table 2.

Table 2
Comparison of performance metrics

Model AUC ACC PRE SN MCC

a - BM 0.945 0.95 0.94 0.945 0.632

MDN 0.84 0.82 0.74 0.45 0.48

SVM 0.75 0.83 0.68 0.59 0.53

Stacked SVM 0.76 0.84 0.71 0.60 0.55

Hybrid SVM 0.77 0.85 0.73 0.63 0.58

RF 0.76 0.83 0.68 0.61 0.54

Stacked RF 0.85 0.84 0.77 0.50 0.53

XGboost 0.85 0.84 0.77 0.50 0.54

Adaboosting 0.83 0.84 0.77 0.50 0.53

Ensemble learning 0.83 0.86 0.76 0.49 0.52

Model	AUC	ACC	PRE	SN	MCC
a - BM	0.945	0.95	0.94	0.945	0.632
MDN	0.84	0.82	0.74	0.45	0.48
SVM	0.75	0.83	0.68	0.59	0.53
Stacked SVM	0.76	0.84	0.71	0.60	0.55
Hybrid SVM	0.77	0.85	0.73	0.63	0.58
RF	0.76	0.83	0.68	0.61	0.54
Stacked RF	0.85	0.84	0.77	0.50	0.53
XGboost	0.85	0.84	0.77	0.50	0.54
Adaboosting	0.83	0.84	0.77	0.50	0.53
Ensemble learning	0.83	0.86	0.76	0.49	0.52

The Bayesian method is used after tuning the parameters, which is a model with extraordinary predictive power for predicting 5-year a-BM. The better performance for prediction is achieved (See Figs. 3 6) by having the 0.8451 as AUC that is maximized by 3%. The Youden index is considered the cut-off point, which is more prominent and the cut-off point is obtained as 0.235. On the other hand, the AUC, F-measure, specificity, and sensitivity are 0.8451, 0.605, 0.794, and 0.742. Since the harmonic average between the recall and precision, a particular explanation is needed that the critical metric is considered an F-measure in machine learning. The model’s performance attains worst at F-measure=0 and reaches best at F-measure=1 when compared with Adaboost, SVM, and Random Forest, the F-measure increases by 0.04, 0.037, and 0.01.

Fig. 3

Accuracy evaluation.

Fig. 4

Precision evaluation.

Fig. 5

Sensitivity evaluation.

Fig. 6

MCC evaluation.

One of the most generally used models is the regression algorithm in the survival analysis field, which is varied from the classic classification algorithms described earlier in time and survival status as the variables result that need to predict at the same time. The regression model is built using the training set to be fitted, which is utilized in the existing algorithms after the observation time of the patients. The survival time of 5-year is developed to predict the probability with the testing set. The higher AUC value in the proposed model is compared with the regression model. The considerable difference as p < 0.005 is obtained as the ROC curves between the two models presented in Fig. 7. The higher AUC value is achieved in the proposed model compared with other models from the mentioned experimental results. Since p values are all less than 0.005, the ROC curves among the models attain a considerable difference. The proposed technique is better in an obvious way. Because the regularization items are added in the XGBoost algorithm for penalizing the value and the number of leaf nodes, it is partly better. The model is presented at the over-fitting risk. On the other hand, the traditional algorithms have attained the demerits due to the lack of regularisation. When training the model, it is straightforward to fit well and is not robust enough. When testing the model, it is not good.

Fig. 7

AUC evaluation.

Table 3 shows the performance of the anticipated model with various other datasets like Kaggle, UCI machine learning repository, NKI and UTA4 respectively. The performance of the anticipated model is 95% which is substantially higher than other approaches. The proposed model shows 5%, 4%, 2% and 35% higher accuracy compared to Kaggle, UCI, NKI and UTA4 datasets. From this it is proven that the model works well with the provided dataset.

Table 3

Comparison of performance metrics

Model	CRCB	Kaggle	UCI	NKI	UTA4
	ACC
a - BM	0.95	0.90	0.91	0.93	0.60
MDN	0.82	0.78	0.70	0.43	0.45
SVM	0.83	0.80	0.65	0.55	0.51
Stacked SVM	0.84	0.81	0.64	0.55	0.50
Hybrid SVM	0.85	0.82	0.69	0.60	0.54
RF	0.83	0.83	0.65	0.59	0.51
Stacked RF	0.84	0.83	0.72	0.48	0.52
XGboost	0.84	0.81	0.73	0.45	0.54
Adaboosting	0.84	0.81	0.71	0.43	0.51
Ensemble learning	0.86	0.86	0.70	0.50	0.53

4.2 Feature analysis

More inclination is needed for choosing the model from a practical point of view, having fewer input variables without not having a considerable effect on the system’s performance. The clinical medical data consists of sparse features and some redundant features simultaneously. It is considered excellent for finding and avoiding clinical data. The features are lowered from 89 to 23 through EFS and SFS feature selection. The model’s performance is not affected considerably. Every patient or sample has 89 features and the binary outcome variable. Out of 89 features, 51 features having considerable consequences on the outcome variable as p < 0.05 are chosen through the statistical feature selection (SFS). Hence, the ensemble feature selection (EFS) technique filters the 23 features with greater scores to measure the importance of the variable. The original 89 features help build the models presented with no FS, and using the 23 features selected with EFS contains no considerable difference, such as p = 0.28248 having 0.8486 for AUC. Since the AUC is 0.8446 and the p-value is 0.45857, SFS does not affect the model’s performance. Hence, the optimum model is obtained for 5-year a-BM prediction after selecting EFS and SFS features. The AUC values are 0.8446, 0.8451, and 0.8486 are attained at the various cut-off points according to the models with statistical feature selection (After SFS), having ensemble feature selection (After EFS), and devoid of feature selection (without FS). AUC loss is 0.0035 as 0.4% relative to the model with no feature selection after completing the ensemble feature selection. The F1 loss of 2.7% is lesser at the same time. Hence, the model is affected much less after avoiding some unwanted features using the EFS technique. On the other hand, the EFS and SFS techniques delete many features. The predictability and applicability of the model are significantly improved with the number of features are reduced from 89 to 23.

4.3 Hyper-parameter analysis

This section gives the parameter analysis of the proposed model with a max_depth with a range of [2, 6], and step range is 1. The min_weight is [0.0, 1.0], sub-samples of [0.4, 1.0] and step is 0.1. The lambda value ranges from [0.0, 1.0] and the gamma value is [0.0, 1.0]. Based on these hyper-parameters of the approximation model, the p-value falls into the range to project the models significance, i.e. p = 0.001 . With these parameters, the t-distribution is extremely applicable even for small-sized samples.

4.4 Threshold analysis

After SFS, the essential features are selected from the 51 features mentioned above. The threshold is identified as the cut-off score if the feature is necessary or not, which creates the major part of EFS. Both the predictability and applicability are considered depending on the final importance score, and the backward selection technique is used for evaluating the features having the scores with importance. The seven thresholds are used as the importance score for screening the essential features using the backwards selection step by step. The selected feature set is used for establishing the model, and the parameters are tuned based on the optimization of the TPE parameter. A similar dataset is trained and tested, and the results are obtained. Consider an instance where 10-fold CV is repeated five times when the 0.005 is considered the threshold. When training the model, the average AUC value is 0.8459 is the maximum AUC value. On the other hand, when testing the model, the AUC has a value of 0.8452. The threshold of 0.016 and the loss of AUC is around 0.6%, with fewer features attained. Yet, only three features are minimized. Hence, applicative ability and predictive ability are considered. The score of 0.016 is chosen as the importance threshold in EFS for filtering the features.

4.5 Statistical analysis

The proposed system is used for measuring the breast cancer prognosis and t. It is performed across the a-BM performance measures in the proposed method. The t-distribution is followed by a t-test available for the small-sized sample. Hence the performance metric has the accuracy for a-BM is used. The t-test has a p-value, and the t-value is 0.00 and -14.93. The proposed method has differences in performance, and the previous techniques are considered in a statistical way that gives the outcomes are considerable statistically. The t-test is performed on the validation dataset; the p-value and t-value are 0.00 and 109.54, respectively. The inbuilt library function is used in the proposed system for the test as stats.

5 Conclusion

Millions of death all over the world is because of breast cancer where gene signature complexity and the breast cancers’ gene are some other factors related to it. Hence, this is imperative for designing an efficient and fast model for the prognosis prediction of breast cancer. The learning model is developed in the proposed system having the assistance of a - BM for predicting the expectancy of life of breast cancer patients. This model, like the clinical data, uses multi-modal inputs. The designing is done for various input sources, and features extraction is done for the provided input. The features are passed to extract the features for the final breast cancer survival prediction through the a - BM classifier. The architecture is used for the same diseases that need multi-modal inputs. The model needs to be validated across breast cancer datasets even though the proposed system performs better than other previous prediction models. In this proposed system, the dataset has the 1980 samples only that are very small for machine learning research. Hence, more enhancements are needed in the outcomes when the dataset is more extensive, which is required for mIoT. The integration of the breast cancer tissue images is done as the different modalities in the dataset as the extension of the proposed model. Also, a few modalities are added to the dataset, like miRNA expression values and gene methylation. The major research constraint is the samples over the provided dataset. Generally, the learning models give higher prediction accuracy when the numbers of samples are higher. This can be rectified in the future with the construction of real-time dataset by acquiring data from the Cancer centers and future it can be set a trademark to fix the benchmark standard. Also, while handling with huge samples, deep learning approaches are best suited compared to machine learning approaches. The computational complexity is higher when the samples are huge, while the complexity can be reduced with deep learning approaches with higher accuracy.

References

Sun , Goodison

, Li

, Liu

and Farmerie

, Improved breast cancer prognosis through the combination of clinical and genetic markers, Bioinf 23(1) (2006), 30–37.

Xu , Zhang

, Zou

, Wang

and Li

, A gene signature for breast cancer prognosis using support vector machine, in Biomedical Engineering and Informatics (BMEI), 2012 5th International Conference on. IEEE (2012), 928–931.

Khademi and Nedialkov

N.S.

, Probabilistic graphical models and deep belief networks for prognosis of breast cancer, in Machine Learning and Applications (ICMLA), 2015 IEEE 14th International Conference on. IEEE (2015), 727–732.

Nguyen , Wang

and Nguyen

H.N.

, Random forest classifier combined with feature selection for breast cancer diagnosis and prognostic, Journal of Biomedical Science and Engineering 6(05) (2013), 551.

Gao , Aksoy

B.A.

, Dogrusoz

, Dresdner

, Gross

, Sumer

S.O.

, Sun

, Jacobsen

, Sinha

, Larsson

, et al., Integrative analysis of complex cancer genomics and clinical profiles using the cbioportal, Sci. Signal 6(269) (2013), pl1–pl1.

Zhang , Li

, He

and Wang

, A novel mkl method for gbm prognosis prediction by integrating histopathological image and multi-omics data, IEEE Journal of Biomedical and Health Informatics PP (2019), 1–1.

Sun , Li

, Tang

and Wang

, Integrating genomic data and pathological images to effectively predict breast cancer clinical outcome, , Computer Methods and Programs in Biomedicine 161 (2018), 04.

Sohn , Shang

and Lee

, Improved multi-modal deep learning with variation of information, in Advances in Neural Information Processing Systems (2014), 2141–2149.

Wang and Han

, Multi-modal biometric authentication based on score level fusion using support vector machine, Optoelectronics Review 17(1) (2009), 59–64.

10.

Ju , Bibaut

and van der Laan

M.J.

, The Relative Performance of Ensemble Methods with Deep Convolutional Neural Networks for Image Classification, arXiv e-prints, p. arXiv:1704.01664, Apr. 2017.

11.

Kumar , Gupta

, Chandra

S.S.

, Raman

and Channappayya

S.S.

, No-reference quality assessment of tone mapped high dynamic range (hdr) images using transfer learning, in 2017 Ninth International Conference on Quality of Multimedia Experience (QoMEX) (2017), 1–3

12.

Pandya , Park

, Chen

, Goodell

L.A.

and Foran

D.J.

, Toward a portable cancer diagnostic tool using a disposable MEMS-based biochip, IEEE Trans. on Biomedical Engineering 63(7) (2017), 1347–1353.

13.

A1-Badrashiny and Bellaachia

, Breast cancer survivability prediction via classifier ensemble, Inter. J. of Computer and Information Engineering 10(5) (2016), 833–837.

14.

Porter , Coates

and Popovic

, An early clinical study of time-domain microwave radar for breast health monitoring, IEEE Trans. on Biomedical Engineering 63(3) (2016), 530–539.

15.

Yoon , Davtyan

and Schaar

, Discovery and clinical decision support for personalized healthcare, IEEE J. of Biomedical and Health Informatics 21(4) (2017), 1133–1145.

16.

Hazra , Bera

and Mandal

, Predicting lung cancer survivability using SVM and logistic regression algorithms, Inter. J. of Computer Applications 174(2) (2017), 19–24.

17.

Mahrooghy , Ashraf

A.B.

, Daye

, et al. Pharmacokinetic tumour heterogeneity as a prognostic biomarker for classifying breast cancer recurrence risk, IEEE Trans. on Biomedical Engineering 62(6) (2015), 1585–1593.

18.

Wu , Burnside

E.S.

, Cox

, Fan

, Yuan

and Yin

, Breast cancer risk prediction using electronic health records, in IEEE Inter. Conf. on Healthcare Informatics (2017), 23–26.

19.

Wang , Makonde

, Chen

K.H.

and Wang

K.M.

, A hybrid classifier combining SMOTE with PSO to estimate 5-year survivability of breast cancer patients, Applied Soft Computing Journal 20(7) (2014), 15–24.

20.

Thornton , Hutter

, Hoos

H.H.

and Leytonbrown

, Auto-WEKA: combined selection and hyperparameter optimization of classification algorithms, Computer Science (2013), 847–855.

21.

Sehhati

, Mehridehnavi

, Rabbani

and Pourhossein

, Stable gene signature selection for breast cancer recurrence prediction using joint mutual information, IEEE/ACM Trans Computational Biology and Bioinformatics 12(6) (2015), 1440–1448.

22.

Xu , Xiang

, Liu

, et al. Stacked sparse autoencoder (SSAE) for nuclei detection on breast cancer histopathology images, IEEE Trans. on Medical Imaging 35(1) (2016), 119–130.

23.

Zhang , Liu

, Huang

, Wang

, Chen

and Meng

, Cancer Progression prediction using gene interaction regularized Elastic Net, IEEE/ACM Trans. on Computational Biology and Bioinformatics 14(1) (2017), 145–154.

24.

Lambertini , Santoro

, Mastro

L.D.

, et al. Reproductive behaviors and risk of developing breast cancer according to tumor subtype: A systematic review and meta-analysis of epidemiological studies, Cancer Treatment Reviews 49 (2016), 65–76.

25.

Lee and Minkyung

, Effects of interval between age at first pregnancy and age at diagnosis on breast cancer survival according to menopausal status: a register-based study in Korea, BMC Women’s Health 14(1) (2014), 113.

26.

Peng , Wei

, Lu

, et al. Diagnosis and treatment pattern among rural and urban breast cancer patients in Southwest China from to 2005 to 2009, Oncotarget 7(47) (2016), 78168–78179.

27.

Bhoo-Pathy , Yip

C.H.

, Hartman

, et al. Adjuvant! Online is overoptimistic in predicting survival of Asian breast cancer patients, European J. of Cancer 48(7) (2012), 982–989.

28.

Bohemian , Marateb

H.R.

, Mansourian

, Mananas

M.A.

and Mokarian

, A hybrid computer-aided-diagnosis system for prediction of breast cancer recurrence (HPBCR) using optimized ensemble learning, Computational and Structural Biotechnology Journal 15 (2017), 75–85.

29.

Wu , et al., Deep neural networks improve radiologists’ performance in breast cancer screening, IEEE Transactions on Medical Imaging, Early Access (2019), 1–1.

30.

Turkki , et al., Breast cancer outcome prediction with tumour tissue images and machine learning, Breast Cancer Research and Treatment 177(1) (2019), 41–52.