Research ArticleTwo Artificial Neural Networks for Modeling DiscreteSurvival Time of Censored Data
Taysseer Sharaf1 and Chris P. Tsokos2
1Department of Mathematics and Statistics, University of South Florida, 4202 E. Fowler Avenue, CMC 342, Tampa, FL 33620, USA2Department of Mathematics and Statistics, University of South Florida, 4202 E. Fowler Avenue, CMC 366, Tampa, FL 33620, USA
Correspondence should be addressed to Taysseer Sharaf; [email protected]
Received 17 September 2014; Revised 17 February 2015; Accepted 23 February 2015
Academic Editor: Jun He
Copyright © 2015 T. Sharaf and C. P. Tsokos. This is an open access article distributed under the Creative Commons AttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properlycited.
Artificial neural network (ANN) theory is emerging as an alternative to conventional statistical methods in modeling nonlinearfunctions. The popular Cox proportional hazard model falls short in modeling survival data with nonlinear behaviors. ANN is agood alternative to the Cox PH as the proportionality of the hazard assumption andmodel relaxations are not required. In addition,ANN possesses a powerful capability of handling complex nonlinear relations within the risk factors associated with survival time.In this study, we present a comprehensive comparison of two different approaches of utilizingANN inmodeling smooth conditionalhazard probability function.Weuse realmelanoma cancer data to illustrate the usefulness of the proposedANNmethods.We reportsome significant results in comparing the survival time of male and female melanoma patients.
1. Introduction
Artificial neural network (ANN) is becoming one of themostpopular alternatives to conventional statistical modeling.It is actually conceived as an advanced generalized linearmodel. We have seen various applications of ANN utilizedin different scientific subjects like engineering, economics,environment, and health, among others. For example, vanHinsbergen et al. 2009 [1] applied artificial neural networksto short-term time prediction of traffic travel time. Kingstonet al. 2005 [2] proposed ANN to model water resources.In economics, Baesens et al. 2005 [3] used ANN to predictsurvival time of personal loan data. Baesens et al. comparedthe ANNmodel used with other survival analysis models likelogistic regression and Cox PH and the results came in favorof the ANN.
In themedical sciences,most of the proposed applicationsof ANN were on prognostic models. For example, one of themost paramount research entities is cancer. Classifying atumor asmalignant or benign is important in cancer research.Chen et al. in 2002 used ANN to diagnose breast cancertumors [4]. Ercal et al. in 1994 presented an ANN modelto distinguish between three benign skin cancer categories
and malignant melanoma [5]. But fitting a complex nonlin-ear modeling such as ANN in regression problems is lessprevalent. Determining the risk factors that cause canceror modeling the survival time of a patient once he/she isdiagnosed with cancer using ANN is less common.
In this present study, we are interested in utilizingANN insurvival time modeling of skin cancer (melanoma) patients.Soong et al. [6] in 2010 developed a statistical model topredict the survival time of localized melanoma patients.They used the proportional hazard model developed by Cox[7], but the assumptions of hazard function proportionalitymay not be applicable to a different set of data.Moreover, theydid not study the effect of interaction terms. Thus, applyingANN is more applicable and efficient, especially when thedata does not satisfy Cox PH assumptions. ANN does notrequire any assumptions that need to be justified, and it ismore precise in fitting nonlinear models [8–10]. One of thebasic approaches in utilizing ANN in survival analysis is byclassification, whether a patient will survive over a fixed timeinterval or not [11]. However, the latter classification methodlacks the information about the survival probability functionestimates. In 1995, Lapuerta et al. proposed the use ofmultipleneural networks one for each time interval [12]. This model
Hindawi Publishing CorporationAdvances in Artificial IntelligenceVolume 2015, Article ID 270165, 7 pageshttp://dx.doi.org/10.1155/2015/270165
2 Advances in Artificial Intelligence
predicts the survival probability of each time period based ona neural network trained on the observations of the same timeperiod only.The pitfall of this approach is the large number ofnetworks that will be trained if one studies the survival timeover immense time intervals.
Other methods of ANN applied to survival time wereproposed by Faraggi and Simon in 1995 [8] and by Ohno-Machado in 1996 [13].We consider in this study the approachrepresented by Biganzoli et al. in 1998 [9], which was amodification of a study done by Ravdin and Clark in 1992[14], in addition to the approach represented by Mani et al.in 1999 [10]. Thus, in this study a comparison between thetwo methods Biganzoli and Mani is given. Also, we studythe difference between the survival time of male and femalemelanoma patients.
In the following section, we discuss the data used to per-form our comparison, along with significant results exhibit-ing the differences between male and female melanomapatient survival times. In the third section we discuss brieflythe two methods emphasizing the differences, advantages,and disadvantages of both. In the fourth section we presentour results and identify the model that gave the best perfor-mance in estimating the survival probability function withless error.
2. Materials and Methods
2.1. Data. We have 130,006 patients diagnosed with mela-noma between the years 2000 and 2009 in the USA. Dataaccumulated from 13 registers of the Surveillance, Epidemi-ology, and end results program (SEER) [15]. We filter out thislarge dataset to contain only consummate information withrespect to the patient’s age at diagnosis, tumor thickness, stageof cancer, and ulceration. Soong et al. [6] in 2010 used thesefour variables, but their study did not consider the differencebetween male and female survival. We found that there existsa significant difference between the median survival timeof males and females based on a 5% level of significanceusing the Kruskal-Wallis test. Thus, studying the effect ofgender on survival time by making one model for both malesand females is not statistically correct, as the survival timefor male and females does not have the same distribution.Figure 1 represents a schematic diagram of the distributionof the complete data with respect to gender and cancer stage.
We train neural network model separately for malesand females. Figure 1 also includes the total number ofpatients with complete informationwhowere diagnosedwithmelanoma between the years 2000 and 2009. Patients wereeither alive at the end of the 10-year period (censored) orlost to follow-up during the ten-year interval (censored),in addition to patients who died because of melanoma(uncensored) (95% of male patients were censored and 98.1%of female patients were censored). We omitted patients withincomplete information. For training purposes, each of thegender category data was divided into six groups; five will beused in the cross validation technique to train and validatethe neural network model, while the last group is used forprediction and for accomplishing the comparison between
Total69082
Males39147
Stage 0953
Stage 132980
Stage 24370
Stage 3844
Females29935
Stage 0751
Stage 126093
Stage 22679
Stage 3412
Figure 1: Distribution of complete information of melanomapatients.
the two modeling approaches. More details about the crossvalidation will follow in the next section.
In our modeling procedure, we used age at diagnosis andtumor thickness (in millimeter) as quantitative variablesalong with three dummy variables representing stage 1 (refer-ring to a localized tumor), stage 2 (referring to a regionaltumor), and stage 3 (referring to a distant tumor). The baselevel is referred to as in situ tumor.
2.2. Methodology. ANN was based on the model of one per-ceptron introduced by Rosenblatt in 1957 [16]. Today a mul-tilayer perceptron (MLP) is known now by neural networksand consists of multiple layers of neurons. The first layer(input layer) represents the covariates or the risk factors,which are the inputs of the hidden neurons in the first hiddenlayer. The output of the hidden layer is the input of anotherhidden layer (if more than one hidden layer exists) or theoutput layer.This type of MLP is called feed forward artificialneural network (FFANN).We shall discuss bothmethods andcompare them in the current study using FFANN with onehidden layer.
2.2.1. Method 1: PLANN. Partial logistic artificial neuralnetwork (PLANN) is the approach that was introduced byBiganzoli et al. [9]. PLANN is a three-layer feed forwardartificial neural network with one output unit in the outputlayer. The activation function used in the hidden and outputlayer is the logistic function given by
𝑓 (𝑧) =exp (𝑧)
1 + exp (𝑧). (1)
PLANN estimates the conditional hazard function that isbased on the discrete survival method. The discrete survivalmethod was introduced by Allison [17] in 1982 and then
Advances in Artificial Intelligence 3
Singer and Willett [18] in 1993. The discrete survival methodconsiders grouping the continuous survival time into 𝑘 =1, 2, . . . , 𝐾 disjoint intervals, in which the individual recordswill be replicated 𝑙 times, where 𝑙 is the number of time periodin which the event occurred. The discrete hazard probabilityfunction for time period 𝑘 given a vector of covariates xi isgiven by
ℎ𝑘 (xi,𝑡𝑘) =1
1 + EXP {− (𝛼𝑘𝑡𝑘 + 𝛽𝑡xi)}
, (2)
where 𝑡𝑘 represents the time input for time period 𝑘. Toestimate the conditional hazard in (2) PLANN uses three-layer FFANN with an activation function for the hidden andoutput layers given in (1). The output of the network with an𝐻 number of hidden units is given by
ℎ (xi,𝑡𝑘) = 𝑓(𝑏 +𝐻
∑ℎ=1
𝑤ℎ𝑓(𝑎ℎ +𝑃
∑𝑝=1
𝑤𝑝ℎ𝑥𝑖𝑝)) , (3)
where𝑤𝑝ℎ and𝑤ℎ are the weights of the ANN to be estimatedfor the first layer and second layer, respectively, and also 𝑎ℎand 𝑏 are the weights for the bias connection with the hiddenunits and with the output unit, respectively. The target ofthis network is the censoring indicator 𝑐𝑖𝑘, which is equalto 1 if the event occurred for subject 𝑖 and 0 otherwise. Thecost function used in PLANN is the cross entropy functionwhich is appropriate for binary classification problems [19].The weights of PLANN can be estimated by minimizing thecost function given by
𝐸 = −𝑛
∑𝑖=1
𝑘𝑖
∑𝑘=1
{𝑐𝑖𝑘 log [ℎ (xi,𝑡𝑘)]
+ (1 − 𝑐𝑖𝑘) log [1 − ℎ (xi,𝑡𝑘)]} .
(4)
Once the network weights are estimated, the monotonesurvival probabilities can be easily found by converting thediscrete hazard rate estimates obtained from the networkoutput by the following equation:
𝑆 (𝑡𝑘) =𝑘
∏𝑙=1
(1 − ℎ (𝑡𝑙)) . (5)
The advantage of this approach is that the time dependentcovariates can easily be introduced in the model, as the indi-vidual records are available for each time period.However, forlarge datasets or studies conducted over a long period of timethis approach is inaccessible due to the immense number ofreplication requisites [3]. Figure 2 shows the architecture ofthe PLANN introduced by Biganzoli.
The first layer in Figure 2 contains the bias and one nodefor the time period and the rest of the nodes for the covariates.PLANN uses one input for the time to estimate smoothdiscrete hazard rates. However, we have used 10 nodes for thetime (one for each time period) to be able to compare it withthe second method.
Bias
HiddenInput
Covariates
Bias
Output
Con
ditio
nal h
azar
d pr
obab
ility
.
.
....
P
1 1
22
H
Figure 2: FFANN for partial logistic artificial neural network, withthree layers. The input layer has 𝑝 covariates and hidden layer with𝐻 hidden units and one output unit in the output layer. Activationfunction used in both hidden and output layer is logistic function(1).
2.2.2. Method 2. Mani et al. developed an approach whichwas utilized by Street [20] that predicts the survival functionusing a neural network with 𝐾 outputs, where 𝐾 is thenumber of time periods. He trained his network utilizing atarget vector derived by Kaplan-Meier survival curves [21].Mani used the same neural architecture as Street, but toestimate the hazard function instead. In order to estimate thehazard function, each individual or subject would have atraining vector (1 by 𝐾) target of hazard probabilities ℎ𝑖𝑘 asfollows:
ℎ𝑖𝑘 =
{{{{{{{{{{{
0 for 1 ≤ 𝑘 ≤ 𝐾
1 for 𝑡 ≤ 𝑘 ≤ 𝐾 and event = 1𝑟𝑘𝑛𝑘
for 𝑡 ≤ 𝑘 ≤ 𝐾 and event = 0.
(6)
Here, ℎ𝑖𝑘 = 0 for each time period if patient 𝑖 survived. ℎ𝑖𝑘 = 1from time interval 𝑡 to𝐾 if patient died because of melanomaat duration 𝑡 within the study time. And, for those patientswho are lost to follow-up during the study of duration 𝑡 < 𝐾,their hazards are equal to the ratio 𝑟𝑘/𝑛𝑘, which is the Kaplan-Meier hazard estimate for time interval 𝑘. 𝑟𝑘 is the number ofpatients who died because of melanoma in time period 𝑘, and𝑛𝑘 is the number of patients that are at risk in time interval𝑘. For training the neural network, Mani used the logisticsigmoid function given by
Φ (𝑥) =1
1 + 𝑒−𝑥. (7)
The network weights are estimated by minimizing the costfunction, which is the cross entropy function. Figure 3 showsthe neural network architecture utilized in Method 2. Thenumber of units 𝑝 of the input layer is equivalent to thenumber of independent variables or risk factors. The 𝐾output units of the output layer learn to estimate the hazardprobability of each individual. Once the ANN is trained andthe hazard estimates are predicted, we convert those hazard
4 Advances in Artificial Intelligence
Bias
HiddenInput
Cov
aria
tes
Bias
Output
.
.
.
.
.
.
.
.
.
1
11
22
2
P
KH
hi1
hi2
hiK
Figure 3: Three-layer network with 𝐾 output units in the outputlayer where 𝐾 is equal to the number of time intervals.
estimates to the survival estimates by using (5) (for eachmethod). We have trained the weights of the ANN in bothmethods using the quasi-Newton algorithm.
2.3.Model Selection. Now, we are concernedwith the optimalnumber of hidden units in the hidden layer that will give usthe best neural network model. There are several methods inthe literature that we can use to select the best neural network.The most popular method is the V-fold cross validationmethod as it does not rely on any probabilistic assumptionsand helps in determining when overfitting occurs. Otherstatistical methods like hypothesis testing or informationcriteria were introduced and examined by Anders and Kornin 1999 [22] for neural network model selection, and theysuggested that those statistical methods should take part inneural network modeling. However, since their proposedmethods were based on certain probabilistic assumptions, itmay not be always applicable in modeling real phenomena.
In order to do our comparison we took the best neuralnetwork for each method and then tested their performanceon the same set of data (this set of data was removed fromthe training dataset). In the current study, we have used5-fold cross validation to select the best model for eachmethod (Methods 1 and 2). We divide the male and femaledatasets into six groups. Five were used in the training andvalidation, and the last groupwas used for comparing the bestmodels from the twomethods together (hold-out dataset). Inaddition, we use the weight decay that helps avoid overfittingand penalize large weight solutions to help in generalization.Asmentioned byRipley [23, 24] aweight decay value between𝛼 = 0.01 and 0.1 would be more appropriate depending onthe degree of fit that is expected. We have used the crossvalidation method along with four different values of weightdecay 𝛼 = {0.025, 0.05, 0.075, 0.1}, to pick the best model.Thesame procedure of trying different weight decay values wasused in [9].
The cross validation method will help us in finding theoptimal number of hidden nodes. In addition, we consider
Table 1: Mean prediction error for the eight competing neuralnetwork models for estimating the survival time of male melanomapatients.
Mean St. dv. CountElia’s method𝛼 = 0.025 −0.3776 1.7876 85460𝛼 = 0.050 −0.4113 1.9388 85460𝛼 = 0.075 −0.308 1.4295 85460𝛼 = 0.1 −0.3366 1.5329 85460
Mani’s method𝛼 = 0.025 −0.0236 0.2029 85460𝛼 = 0.050 −0.0215 0.1710 85460𝛼 = 0.075 −0.0197 0.1539 85460𝛼 = 0.1 −0.0197 0.1485 85460
the model with the lowest prediction error when appliedto a new data. Therefore, for each method, we picked thebest model (with lowest cross validation error) and thencompared its performance on the hold-out dataset. Werepeated our comparison for the four values of weight decay,since for the same data two factors affect ANN performance(number of hidden units and weight decay value).
3. Results
In our analysis, we used ten time intervals (12 months each)and in order to do the comparison between the PLANN andMani’s method we used 10 inputs for the ten time intervals inPLANN instead of one so that we can compare the output ofthe PLANN with the second method.
After training, the cross validation method resulted inchoosing the networks with 52 hidden units (number ofhidden nodes seems to be large but, by taking into accountthe number of output units, we have 10 parallel networks withfive hidden nodes each) for both methods as the best model.We obtain similar results from ANN trained with the fourdifferent values of weight decay. Still, we want to examine theprediction accuracy for all eight available models to chooseour best-fit model. Table 1 exhibits the comparison betweenthe eight models for male melanoma patients and Table 2exhibits the comparison for the female melanoma patients.
It is clear from Table 1 that using Mani’s method yieldsbetter predictive neural network model than the PLANNproposed by Elia. Among the four competing models ofMani’s method we have chosen the model with weight decay𝛼 = 0.1 as the best-fit model for predicting survival times ofmale melanoma patients that yield smaller mean error andstandard deviation.
The results in Table 2 support the decision we have madefor male melanoma patients that Mani’s method has a betterpredictive accuracy than that of the PLANN. But the bestmodel for predicting the female melanoma patient’s survivaltime is the model with weight decay value 𝛼 = 0.075.
Advances in Artificial Intelligence 5
24
68
10
2040
6080
100
0.75
0.8
0.85
0.9
0.95
1
Time
Age at diagnosis
Surv
ival
pro
babi
lity
(a)
24
68
10
2040
6080
100
0.95
0.96
0.97
0.98
0.99
1
TimeAge
Surv
ival
pro
babi
lity
(b)
Figure 4: Survival probability function surface plot results for age: (a) for male melanoma patients and (b) for female melanoma patients.Thesurvival probability is estimated as a function of time (ten-year period) and age. Other risk factors like tumor thickness are fixed at 0.58mmwith no ulceration and in the initial stage.
Table 2: Mean prediction error for the eight competing neural net-work models for estimating the survival time of female melanomapatients.
Mean St. dv. CountElia’s method𝛼 = 0.025 −0.3384 2.1326 69130𝛼 = 0.050 −0.2274 14451 69130𝛼 = 0.075 −0.1882 1.1672 69130𝛼 = 0.1 −0.2011 1.2199 69130
Mani’s method𝛼 = 0.025 −0.0208 0.2310 69130𝛼 = 0.050 −0.0183 0.1849 69130𝛼 = 0.075 −0.0167 0.1606 69130𝛼 = 0.1 −0.0246 0.3022 69130
4. Discussion
In modeling survival data with ANN, it is more prevalent toutilize Method 2 (Mani’s method). The prediction accuracyis much better compared to the PLANN. However, theresults may change if somehow the survival data containstime varying covariates (risk factors). One can attempt toamend the PLANN model by differentiating between theindividuals who survived the whole duration time andthose who dropped out during the duration time, whichopens another area of research in this field. With regardto learning techniques for ANN, Lisboa et al. [25] haveamended the PLANN by adapting the Bayesian learning forneural networks, developed by Mackay in 1995 [26]. It isstill an open problem: how the Bayesian learning will affectthe performance of Mani’s ANN? Is it going to change thecomparison results with the PLANN?These are among other
questions that we need answers to and we open more areas ofresearch on this type of problems.
It is clear to us that male and female melanoma patientsneed to be treated differently as shown by the survival plotsin Figure 4, which displays the surface plot of survival prob-ability of males (Figure 4(a)) and females (Figure 4(b)). Inthis figure, the survival is estimated as a function of age atdiagnosis and time in years. The tumor thickness is 0.58mmand ulceration variable is set to 0, considering that the patientwas diagnosed in the initial stage. Male infant patients haveless survival probability than that of female infant patientsover a 10-year period, whereas a male patient at the age of 40to 50 seems to have higher survival probabilities compared tofemale patients at the same age.
Figure 5 displays the surface plot of survival results formale melanoma patients for tumor thickness ranging from0.01mm to 9mm. The surface plot (Figure 5(a)) is formale patient diagnosed at 20 years of age, whereas the plot(Figure 5(b)) is for male patient diagnosed at the age of 60years. As we can see, survival estimate for young men isfarther away and lower than that of older men and thesefindings were found similar to a recent study by Fisher andGeller in 2013 [27].
Fisher and Geller mentioned that more attention wasgiven to older men over the past years and suggested thatmore awareness is needed to be addressed to young men tohelp in early detection of melanoma. They also mentionedthe difference between young men and young women, whichwe can figure out by comparing the two left plots of Figures5 and 6. The survival probability for young men (diagnosedwith tumor thickness larger than 4mm) within two years ofdiagnosis is too low (almost 0) compared to that of youngwomen.
6 Advances in Artificial Intelligence
200400
600800
100000.10.20.30.40.50.60.70.80.9
Tumor thickness
Surv
ival
pro
babi
lity
24
68
10
Time
(a)
24
68 10
200400600800
1000
0.75
0.8
0.85
0.9
0.95
1
TimeTumor thickness
Surv
ival
pro
babi
lity
(b)
Figure 5: Survival probability function surface plot results for tumor thickness: (a) for the male patient diagnosed at age of 20 years old and(b) for male diagnosed at the age of 60 years.
12
34
56
78
9 10
200400
600800
1000
0.4
0.5
0.6
0.7
0.8
0.9
1
Time
Tumor thickness
Surv
ival
pro
babi
lity
(a)
24
68
10
200400
6008001000
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
TimeTumor thickness
Surv
ival
pro
babi
lity
(b)
Figure 6: Survival probability function surface plot results for tumor thickness: (a) for the female patient diagnosed at the age of 20 yearsand (b) for female diagnosed at the age of 60 years.
Some of our significant findings were found to be similarto those found in another study by Gamba et al. in 2013 [28].However, more investigation and statistical data analysis arerequired to better understand the causes of the differencesbetween young males and females and to plan new strategiesto fight themajor pernicious form of skin cancer (melanoma)[29].
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper.
References
[1] C. P. I. van Hinsbergen, J. W. C. van Lint, and H. J. vanZuylen, “Bayesian committee of neural networks to predicttravel times with confidence intervals,” Transportation ResearchPart C: Emerging Technologies, vol. 17, no. 5, pp. 498–509, 2009.
[2] G. B. Kingston, M. F. Lambert, and H. R. Maier, “Bayesiantraining of artificial neural networks used for water resourcesmodeling,” Water Resources Research, vol. 41, no. 12, Article IDW12409, pp. 1–11, 2005.
[3] B. Baesens, T. van Gestel, M. Stepanova, D. van den Poel, andJ. Vanthienen, “Neural network survival analysis for personal
Advances in Artificial Intelligence 7
loan data,” Journal of the Operational Research Society, vol. 56,no. 9, pp. 1089–1098, 2005.
[4] D.-R. Chen, R.-F. Chang, W.-J. Kuo, M.-C. Chen, and Y.-L.Huang, “Diagnosis of breast tumors with sonographic textureanalysis using wavelet transform and neural networks,” Ultra-sound in Medicine and Biology, vol. 28, no. 10, pp. 1301–1310,2002.
[5] F. Ercal, A. Chawla, W. V. Stoecker, H.-C. Lee, and R. H. Moss,“Neural network diagnosis of malignant melanoma from colorimages,” IEEE Transactions on Biomedical Engineering, vol. 41,no. 9, pp. 837–845, 1994.
[6] S.-J. Soong, S. Ding, D. Coit et al., “Predicting survival outcomeof localized melanoma: an electronic prediction tool based ontheAJCCmelanoma database,”Annals of Surgical Oncology, vol.17, no. 8, pp. 2006–2014, 2010.
[7] D. R. Cox, “Regression models and life-tables,” Journal of theRoyal Statistical Society, Series B: Methodological, vol. 34, no. 2,pp. 187–220, 1972.
[8] D. Faraggi and R. Simon, “A neural network model for survivaldata,” Statistics in Medicine, vol. 14, no. 1, pp. 73–82, 1995.
[9] E. Biganzoli, P. Boracchi, L.Mariani, andE.Marubini, “Feed for-ward neural networks for the analysis of censored survival data:a partial logistic regression approach,” Statistics inMedicine, vol.17, no. 10, pp. 1169–1186, 1998.
[10] D. R. Mani, J. Drew, A. Betz, and P. Datta, “Statistics and datamining techniques for lifetime value modeling,” in Proceedingsof the 5th ACM SIKGKDD International Conference on Knowl-edge Discovery and Data Mining (KDD ’99), pp. 94–103, SanDiego, Calif, USA, August 1999.
[11] L. Bottaci, P. J. Drew, J. E. Hartley et al., “Artificial neuralnetworks applied to outcome prediction for colorectal cancerpatients in separate institutions,”The Lancet, vol. 350, no. 9076,pp. 469–472, 1997.
[12] P. Lapuerta, S. P. Azen, and L. LaBree, “Use of neural networksin predicting the risk of coronary artery disease,”Computers andBiomedical Research, vol. 28, no. 1, pp. 38–52, 1995.
[13] L. Ohno-Machado, “Sequential use of neural networks forsurvival prediction inAIDS,” inProceedings of theAMIAAnnualSymposium Proceedings Archive, pp. 170–174, 1996.
[14] P. M. Ravdin and G. M. Clark, “A practical application of neuralnetwork analysis for predicting outcome of individual breastcancer patients,” Breast Cancer Research and Treatment, vol. 22,no. 3, pp. 285–293, 1992.
[15] Surveillance Epidemiology and End Results (SEER) Pro-gram, Research Data (1973–2009), Division of Cancer Controland Population Sciences, National Cancer Institute, Surveil-lance Research Program, Surveillance Systems Branch, 2012,http://www.seer.cancer.gov/.
[16] F. Rosenblatt, The Perceptron: A Perceiving and RecognizingAutomaton, Cornell Aeronautical Laboratory, 1957.
[17] P. D. Allison, “Discrete-time methods for the analysis of eventhistories,” Sociological Methodology, vol. 13, pp. 61–98, 1982.
[18] J. D. Singer and J. B.Willett, “It’s about time: using discrete-timesurvival analysis to study duration and the timing of events,”Journal of Educational and Behavioral Statistics, vol. 18, no. 2,pp. 155–195, 1993.
[19] C. M. Bishop, Neural Networks for Pattern Recognition, OxfordUniversity Press, New York, NY, USA, 1995.
[20] W. N. Street, “A neural network model for prognostic pre-diction,” in Proceeding of the 15th International Conference onMachine Learning, San Francisco, Calif, USA, 1998.
[21] E. L. Kaplan and P. Meier, “Nonparametric estimation fromincomplete observations,” Journal of the American StatisticalAssociation, vol. 53, pp. 457–481, 1958.
[22] U. Anders and O. Korn, “Model selection in neural networks,”Neural Networks, vol. 12, no. 2, pp. 309–323, 1999.
[23] B. D. Ripley, “Neural networks and flexible regression anddiscrimination,” Journal of Applied Statistics, vol. 21, no. 1-2, pp.39–57, 1994.
[24] B. D. Ripley, Pattern Recognition and Neural Networks, Cam-bridge University Press, Cambridge, Cambridge, UK, 1996.
[25] P. J. G. Lisboa, H. Wong, P. Harris, and R. Swindell, “A Bayesianneural network approach for modelling censored data withan application to prognosis after surgery for breast cancer,”Artificial Intelligence in Medicine, vol. 28, no. 1, pp. 1–25, 2003.
[26] D. J. C. MacKay, “Probable networks and plausible predic-tions—a review of practical Bayesian methods for supervisedneural networks,”Network: Computation in Neural Systems, vol.6, no. 3, pp. 469–505, 1995.
[27] D. E. Fisher and A. C. Geller, “Disproportionate burden ofmelanoma mortality in young US men: the possible role ofbiology and behavior,” JAMA Dermatology, vol. 149, no. 8, pp.903–904, 2013.
[28] C. S. Gamba, C. A. Clarke, T. H. M. Keegan, L. Tao, and S.M. Swetter, “Melanoma survival disadvantage in young, non-hispanic white males compared with females,” JAMA Derma-tology, vol. 149, no. 8, pp. 912–920, 2013.
[29] A. C. Society, “Cancer Facts & Figures 2013,” 2013, http://www.cancer.org/acs/groups/content/@epidemiologysurveilance/doc-uments/document/acspc-036845.pdf.
Submit your manuscripts athttp://www.hindawi.com
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttp://www.hindawi.com
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Applied Computational Intelligence and Soft Computing
Advances in
Artificial Intelligence
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporation
http://www.hindawi.com Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Modelling & Simulation in EngineeringHindawi Publishing Corporation http://www.hindawi.com Volume 2014
The Scientific World JournalHindawi Publishing Corporation http://www.hindawi.com Volume 2014
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttp://www.hindawi.com Volume 2014