A scoring system based on artificial neural network for predicting 10-year survival in stage II A colon cancer patients after radical surgery

Nearly 20% patients with stage II A colon cancer will develop recurrent disease post-operatively. The present study aims to develop a scoring system based on Artificial Neural Network (ANN) model for predicting 10-year survival outcome. The clinical and molecular data of 117 stage II A colon cancer patients from Sun Yat-sen University Cancer Center were used for training set and test set; poor pathological grading (score 49), reduced expression of TGFBR2 (score 33), over-expression of TGF-β (score 45), MAPK (score 32), pin1 (score 100), β-catenin in tumor tissue (score 50) and reduced expression of TGF-β in normal mucosa (score 22) were selected as the prognostic risk predictors. According to the developed scoring system, the patients were divided into 3 subgroups, which were supposed with higher, moderate and lower risk levels. As a result, for the 3 subgroups, the 10-year overall survival (OS) rates were 16.7%, 62.9% and 100% (P < 0.001); and the 10-year disease free survival (DFS) rates were 16.7%, 61.8% and 98.8% (P < 0.001) respectively. It showed that this scoring system for stage II A colon cancer could help to predict long-term survival and screen out high-risk individuals for more vigorous treatment.


INTRODUCTION
Colorectal cancer (CRC) has been ranked as the fourth most common malignancies among women and the fifth among men in China [1]. For stage II colon cancer patients, the 5-year survival rate was approximately 80% after surgical resection [2][3][4]. According to the results of previous studies, 10% to 30% of these patients would eventually develop recurrent disease despite of receiving radical treatment [5,6]. Several clinical indicators have been identified as risk factors including tumor stage T4, poor tumor grading, elevation of pre-operative carcinoembryonic antigen(CEA), presence of lymphovascular or perineural invasion [7][8][9]. Recently, certain molecular markers have been applied in predicting prognosis for stage II colon cancer. Lower-expression p53 was considered as a risk predictor of disease recurrence and bad survival outcome at 10 years [10]. Over-expression of survivin also significantly correlated with worse overall survival [11,12]. However, any single clinical indicator or biomarker seems not of enough power to well predict the long-term survival outcome, and a predictive model by integrating both clinical indicators and molecular biomarkers is hopeful to meet the clinical needs.
Artificial neural network (ANN) is a mathematical and computational method that has been applied for diagnosis and prognosis prediction in several types of cancer in previous studies achieving higher sensitivity and specificity than the traditional procedures [13,14,15]. The objective of this study is to construct a scoring system with ANN model to predict 10-year survival for stage II A colon cancer patients after radical surgery and then screen prognostic high-risk subgroup out of them.

Patient characteristics
Of the total 151 patients, complete results of Tissue Microarrays (TMA) analysis were obtained in 138 patients. With the median 12 years' follow-up, 21 patients (15.2%, 21/138) were lost to obtain a specific survival status. Thus, 117 cases (72 males and 45 females, with median age of 57 years) with definite survival outcome were included for analysis. Most of the tumors were moderately differentiated adenocarcinoma with median tumor size of 5 cm. Post-operative metastases occurred in 23 patients (19.6%, 23/117), mostly entailing liver metastases (12.8%, 15/117). Within 10-year follow-up, 16 patients (13.6%, 16/117) died of colon cancer progression. The 10-year overall survival (OS) and disease free survival (DFS) rate of total patients were both 86.8%. The 117 cases were randomly divided into a training set (n = 59) and an independent test set (n = 58). The clinical and pathological data are shown in Table 1. Table 2 lists the cut-off points of binary classification for all clinical and molecular indicators and the results of univariate analyses on 10-year survival status. Figure 1 shows Receiver Operating Characteristic curve (ROC) for 5 clinical variables ( Figure 1A), total molecular indicators in tumor tissue ( Figure 1B) and those in normal mucosa ( Figure 1C). We developed 4 ANN models. Model 1 started from 9 clinical indicators (Table 2). After analysis, tumor size, pathological grading, body mass index (BMI) and postoperative liver metastasis were selected as significant prognostic indicators. This model was moderately associated with 10-year survival outcome, with sensitivity 53.8%, specificity 97.8% and overall accuracy 87.9% ( Figure 1D).
Model 4 integrated all the significant clinical indicators and biomarkers found by Model 1, Model 2 and Model 3. Pathological grading, TGFBR2, TGF-β, MAPK, pin1, β-catenin in tumor tissue and TGF-β in normal mucosa were ultimately identified as significant prognostic predictors. Model 4 was strongly correlated with the 10year survival outcome, with sensitivity 92.3%, specificity 93.3% and overall accuracy 93.1% ( Figure 1D).

Scoring system development
The scoring system was developed based on 7 significant predictors in Model 4 (Table 3). In this system, 49 credits were assigned to poor pathological grading, 33 to reduced expression of TGFBR2, 45 to over-expression of TGF-β, 32 to over-expression of MAPK, 100 to overexpression of pin1, 50 to over-expression in tumor tissue and 22 to reduced expression of TGF-β in normal mucosa. For instance, a stage II A colon patient with a poor pathological grading, pin1 >9 and β-catenin >9 in tumor tissue would obtain a total score of 49+100+50 = 199. The relation between 10-year survival probability and the total score is shown in Table 4. Patients with total score ranging from 0 to 100 were classified into lower risk subgroup, 101 to 220 into moderate risk subgroup, and 221 to 331 into higher risk subgroup. The average 10-year OS rate in moderate risk subgroup was significantly larger than that in higher risk subgroup (62.9% vs. 16.7%; P < 0.001), but significantly smaller than that in lower risk subgroup (62.9% vs. 100%; P = 0.002) ( Figure 2A); And similarly, the average 10-year DFS rate in moderate risk subgroup was significantly larger than that in higher risk subgroup (61.8% vs. 16.7%; P < 0.001) , but significantly smaller than that in lower risk subgroup (61.8% vs. 98.8%; P = 0.003) ( Figure 2B).

DISCUSSION
Although 5-fluorouracil (5-FU)-based adjuvant chemotherapy is recommended for stage II B-C (T4N0) colon cancer patients, the necessity of post-operative chemotherapy in patients with stage II A colon cancer remains controversy [16][17][18]. Identification of the higher risk subgroup with poorer prognosis is a crucial strategy to optimize post-operative treatment for stage II A colon cancer patients. In current study, we developed a new approach to predict 10-year prognosis and screen out the higher risk individuals by integrating clinical factors and   [19,20].
Additionally, previous studies have showed that TGFBR2 was a prognostic marker and its down regulation exhibited poor overall survival in oral cancer and breast cancer [21,22]. In our data, TGFBR2 was also served as a protective prognostic factor for stage II A colon cancer. However, we  found an opposite predicting result of TGF-β expression between tumor tissue and normal mucosa of stage II A colon cancer. Our study shows that over-expression of TGF-β in tumor tissue is a high risk factor for long-term survival, whereas over-expression of TGF-β in normal mucosa is a protective prognostic factor. It is noted that TGF-β has bidirectional functions in the progression of cancer: it acts as a tumor suppressor during the early stages by inhibiting cell proliferation while it functions as a pro-oncogenic factor during tumor progression through stimulating induction of epithelial-mesenchymal transition (EMT) [23,24].
MAPK signaling pathway is correlated with proliferation, survival and apoptosis of tumor cells, and therefore activates the carcinogenesis [25]. Zolota et al. found high expression of MAPK signaling pathway components indicated poor overall survival of patients with gliomas, suggesting that MAPK attributed to promoting biologic behavior of gliomas [26]. Similarly, Yu et al. have reported that high MAPK expression in colorectal cancer specimen predicted poorer prognosis [27]. Nevertheless, Yu et al. did not assess MAPK expression in normal colon tissue of the patients. By evaluating the staining image of immunohistochemistry (IHC) in both tumor and normal tissue, our model indicated that MAPK expression in tumor tissue instead of normal tissue was negatively correlated to better longterm outcome in stage II A colon cancer .
Pin1 has been reported that it contributed to tumor development and promoted tumor aggressiveness [28]. Its over-expression is related with poorer survival outcome in gastric cancer and non-small cell lung cancer [29,30]. Additionally, Kuramochi et al. found that high expression of pin1 was correlated with poor histological type (P = 0.0240), more aggressive invasion (P = 0.0051), and worse staging (P = 0.0027) of colorectal tumors [31]. In present study, we observed a strong correlation between pin 1 expression and survival outcome, which inferenced pin1 expression as an excellent prognostic predictor for stage II A colon cancer. On the other hand, it has been widely reported that pin 1 could upregulate β-catenin expression in tumor progression [31][32][33]. As a vital member of Wnt/β-catenin pathway, β-catenin activates tumorigenic behaviors, such as migration, stemness, anchorage-independent growth and chemo-sensitivity [34]. However, the prognostic significance of β-catenin expression in CRC remains unclear. Previous studies as well as our study showed the worse survival outcome for patients with high expression of nuclear β-catenin [35,36] , whereas some showed that the high nuclear expression was associated with a better prognosis [37,38], and other studies even failed to find any predictive value of nuclear β-catenin in CRC [39,40]. Further studies are needed to verify the predictive value of β-catenin in CRC.
The scoring system we have constructed can convert the result of ANN models into a visualized scale, thus making it easier to calculate the probability for 10-year survival. Further more, the subgroups defined by this scoring system revealed that the patients in subgroup with higher score had a poorest long-term prognosis, thus proving its strong discriminate power. The system will be helpful for making clinical decision by guiding postoperative treatment for the higher risk subgroup, such as adjuvant chemotherapy and more normative follow-up.
Several potential limitations of this scoring system should be acknowledged. First of all, only the final survival status after 10 years post-operatively was investigated, while the time dimension was not taken into account. Second, the complete data of post-operative treatment for patients failed to be obtained, which impeded us to predict the long-term survival outcome by post-operative treatment factors. Post-operative treatment regimens and duration were distinguished in several patients, which might impair the analysis accuracy without concerning of the situation. Third, since this scoring system was generated based on a limited number of patients from a single institution, the validity of this scoring system should be improved by large-volume and prospective studies.
In conclusion, this scoring system based on ANN model have identified reduced expression of TGFBR2, over-expression of TGF-β, MAPK, pin1, β-catenin in tumor tissue and reduced expression of TGF-β in normal mucosa as significant risk predictors for 10-year survival outcome in stage II A colon cancer patients after radical surgery. We considered this scoring system can better predict long-term survival outcome and then screen out high-risk individuals with stage II A colon cancer for further aggressive post-operative treatment.

Patient selection
The study included 151 patients with stage II A colon cancer who underwent lesions radical resection from August 1996 to May 2003 at Department of Colorectal Surgery in Sun Yat-sen University Cancer center. The detailed clinical information of the eligible patients was collected through the electronic medical records system. The inclusion criteria were: (1) The histological diagnosis should be proven as colon adenocarcinomas; (2) The disease stage should be confirmed as T3N0M0 according to 2010 American Joint Committee on Cancer/ International Union Against Cancer staging system (AJCC/UICC); (3) No adjuvant therapy was received by any patient; (4) No evidence showed any history of other active malignancy (except for basal cell carcinoma of the skin) and existence of multiple primary colorectal cancer. All patients were recommended to follow up every 6 months for 10 years post-operatively.
The study was approved by the Medical Ethics Committee of the Cancer Center, Sun Yat-sen University (NO. GZR2011-10).

Statistical analysis
In this study, each continuous or ordinal variable was dichotomized by a cut-off point chosen via ROC analysis and then went through a univariate analysis for 10-year survival outcome. The ANN applied in this study was a Multi-layer Perceptron (MLP) neural network including three layers: an input layer, a hidden layer and an output layer. The first layer has input neurons which sends data via synapses to the hidden layer of neurons, and then via more synapses to the third layer of output neurons. ANN was constructed to predict 10-year survival status using R packages, including "caret" and "nnet" for ANN model selection with recursive backwards feature selection via bootstrap resampling technique, "pROC" for ROC analysis and "rms" for scoring system development. Basic ANN model of clinical indicators and biomarkers were generated respectively, which was subsequently integrated to establish an ultimate model. The data of each ANN model was randomly and equally distributed into training set and test set by sample function of R packages. All ANN models were modeled with training set and evaluated their accuracy with test set. Scoring system was developed by converting the ultimate ANN model into principal component regression model with predicted values from the ANN model as responses. The score for each risk factor was assigned according to the regression coefficient. The total score for each individual was calculated by adding the allocated scores corresponding to a certain 10-year survival probability. According to the distinguishing survival probability, the patients were divided into distinct risk groups and the difference of 10-year OS and DFS were compared among groups by applying Log-rank test respectively.