Predicting two-year quality of life after breast cancer surgery using artificial neural network and linear regression models 

C L I N I C A L T R I A L

Predicting two-year quality of life after breast cancer surgery

using artificial neural network and linear regression models

Richard Culbertson• _{Hong-Tai Chang}•

Ming-Feng Hou

Received: 11 April 2012 / Accepted: 17 July 2012 / Published online: 27 July 2012 Ó Springer Science+Business Media, LLC. 2012

Abstract The purpose of this study was to validate the use of artificial neural network (ANN) models for pre-dicting quality of life (QOL) after breast cancer surgery and to compare the predictive capability of ANNs with that of linear regression (LR) models. The European Organi-zation for Research and Treatment of Cancer Quality of Life Questionnaire and its supplementary breast cancer measure were completed by 402 breast cancer patients at baseline and at 2 years postoperatively. The accuracy of the system models were evaluated in terms of mean square error (MSE) and mean absolute percentage error (MAPE). A global sensitivity analysis was also performed to assess the relative significance of input parameters in the system model and to rank the variables in order of importance. Compared to the LR model, the ANN model generally had smaller MSE and MAPE values in both the training and

testing datasets. Most ANN models had MAPE values ranging from 4.70 to 19.96 %, and most had high predic-tion accuracy. The ANN model also outperformed the LR model in terms of prediction accuracy. According to global sensitivity analysis, pre-operative functional status was the best predictor of QOL after surgery. Compared with the conventional LR model, the ANN model in the study was more accurate for predicting patient-reported QOL and had higher overall performance indices. Further refinements are expected to obtain sufficient performance improvements for its routine use in clinical practice as an adjunctive decision-making tool.

Keywords Breast cancer Quality of life  Artificial neural network Linear regression  Global sensitivity analysis

H.-Y. Shi

Department of Healthcare Administration and Medical Informatics, Kaohsiung Medical University, Kaohsiung 80708, Taiwan

e-mail: hshi@kmu.edu.tw J.-T. Tsai

Department of Computer Science, National Pingtung University of Education, Pingtung 900, Taiwan

Y.-M. Chen

Faculty of Nursing, Kaohsiung Medical University, Kaohsiung 80708, Taiwan

Y.-M. Chen

Department of Nursing, Kaohsiung Municipal Hsiao-Kang Hospital, Kaohsiung 812, Taiwan

R. Culbertson

Department of Global Health Systems and Development, Tulane University, New Orleans, LA 70112, USA

H.-T. Chang

Division of General & Gastroenterological Surgery,

Department of Surgery, Kaohsiung Veterans General Hospital, Kaohsiung 80708, Taiwan

M.-F. Hou

Division of General & Gastroenterological Surgery, Department of Surgery, Kaohsiung Medical University Hospital, Kaohsiung 80708, Taiwan

M.-F. Hou (&)

Cancer Center, Kaohsiung Medical University Hospital and Institute of Clinical Medicine, College of Medicine, Kaohsiung Medical University, 100 Shi-Chuan 1st Road, Kaohsiung 80708, Taiwan, ROC

e-mail: mifeho@kmu.edu.tw M.-F. Hou

National Sun Yat-Sen University-Kaohsiung Medical University Joint Research Center, 804 Kaohsiung, Taiwan

Abbreviations

BCS Breast-conserving surgery MRM Modified radical mastectomy TRAM Transverse rectus abdominus muscle QOL Quality of life

ANNs Artificial neural networks LR Linear regression

MLP Multilayer perceptron MSE Mean square error

MAPE Mean absolute percentage error VSR Variable sensitivity ratios

Introduction

Women with early stage breast carcinoma generally have three equally effective surgical options: breast-conserving surgery (BCS), modified radical mastectomy (MRM), or transverse rectus abdominus muscle (TRAM) flap surgery. Since the three procedures have comparable survival rates, patients typically select the procedure that optimizes quality of life (QOL) [1–3]. Artificial neural networks (ANNs) are complex and flexible nonlinear systems with properties not found in other modeling systems. These properties include robust performance in dealing with noisy or incomplete input patterns, high fault tolerance, and the capability to generalize from the input data [4–6]. The computational power of an ANN is derived from the distributed nature of its connections. Once a model is trained, it can be tested against novel records to predict outputs [4–6].

Although models proposed in the literature so far have contributed to the growing understanding of breast cancer surgery outcomes, they have had major shortcomings [7–10]. First, few studies of breast cancer outcomes have used longi-tudinal data for more than 2 years. Second, most studies have analyzed populations in the United States (US) or other coun-tries, which may substantially differ from those in Taiwan. Third, no studies have considered group differences in factors other than outcomes such as age and nonsurgical treatment. Finally, almost all published articles agree that the essential issue of the internal validity (reproducibility) of the ANN and regression models has not been adequately addressed.

Therefore, the primary aim of the study was to validate the use of ANN models in predicting patient-reported QOL after breast cancer surgery, and the secondary aim was to compare the predictive capability of ANNs with that of linear regression (LR) models.

Materials and methods

Study design and population

The study included all patients who had been diagnosed and treated for incidental breast cancer between August,

2007 and September, 2009 at either of two participating tertiary academic hospitals in southern Taiwan. Patients who presented with curable diseases (i.e., no distant metastasis) were offered counseling regarding their surgi-cal options (BCS, MRM, or TRAM flap surgery). After excluding patients with benign tumor (n = 342) or cogni-tive impairment (n = 4), 479 patients who gave written consent were enrolled in the study. At 2 years postopera-tively, seventy-six patients were excluded due to loss to follow-up (n = 57) or refusal to participate (n = 19). The remaining 403 patients completed two surveys.

Instruments

The European Organization for Research and Treatment of Cancer (EORTC) QLQ-C30 and QLQBR23 questionnaires were used to assess QOL [11,12]. The Chinese versions of the EORTC QLQ-C30 and EORTC QLQ-BR23 have been validated in breast cancer surgical patients in Taiwan [13]. Since most symptom subscales of the QLQ-C30 and the QLQ-BR23 refer to systematic treatment effects, the analysis in this study was limited to the function subscales and global quality of life.

Before performing this study of human subjects, approval was obtained from all participating institutions. In all sub-jects, the QLQ-C30 and the QLQ-BR23 were administered by the same two research assistants before and after surgery.

System model development

The factors used in the LR model to predict long-term QOL of breast cancer surgery patients included both patient characteristics and hospital characteristics. The LR model can be formulated as the following linear equation:

^

Y ¼ b0þ biXiþ ei; i¼ 1; 2; . . .; m:

where ^Y is the actual output value, b0is the intercept, biis

the model coefficient parameter, Xi is the independent or

input variable, eiis the random error, and m is the number

of variables.

The ANN used in this study was a standard feed-for-ward, back-propagation neural network with three layers: an input layer, a hidden layer, and an output layer. The multilayer perceptron (MLP) network is an emerging tool for designing special classes of layered feed-forward net-works [14]. The cross-validation approach typically used to optimize the time when an MLP network training session ‘‘stops’’ is to include one estimation subset for training the model and one validation subset for evaluating the model performance. A neural network is optimized using a training dataset. A separate test dataset is used to halt training to mitigate overfitting. The training cycle is repe-ated until the test error no longer decreases [6,15].

Statistical analysis

The dataset was randomly divided into two sets: one set of 322 cases (80 % of the overall dataset) for training the model and another set of 81 cases for testing the model. The model was built using the training set. Patient char-acteristics and hospital charchar-acteristics were the indepen-dent variables, and the outcome (QOL) was the depenindepen-dent variable. The LR and ANN models were then tested using the 81 cases in the testing dataset.

The model fit and prediction accuracy of the system models were measured in terms of mean square error (MSE) and mean absolute percentage error (MAPE), respectively. The prediction accuracy of a model is con-sidered excellent if its MAPE value is lower than 10 %. Values between 10 and 20 %, between 20 and 50 %, and higher than 50 % are considered indicators of high, aver-age, and low prediction accuracy [16]. The formulas used to calculate MSE and MAPE were

MSE ¼ 1 n Xn i¼1 ðYi ^YiÞ 2 ; and MAPE ¼ 1 n Xn i¼1 Yi ^Yi    Yi  100 %;

where n is the number of observations, Yi is the desired

(target) value of the ith observation, and ^Yi is the actual

output value of the ith observation.

The change rate was also used to compare model per-formance between the training and testing sets. This cri-terion was used to calculate the difference in MSE index between the test and the training sets so that the better model could be identified. Absolute value was defined as [(the MSE value from testing set—the MSE value from training set)/(the MSE value from training set)] 9 100 %. Low change rates and low MSE values were considered indicators of good model performance.

The unit of analysis in this study was the individual breast cancer surgery patient. The data analysis was per-formed in several stages. First, continuous variables were tested for statistical significance by one-way analysis of variance (ANOVA), and categorical variables were tested by Fisher exact analysis. Univariate analyses were applied to identify significant predictors (p \ 0.05). Second, STATISTICA 10.0 (StatSoft, Tulsa, OK) software was used to construct the MLP network model and the LR model of the relationship between the identified predictors and QOL. Finally, a global sensitivity analysis was also performed to assess the relative significance of input parameters in the system model and to rank the variables in order of importance. The global sensitivity of the input

variables against the output variable was expressed as the ratio of the network error (variable sensitivity ratios, VSR) with a given input omitted to the network error with the input included. A ratio of one or lower indicates that the variable degrades network performance and should be removed [17].

Results

Table1 shows the patient characteristics and hospital characteristics in this study. The mean age of the study population was 52.21 (±9.59) years. On average, 88.18 % of female patients were married, and the overall CCI was 0.59 ± 0.99. Of these 403 patients, eight patients in stage IV showed confirmed true metastatic disease, including two lung metastases, two liver metastases, and four bone metastases; and forty-eight patients showed a breast cancer history before receiving the surgical procedure. The sig-nificant variables ultimately selected for inclusion in the LR models were education, menopause status, surgical type, chemotherapy, radiotherapy, hormone therapy, post-operation LOS, complications, and pre-post-operation func-tional status (p \ 0.05) (Table2).

In this study of the MLP network, 80 % training and 20 % testing samples are randomly selected to analyze the database in each run. In order to make MLP learning perform better, the neuron activation functions for the hidden and output neurons availably are given as follows, such as identity, hyperbolic tangent, logistic sigmoid, exponential, and Sine. The optimal number of neurons in the hidden layer and the type of the activation functions are iteratively determined by developing 50 neural networks and observing the MSE index of the output error. The training process would continue training the network for as many cycles as needed so long as the training and testing errors are on the decrease, otherwise it would stop training as the test error increases. The ANN-based approaches provided the 3-layer networks and the relative weights of neurons used for predicting QOL. For example, the MLP 9-13-1 model for the QLQ-BR23 body image score pre-diction included nine inputs, one bias neuron in the input layer, 13 hidden neurons, one bias neuron in the hidden layer, and one output neuron (Table3). The activation functions of logistic sigmoid and hyperbolic tangent were used in each neuron of the hidden layer and output layer, respectively.

For predicting QOL, the ANN model had relatively larger change rates and MSE values in the testing set with the exception of MSE for the testing set at year two (Tables4, 5). Apparently, the ANN model also outper-formed the LR model in terms of predictive accuracy. Most MAPE values obtained by the ANN model were lower than

20 %, which indicated that the ANN model had excellent accuracy for predicting QOL.

Table6 presents the VSR values for the outcome able (QOL) in relation to the three most influential vari-ables. In the global sensitivity analysis, the most influential (sensitive) parameter in terms of its effects on most QLQ-BR23 and QLQ-C30 subscales was pre-operative func-tional status followed by surgical type. All VSR values exceeded one, indicating that the network performed better when all variables were considered.

In order to verify the predictive accuracy of the models, the 40 datasets shown in Table7were collected. Compared to the LR model, the ANN model consistently obtained higher performance indices in the BR23 and QLQ-C30 subscales.

Discussion

Floyd et al. [18] was the first to develop an ANN model for predicting breast cancer based on mammographic findings. They concluded that the ANN model could be trained to predict malignancy based on mammographic findings with accuracy exceeding that of experienced radiologists. Ayer et al. [19] constructed LR and ANN models for estimating breast cancer risk based on mammographic descriptors and demographic risk factors. They concluded that the ANN model can be viewed as a generalization of the LR model and that the main advantage of ANN models over LR models is their hidden layers of nodes. Orr developed a Table 1 Patient and hospital characteristics of the study (N = 403)

Variables Mean ± SD N (%)

Patient characteristics

Age at operation (years) 52.21 ± 9.59 Married

No 48 (11.82)

Yes 355 (88.18)

Education (years) 9.43 ± 4.58 Living with immediate family

No 16 (3.97)

Yes 387 (96.03)

Body mass index (kg/m2_{) 23.85 ± 3.61}

Smoker No 389 (96.53) Yes 14 (3.47) Drinker No 391 (97.02) Yes 12 (2.98) Menopause status No 200 (49.63) Yes 203 (50.37)

Number of fetuses (cases) 2.35 ± 1.17 Breast cancer history

No 355 (88.09)

Yes 48 (11.91)

Other breast disease history

No 333 (82.63)

Yes 70 (17.37)

Charlson co-morbidity index 0.59 ± 0.99 Tumor pathology differentiation

High 56 (13.90) Medium 270 (67.00) Low 77 (19.10) Tumor stage Stage 0/I 159 (39.45) Stage II 147 (36.48) Stage III/IV 97 (24.17) Hospital characteristics Surgical procedure MRM 234 (58.06) BCS 113 (28.04) TRAM 56 (13.90)

Operation time (min) 166.92 ± 107.36 Anesthesia time (min) 197.71 ± 117.06 ASA class I 40 (9.93) II 312 (77.42) III 51 (12.65) Chemotherapy No 103 (25.56) Table 1 continued Variables Mean ± SD N (%) Yes 300 (74.44) Radiotherapy No 232 (57.57) Yes 171 (42.43) Hormone therapy No 228 (56.57) Yes 175 (43.43)

Post-operation LOS (days) 3.09 ± 1.49 Post-hospitalization 30 days No 324 (80.40) Yes 79 (19.60) Complications No 351 (87.10) Yes 52 (12.90)

MRM modified radical mastectomy; BCS breast-conserving surgery; TRAM transverse rectus abdominus muscle mastectomy with recon-struction; ASA American society of anesthesiologists; LOS length of stay; SD standard deviation

simplified and standardized method of classifying patients with abnormal mammograms by incorporating quantitative risk assessment [20]. His performance comparisons of ANN models and conventional LR models used for mammographic classification showed better discrimination in the ANN model. From a practical standpoint, however, the models showed similar performance in identifying malignant cases misclassified by clinical impression.

This study confirmed that, compared to LR models, ANN models are dramatically more accurate in predicting

patient-reported outcomes (QOL). To the best of our knowledge, this study is the first to use ANNs for analyzing predictors of QOL after breast cancer surgery. This model was tested against actual outcomes obtained by models constructed using identical inputs, including a neural net-work model and a linear regression model. We also showed that, given the same numbers of inputs for patient char-acteristics and hospital charchar-acteristics and the same two outcome measures, the predictive accuracy of ANN is superior to that of LR.

Multiple outcome-predicting models have been devel-oped with conventional statistical procedures, but their application at the individual level is hampered by the highly interdependent clinical variables involved, which may potentially interact with each other and have reci-procal enhancing effects [6,10]. Hence, conventional sta-tistical approaches have intrinsic limitations in handling this complex nonlinear information [4–6].

The ANNs are adaptive models that use a dynamic approach for analyzing outcome risks and can modify the internal structure in relation to a functional objective [4–6]. Although conventional statistics reveal significant param-eters only for the overall population, ANNs include parameters that are significant at the individual patient level even if they are not significant in the overall popu-lation [5,6]. We believe that the large and homogeneous Table 2 Coefficients of selected significant variables in each quality of life subscale of linear regression model (N = 403)a

Variable BRBI BRSEF BRSEE BRFU QL PF RF EF CF SF

Education 0.07 0.04 0.19 -0.02 0.14 0.13 8.31 0.11 0.27 0.07 Menopause status Yes versus no 2.02 0.57 2.39 -0.22 0.38 -0.17 -107.67 0.24 -0.31 0.58 Surgical type BCS versus MRM -1.38 -2.27 -0.48 0.29 1.76 -1.17 1.50 0.41 0.16 -1.24 TRAM versus MRM 1.25 -0.20 -0.70 0.03 1.69 -1.69 -8.87 1.47 1.40 -1.05 Chemotherapy Yes versus no 1.93 0.44 5.97 -1.07 0.22 0.31 -55.09 -0.86 -0.10 0.76 Radiotherapy Yes versus no 4.66 0.28 -3.68 0.51 0.62 -0.06 99.03 1.19 1.67 1.49 Hormone therapy Yes versus no -1.37 -0.11 -2.99 -0.59 0.88 0.89 -10.13 -0.42 -0.13 -0.07 Post-operation LOS -0.09 0.79 1.27 -0.34 -0.15 0.19 23.26 -0.46 -0.75 -0.89 Complications Yes versus no -21.41 -1.09 -0.55 0.15 -0.84 -0.77 -119.57 0.74 0.06 -0.84 Pre-operation functional status -0.27 -0.25 -0.95 -0.04 -0.11 -0.24 -9.17 -0.15 -0.24 -0.22 Constant 16.31 14.71 56.98 2.61 4.01 18.49 664.70 11.27 16.76 17.54 BRBI QLQ-BR23 body image; BRSEF QLQ-BR23 sexual functioning; BRSEE QLQ-BR23 sexual enjoyment; BRFU QLQ-BR23 future per-spective; QL QLQ-C30 global quality of life; PF QLQ-C30 physical functioning; RF QLQ-C30 role functioning; EF QLQ-C30 emotional functioning; CF QLQ-C30 cognitive functioning; SF QLQ-C30 social functioning; MRM modified radical mastectomy; BCS breast-conserving surgery; TRAM transverse rectus abdominus muscle mastectomy with reconstruction

a _{All regression coefficients are statistically significant (p \ 0.05)}

Table 3 Artificial neural network models at different subscales

Subscale Neta

QLQ-BR23 body image 9–13–1

QLQ-BR23 sexual functioning 9–9–1

QLQ-BR23 sexual enjoyment 9–5–1

QLQ-BR23 future perspective 9–18–1 QLQ-C30 global quality of life 9–11–1 QLQ-C30 physical functioning 9–4–1

QLQ-C30 role functioning 9–13–1

QLQ-C30 emotional functioning 9–18–1 QLQ-C30 cognitive functioning 9–10–1 QLQ-C30 social functioning 9–4–1

dataset in the present study, which included all demo-graphic and clinical variables shown to affect QOL in previous linear regression models, provided a sufficiently robust basis for training the network [5,6].

Throughout this two-year follow-up study, the best single predictor of QOL subscale scores was pre-operation functional status, which is consistent with reports that pre-operation functional scores are the best predictors of postoperative QOL [2,21]. Therefore, effective counseling is essential for apprising patients of expected post-surgery impairments. If QOL outcomes are considered as bench-marks then pre-operation functional status, which is a major predictor of postoperative QOL, is crucial. Patients should also be advised that their postoperative QOL might depend not only on the success of their operations, but also on their pre-operation functional status.

Furthermore, recent findings suggest that BCS outper-forms MRM for measuring role functioning, emotional functioning, cognitive functioning, and body image [2]. Compared with the BCS groups, however, the TRAM groups revealed significantly larger subjective improve-ments in physical functioning, emotional functioning, sexual functioning, and sexual enjoyment. One study found

that aspects of QOL, other than body image, were no better in women who underwent breast-conserving surgery or mastectomy with reconstruction than in women who had mastectomy alone [22]. Mastectomy with reconstruction was associated with greater mood disturbance and poorer health. However, the results of a 5-year prospective study on QOL following breast-conserving surgery or mastec-tomy indicated that mastecmastec-tomy patients had a significantly worse body image, role, and sexual functioning, and their lives were more disrupted [23].

In addition, to reduce the risk of recurrence and death, breast cancer patients usually receive systemic thera-pies (chemotherapy, hormone therapy, radiotherapy, and biological treatments) after surgery. Several studies Table 4 Comparison of artificial neural network (ANN) model and

linear regression (LR) model in predicting QLQ-BR23 subscale scores

Index Model Training set (A) Testing set (B) Change ratea QLQ-BR23 body image score

MSE ANN 66.67 84.85 27.27 %

LR 70.00 88.89 26.99 %

MAPE ANN 17.14 % 19.57 % –

LR 20.46 % 28.23 % –

QLQ-BR23 sexual functioning score

MSE ANN 66.67 50.00 25.00 %

LR 72.22 57.14 15.08 %

MAPE ANN 12.50 % 8.79 % –

LR 22.24 % 10.32 % –

QLQ-BR23 sexual enjoyment score

MSE ANN 75.00 50.00 33.33 %

LR 83.33 66.67 19.99 %

MAPE ANN 16.81 % 8.84 % –

LR 27.31 % 12.83 % –

QLQ-BR23 future perspective score

MSE ANN 83.33 68.18 18.18 %

LR 90.32 75.00 16.96 %

MAPE ANN 19.96 % 16.10 % –

LR 34.71 % 22.87 % –

MSE mean square error; MAPE mean absolute percentage error

a _{Change rate =j½ðB  AÞ=ðAÞj  100 %}

Table 5 Comparison of artificial neural network (ANN) model and linear regression (LR) model in predicting QLQ-C30 subscale scores Index Model Training set (A) Testing set (B) Change ratea QLQ-C30 global quality of life score

MSE ANN 29.31 14.81 49.47 %

LR 34.67 18.24 47.38 %

MAPE ANN 10.75 % 6.14 % –

LR 14.63 % 8.06 % –

QLQ-C30 physical functioning score

MSE ANN 83.87 66.67 20.51 %

LR 92.77 77.59 16.36 %

MAPE ANN 17.14 % 15.81 % –

LR 35.57 % 19.31 % –

QLQ-C30 role functioning score

MSE ANN 18.68 6.24 66.60 %

LR 24.07 11.20 53.47 %

MAPE ANN 7.09 % 4.84 % –

LR 9.31 % 6.50 % –

QLQ-C30 emotional functioning score

MSE ANN 83.87 75.00 4.54 %

LR 92.77 88.89 4.18 %

MAPE ANN 17.43 % 16.57 % –

LR 40.56 % 24.23 % –

QLQ-C30 cognitive functioning score

MSE ANN 75.00 50.00 33.33 %

LR 83.33 59.09 29.09 %

MAPE ANN 16.01 % 10.46 % –

LR 18.14 % 11.19 % –

QLQ-C30 social functioning score

MSE ANN 17.47 5.84 66.57 %

LR 17.47 10.42 16.96 %

MAPE ANN 8.64 % 4.70 % –

LR 9.04 % 5.64 % –

MSE mean square error; MAPE mean absolute percentage error

evaluated QOL on breast cancer patients who have received systemic therapies [1–3]. Chemotherapy has considerable effects on QOL for breast cancer patients. Notedly, a complication is a well-recognized risk factor

with adverse outcomes in breast cancer surgery. Our sta-tistical data also show a strong and positive association with poor QOL, which is consistent with previous findings [2,24].

Table 6 Global sensitivity analysis of QLQ-BR23 and QLQ-C30 subscales of artificial neural network (ANN) model

Subscale First Second Third

(VSR) (VSR) (VSR)

QLQ-BR23 body image Surgical type Pre-operation functional status Radiotherapy

(1.54) (1.44) (1.35)

QLQ-BR23 Pre-operation functional status Surgical type Complication

Sexual functioning (1.86) (1.56) (1.21)

QLQ-BR23 Pre-operation functional status Complication Surgical type

Sexual enjoyment (6.49) (1.27) (1.17)

QLQ-BR23 Surgical type Chemotherapy Pre-operation functional status

Future perspective (4.20) (3.45) (2.37)

QLQ-C30 Pre-operation functional status Surgical type Complication

Global quality of life (2.21) (2.19) (1.49)

QLQ-C30 Pre-operation functional status Complication Surgical type

Physical functioning (2.60) (1.79) (1.71)

QLQ-C30 Pre-operation functional status Complication Chemotherapy

Role functioning (82.89) (12.54) (9.46)

QLQ-C30 Pre-operation functional status Surgical type Complication

Emotional functioning (1.36) (1.14) (1.10)

QLQ-C30 Pre-operation functional status Complication Surgical type

Cognitive functioning (1.86) (1.68) (1.55)

QLQ-C30 Surgical type Complication Pre-operation functional status

Social functioning (8.97) (7.91) (6.86)

VSR variable sensitivity ratios

Table 7 Comparison of performance indices of artificial neural network (ANN) model and linear regression (LR) model for predicting QLQ-BR23 and QLQ-C30 subscale scores based on forty new datasets

Subscale ANN model LR model

Sensitivity 1-Specificity PPV NPV Accuracy AUC Sensitivity 1-Specificity PPV NPV Accuracy AUC QLQ-BR23 BRBI 100.00 100.00 1.00 1.00 100.00 1.00 92.86 92.86 0.92 0.92 93.33 0.92 BRSEF 95.83 90.91 0.96 0.91 90.00 0.92 75.00 90.91 0.75 0.91 86.67 0.83 BRSEE 100.00 95.83 0.86 1.00 96.67 0.98 100.00 95.83 0.86 1.00 96.67 0.98 BRFU 66.67 100.00 1.00 0.92 93.33 0.83 33.33 91.67 0.50 0.85 80.00 0.63 QLQ-C30 QL 100.00 81.25 0.82 1.00 90.00 0.91 92.86 81.25 0.81 0.93 86.67 0.87 PF 66.67 95.83 0.80 0.92 90.00 0.81 66.67 95.83 0.80 0.92 90.00 0.81 RF 100.00 100.00 1.00 1.00 100.00 1.00 100.00 95.83 0.86 1.00 96.67 0.98 EF 66.67 95.83 0.86 0.89 86.67 0.87 66.67 94.44 0.89 0.81 83.33 0.81 CF 66.67 100.00 1.00 0.92 93.33 0.92 40.00 96.00 0.67 0.89 86.67 0.68 SF 100.00 100.00 1.00 1.00 100.00 1.00 83.33 100.00 1.00 0.96 96.67 0.92 BRBI QLQ-BR23 body image; BRSEF QLQ-BR23 sexual functioning; BRSEE QLQ-BR23 sexual enjoyment; BRFU QLQ-BR23 future per-spective; QL QLQ-C30 global quality of life; PF QLQ-C30 physical functioning; RF QLQ-C30 role functioning; EF QLQ-C30 emotional functioning; CF QLQ-C30 cognitive functioning; SF QLQ-C30 social functioning; PPV positive predictive value; NPV negative predictive value; AUC area under the curve

Although all research questions were satisfactorily addressed, several limitations are noted. First, this study col-lected data for breast cancer surgery patients who had been under the supervision of two surgeons in two different med-ical centers, each of whom had performed the highest volume of breast cancer surgery procedures in his respective hospital during the previous 20–30 years. This sample selection pro-cedure ensured that patient outcome data would not be affected by surgeons with limited experience. By focusing the analysis on procedures performed by these two surgeons, the results of this study are more representative of all breast cancer patients compared to one analyzing those performed by a single surgeon. However, a notable limitation is that the first patient in the prospective patient cohort was enrolled in 2007. Therefore, depending on their inclusion date, some surveyed patients had a longer follow-up than others did, which may have caused selection bias. Nonetheless, in most QOL subscales, the characteristics of subjects who continu-ously participated throughout this 2-year study did not sig-nificantly differ from those of subjects who died or dropped out during the study (data not shown).

Conclusions

Compared with the conventional multivariate LR model, the ANN model in the study was more accurate in predicting patient-reported QOL and had higher overall performance indices. The global sensitivity analysis also showed that pre-operation functional status is the most important predictor of the QLQ-BR23 and the QLQ-C30 after breast cancer surgery. The predictors analyzed in this study could be addressed in pre-operative and postoperative health care consultations to educate candidates for breast cancer surgery in the expected course of recovery and expected functional outcomes. Further studies of this model may consider the effect of a more detailed database that includes complications and clinical examination findings as well as more detailed outcome data. Hopefully, the model will evolve into an effective adjunctive clinical decision-making tool.

Acknowledgments This work was supported by the National Sci-ence Council, Taiwan, Republic of China, under grant numbers NSC99-2314-B-037-069-MY3 and by the Department of Health, Executive Yuan, under Grant Numbers DOH101-TD-C-111-002. Conflicts of interest The authors have no personal or professional conflicts of interest in the publication of this study.

References

1. Dawood S, Hu R, Homes MD, Collins LC, Schnitt SJ, Connolly J, Colditz GA, Tamimi RM (2011) Defining breast cancer

prognosis based on molecular phenotypes: results from a large cohort study. Breast Cancer Res Treat 126:185–192

2. Shi HY, Uen YH, Yen LC, Culbertson R, Juan CH, Hou MF (2011) Two-year quality of life after breast cancer surgery: a comparison of three surgical procedures. Eur J Surg Oncol 37:695–702

3. van den Hurk CJ, Eckel R, van de Poll-Franse LV, Coebergh JW, Nortier JW, Ho¨lzel D, Breed WP, Engel J (2011) Unfavourable pattern of metastases in M0 breast cancer patients during 1978–2008: a population-based analysis of the Munich Cancer Registry. Breast Cancer Res Treat 128:795–805

4. Tu JV (1996) Advantages and disadvantages of using artificial neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol 49:1225–1231

5. Zou J, Han Y, So SS (2008) Overview of artificial neural net-works. Methods Mol Biol 458:15–23

6. Sandberg IW, Lo JT, Fancourt CL, Principe JC, Katagiri S, Haykin S (2001) Nonlinear dynamical systems: feedforward neural network perspectives. Wiley, New York

7. Giordano A, Giuliano M, De Laurentiis M, Eleuteri A, Iorio F, Tagliaferri R, Hortobagyi GN, Pusztai L, De Placido S, Hess K, Cristofanilli M, Reuben JM (2011) Artificial neural network analysis of circulating tumor cells in metastatic breast cancer patients. Breast Cancer Res Treat 129:451–458

8. Lancashire LJ, Powe DG, Reis-Filho JS, Rakha E, Lemetre C, Weigelt B, Abdel-Fatah TM, Green AR, Mukta R, Blamey R, Paish EC, Rees RC, Ellis IO, Ball GR (2010) A validated gene expression profile for detecting clinical outcome in breast cancer using artificial neural networks. Breast Cancer Res Treat 120: 83–93

9. Foukakis T, Fornander T, Lekberg T, Hellborg H, Adolfsson J, Bergh J (2011) Age-specific trends of survival in metastatic breast cancer: 26 years longitudinal data from a population-based cancer registry in Stockholm, Sweden. Breast Cancer Res Treat 130:553–560

10. Zujewski JA, Harlan LC, Morrell DM, Stevens JL (2011) Ductal carcinoma in situ: trends in treatment over time in the US. Breast Cancer Res Treat 127:251–257

11. Fayers PM, Aaronson NK, Bjordal K, Groenvold M, Curran D, Bottomley A (2006) EORTC QLQ C30 scoring manual, 3rd edn. EORTC, Brussels

12. Sprangers MA, Groenvold M, Arraras JI, Franklin J, te Velde A, Muller M, Franzini L, Williams A, de Haes HC, Hopwood P, Cull A, Aaronson NK (1996) The European Organization for Research and Treatment of Cancer breast cancer-specific quality-of-life questionnaire module: first results from a three-country field study. J Clin Oncol 14:2756–2768

13. Chie WC, Chang KJ, Huang CS, Kuo WH (2003) Quality of life of breast cancer patients in Taiwan: validation of the Taiwan Chinese version of the EORTC C30 and EORTC QLQ-BR23. Psychooncology 12:729–735

14. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning internal representations by error propagation. In: Rumelhart DE, McCleland JL (eds) Parallel distributed processing: explorations in the microstructure of cognition. MIT Press, Cambridge, pp 318–362

15. Haykin S (1999) Neural networks: a comprehensive foundation, 2nd edn. Prentice-Hall, Englewood Cliffs

16. Woods LM, Coleman MP, Lawrence G, Rashbass J, Berrino F, Rachet B (2011) Evidence against the proposition that ‘‘UK cancer survival statistics are misleading’’: simulation study with national cancer registry data. Br Med J 342:d3399

17. Hunter A, Kennedy L, Henry J, Ferguson I (2000) Application of neural networks and sensitivity analysis to improved predic-tion of trauma survival. Comput Methods Programs Biomed 62:11–19

18. Floyd CE Jr, Lo JY, Yun AJ, Sullivan DC, Kornguth PJ (1994) Prediction of breast cancer malignancy using an artificial neural network. Cancer 74:2944–2948

19. Ayer T, Chhatwal J, Alagoz O, Kahn CE Jr, Woods RW, Burn-side ES (2010) Informatics in radiology: comparison of logistic regression and artificial neural network models in breast cancer risk estimation. Radiographics 30:13–22

20. Orr RK (2001) Use of an artificial neural network to quantitate risk of malignancy for abnormal mammograms. Surgery 129: 459–466

21. Rottmann N, Dalton SO, Christensen J, Frederiksen K, Johansen C (2010) Self-efficacy, adjustment style and well-being in breast cancer patients: a longitudinal study. Qual Life Res 19:827–836

22. Nissen MJ, Swenson KK, Ritz LJ, Farrell JB, Sladek ML, Lally RM (2001) Quality of life after breast carcinoma surgery: a comparison of three surgical procedures. Cancer 91:1238–1246 23. Engel J, Kerr J, Schlesinger-Raab A, Sauer H, Halzel D (2004)

Quality of life following breast-conserving therapy or mastec-tomy: results of a 5-year prospective study. Breast J 10:223–231 24. Deshpande AD, Sefko JA, Jeffe DB, Schootman M (2011) The association between chronic disease burden and quality of life among breast cancer survivors in Missouri. Breast Cancer Res Treat 129:877–886