World Journal of Urology (2020) 38:2469–2476

https://doi.org/10.1007/s00345-020-03080-8

ORIGINAL ARTICLE

Long short‑term memory artificial neural network model

for prediction of prostate cancer survival outcomes according to initial
treatment strategy: development of an online decision‑making
support system
Kyo Chul Koo1 · Kwang Suk Lee1 · Suah Kim2 · Choongki Min2 · Gyu Rang Min1 · Young Hwa Lee1 · Woong Kyu Han1 ·
Koon Ho Rha1 · Sung Joon Hong1 · Seung Choul Yang1 · Byung Ha Chung1 

Received: 30 August 2019 / Accepted: 3 January 2020 / Published online: 10 January 2020
© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Purpose  The delivery of precision medicine is a primary objective for both clinical and translational investigators. Patients
with newly diagnosed prostate cancer (PCa) face the challenge of deciding among multiple initial treatment modalities. The
purpose of this study is to utilize artificial neural network (ANN) modeling to predict survival outcomes according to initial
treatment modality and to develop an online decision-making support system.
Methods  Data were collected retrospectively from 7267 patients diagnosed with PCa between January 1988 and December
2017. The analyses included 19 pretreatment clinicopathological covariates. Multilayer perceptron (MLP), MLP for N-year
survival prediction (MLP-N), and long short-term memory (LSTM) ANN models were used to analyze progression to
castration-resistant PCa (CRPC)-free survival, cancer-specific survival (CSS), and overall survival (OS), according to initial
treatment modality. The performances of the ANN and the Cox-proportional hazards regression models were compared
using Harrell’s C-index.
Results  The ANN models provided higher predictive power for 5- and 10-year progression to CRPC-free survival, CSS, and
OS compared to the Cox-proportional hazards regression model. The LSTM model achieved the highest predictive power,
followed by the MLP-N, and MLP models. We developed an online decision-making support system based on the LSTM
model to provide individualized survival outcomes at 5 and 10 years, according to the initial treatment strategy.
Conclusion  The LSTM ANN model may provide individualized survival outcomes of PCa according to initial treatment
strategy. Our online decision-making support system can be utilized by patients and health-care providers to determine the
optimal initial treatment modality and to guide survival predictions.

Keywords  Artificial intelligence · Decision support techniques · Prostate cancer · Survival

Abbreviations CSS Cancer-specific survival

ANN Artificial neural network CV Cross-validation
AUC​ Area under the curve LSTM Long short-term memory
CRPC Castration-resistant prostate cancer MLP Multilayer perceptron
MLP-N MLP for N-year survival prediction
OS Overall survival
Kyo Chul Koo and Kwang Suk Lee equally contributed to the PCa Prostate cancer
study.

* Byung Ha Chung
chung646@yuhs.ac Introduction
1
Department of Urology, Yonsei University College Patients with newly diagnosed prostate cancer (PCa)
of Medicine, 211 Eonju‑ro, Gangnam‑gu, Seoul 135‑720,
are faced with the challenge of selecting their preferred
Republic of Korea
2
option among multiple initial treatment modalities [1].
Selvas AI, Seoul, Republic of Korea

13
Vol.:(0123456789)

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


2470
The generally available modalities for initial PCa treat-
ment include active surveillance (AS), radical prosta-
tectomy (RP), radiation therapy (RT) with or without
androgen deprivation therapy (ADT), and ADT alone.
Contemporary guidelines recommend treatments based
on the patient’s life expectancy and the average-risk of
the disease; however, the guidelines do not provide a sin-
gle superior treatment option [2–4]. Furthermore, patients
have a range of personal values, preferences, and baseline
function; therefore, preference-sensitive treatment selec-
tion is a complex decision for patients as well as health-
care providers [4, 5].
In the era of precision medicine, stratification of indi-
vidual patients is emphasized to select the most appropri-
ate treatment modality. For PCa treatment, prediction of
cancer-specific survival (CSS) is a critical element in the
decision-making process. D’Amico et al. and Cooperberg
et al. developed risk stratifications tools to predict CSS,
with 80% discrimination in internal validation [6, 7]. How-
ever, these tools did not account for prognostic indicators,
such as performance status or comorbidities that affect
survival and were unable to provide survival outcomes
according to treatment modality. Moreover, the calcula-
tions used conventional linear models, such as survival
analysis and the Cox-proportional hazard model. Given the
nonlinear and multidimensional relationship among dis-
tinct prognostic factors in PCa biology, it would be ardu-
ous to determine survival outcomes using conventional
linear analyses alone.
With improvements in computer-aided techniques,
machine learning methods are playing an increasingly
important role in cancer diagnosis and prognosis algo-
rithms [8]. An artificial neural network (ANN) is a math-
ematical and computational method that has been applied
in several cancers, achieving higher predictive accuracy
for survival than conventional linear discriminant analy-
ses [9, 10]. Given multiple effective treatment modalities
for a particular stage of PCa and the complex and non-
linear disease landscape of PCa, we hypothesized that an
ANN model might better stratify individual patients for
selection of an optimal initial treatment modality. Such an
ANN model may assist patients and health-care providers
to make treatment decisions that are concordant with the
patient’s preferences and may ultimately minimize future
regret and anxiety regarding the treatment selected.
The aims of our study were to (1) develop and compare
three types of ANN models for predicting progression to
castration-resistant PCa (CRPC)-free survival, CSS, and
overall survival (OS) according to initial treatment modal-
ity and (2) to construct an online decision-making support
system that provides probabilities of each survival end-
point according to initial treatment modality.
13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
World Journal of Urology (2020) 38:2469–2476
Material and methods
Study cohort and data collection
Clinicopathological records were collected from 9864 con-
secutive patients diagnosed with biopsy-confirmed PCa
at two high-volume institutions between January 1988
and December 2017. Data were retrieved from the SCaP
database, a longitudinal observational data repository that
prospectively collects clinicopathological data and sur-
vival outcomes of patients diagnosed with PCa. Patients
were excluded from the analyses if they met the following
criteria: (1) incomplete clinical data; (2) adjuvant treat-
ments that defied standard recommendations, (3) lost to
follow-up; or (4) unknown cause of death. The final cohort
consisted of 7267/9864 (73.7%) patients. The distribution
of patients along the time frame according to initial treat-
ment modality is described in Fig. 1. The proportion of
patients who received RT was lower than that in West-
ern collectives, presumably due to the treatment selection
being made mostly by urologists and the widespread belief
of Korean men that RT would result in more side effects.
This study was approved by the institutional ethics com-
mittee (3-2016-0190), which waived the requirement for
informed consent for patients from Gangnam and Shin-
chon Severance Hospitals. All study procedures complied
with the principles of the 1946 Declaration of Helsinki
and its 2008 update.
Pretreatment patient age, body mass index, Eastern
Cooperative Oncology Group (ECOG) performance score,
Charlson comorbidity index, hypertension, diabetes mel-
litus, tuberculosis, liver cirrhosis, cerebrovascular disease,
prostate-specific antigen (PSA) level, prostate volume,
PSA density, positive biopsy core number, % maximum
biopsy core involvement, clinical TNM stage, biopsy Glea-
son score, presence of second primary malignancies, and
type of initial treatment modality were reviewed. ECOG
performance score is a measurement tool to describe
a patient’s level of functioning in terms of their ability
to care for themselves, daily activity, and physical abil-
ity. Clinicopathological variables were selected based on
existing literature regarding significant prognosticators of
PCa survival [2–4, 11, 12]. For accuracy of training, vari-
ables with more than 25% of missing data were excluded.
Initial treatment modalities were categorized as follows:
AS, RP, RT with and without ADT, and ADT alone. Selec-
tion of the initial treatment modality was based on surgeon
discretion and patient preference. The survival status and
cause of death were collected using institutional medical
records or the National Cancer Registry Database. Death
was attributed to PCa if PCa was listed on the death cer-
tificate as the cause of death, progressive metastatic CRPC
World Journal of Urology (2020) 38:2469–2476 2471
Fig. 1  Distribution of patients
along the study timeframe
according to initial treatment
modality
was present, or the patient died of complications of PCa
treatment.
Study endpoints
Co-primary endpoints were progression to CRPC-free sur-
vival, CSS, and OS analyzed by Cox-proportional hazards
and ANN models. The secondary endpoint was the develop-
ment of an online decision-making support system based on
ANN analytics.
Statistical analyses
Development of the artificial neural network
Three ANN models were developed according to predefined
methods: the multilayer perceptron (MLP), the MLP for
N-year survival prediction (MLP-N), and the long short-
term memory (LSTM) models [13–15]. All models were
three-layer, feed-forward ANNs with sigmoid activations.
The networks were trained using a standard back-
propagation algorithm, which contained an input, hidden,
and output layers. The ANNs learned the relationships
between independent and dependent variables. There were
20 nuclear features in the input layer, while the numbers
of hidden layers and nodes for MLP, MLP-N, and LSTM
models were 2 and 64, 3 and 128, and 1 and 64, respec-
tively. The output layer of the MLP model had 10 nodes
representing the survival rates from 1 to 10 years. The
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
output layer of the MLP-N model had one output node
representing the survival of each patient at the 5th or 10th
year. The LSTM model had an output layer with one out-
put node representing the probability of survival at each
time point. The training signal for each node represented
the probability of survival for the example at that time
point. The adaptive learning and dropout rates were 0.002
and 0.5, respectively, for all models.
Fivefold cross-validation (CV) was implemented with
each of the classifiers. After dividing the entire dataset
into five equal-sized groups, the first four groups were
used as training data and the fifth as test data. The second
through fifth groups were then used as training data and
the first group as test data. This procedure was continued
until each group had been used as test data [16]. A scikit-
learn machine learning library and the TensorFlow deep
learning framework were used to perform the analyses.
Development of the Cox‑proportional hazard regression
model
Cox-regression analyses were performed using the same
training set. Independent prognostic indicators associated
with each survival endpoint in the multivariate analyses
were entered into the Cox-regression model. The sum of
the relative risks was used in the Cox-regression model to
predict survival outcomes according to treatment modality.
13

2472 World Journal of Urology (2020) 38:2469–2476

Comparison of predictive performances Table 1  Clinicopathological characteristics

Number 7267
For all ANN and the Cox-regression models, the predictive
Age (years) 68.0 (63.0–73.0)
performance was determined based on Harrell’s C-index, area
Body mass index (kg/m2) 24.1 (22.4–25.8)
under the curve (AUC), accuracy, sensitivity, specificity, and
PSA (ng/mL) 10.3 (5.9–26.4)
positive and negative predictive values at a particular cut-off
PSA density 0.3 (0.17–0.76)
value. The cut-off values were defined according to sensitiv-
Gleason score
ity analyses using Youden’s Index. Statistical analyses were
  ≤ 6 2330 (32.1%)
performed using ­SPSS® software (version 21.0; IBM Corpo-
  7 1948 (26.8%)
ration, Armonk, NY, USA). Differences with a p value < 0.05
  ≥ 8 2989 (41.1%)
were considered statistically significant.
Positive biopsy core number (out of 12 cores) 3.0 (1.0–6.0)
Maximum core involvement (%) 50.0 (25.0–80.0)
Results Clinical T stage
  ≤ T2a 1628 (22.4%)
Patient characteristics
 T2b-T2c 2744 (37.8%)
  ≥ T3b 2895 (39.8%)
The clinicopathological features of the patients at initial diag-
Clinical N stage
nosis are presented in Table 1. During the median follow-up
 N0 6487 (89.3%)
period of 76.0 months (interquartile range 47.0–106.0), the
 N1 780 (10.7%)
leading cause of death was PCa, followed by second primary
Clinical M stage
malignancy and cardiopulmonary disease (Table 2). The 5-
 M0 6180 (85.0%)
and 10-year survival outcomes of our cohort were consistent
 M1 1087 (15.0%)
with contemporary survival estimates.
Presence of a second primary malignancy 303 (4.2%)
Comorbidity
Comparison of predictive performances
 Hypertension 2974 (40.9%)
 Diabetes mellitus 1099 (15.1%)
The ANN models produced higher predictive performance
 Tuberculosis 87 (1.2%)
than the Cox-regression model for all survival endpoints
 Liver cirrhosis 25 (0.3%)
at both the 5- and 10-year time points. Among the ANN
 Cerebrovascular disease 346 (4.7%)
models, the LSTM model achieved the highest C-indices
CCI
and AUC, followed by the MLP-N and the MLP models
 0 4942 (68.0%)
(Table 3).
 1 1343 (18.5%)
  ≥ 2 982 (13.5%)
Development of a web‑based decision‑making ECOG performance score
support system  0 6924 (95.3%)
 1 75 (1.0%)
A decision-making support system was developed based on
  ≥ 2 268 (3.7%)
the LSTM model and is available at https​://scapc​alcul​ator.
Initial treatment
yuhs.ac/. The schematic diagram of the system is shown
 AS 221 (3.0%)
in Fig. 2. The system is comprised of: data input (Fig. 2a);
 Radical prostatectomy 4701 (64.7%)
analytic service for survival endpoints (Fig. 2b); and visu-
 Radiation therapy without ADT 230 (3.2%)
alizations (Fig. 2c). Visualizations included the patient’s
 Radiation therapy with ADT 347 (4.8%)
Kaplan–Meier survival curves in comparison to the average
 ADT alone 1768 (24.3%)
survival of patients in each NCCN risk category. The system
Follow-up period (months) 76.0 (47.0–106.0)
was built on a Flask framework and deployed on an Nginx
server, a remote and secure server with HTTPS protocols. Data are number (%) and median (interquartile range)
ADT androgen-deprivation therapy, AS  active surveillance,
CCI charlson comorbidity index, CRPC  castration-resistant prostate
cancer,ECOG Eastern Cooperative Oncology Group, PSA  prostate-
Discussion specific antigen

With the development of computer-aided techniques, the

application of ANN modeling is playing an increasingly
important role in improving cancer diagnosis and survival.

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

World Journal of Urology (2020) 38:2469–2476 2473

Table 2  Causes of death and 5- and 10-year survival outcomes Our MLP and MLP-N models used a feature vector of
Deaths, number
aggregated measures to represent time-series data. In the
MLP-N model, training was performed without censored
  Overall 1795 (24.7%) cases from a certain time point without survival data. The
 Cancer-specific 1228 (16.9%) drawback of these models is that a large part of the temporal
 Other-cause 567 (7.8%) information is lost. The LSTM model was a novel approach
  Second primary malignancy 334 (4.6%) capable of accurately identifying temporal correlation in
  Cardiopulmonary disease 90 (1.2%) time-series data. Since there was no need for aggregated
  Cerebrovascular disease 41 (0.57%) measures, all data were presented to the ANN model and
  Renal disease 19 (0.27%) their original time-dependent order was preserved. To obtain
  Liver disease 17 (0.24%) the most accurate estimates from the ANN models, a fivefold
  Self harm 11 (0.15%) CV was implemented with each of the classifiers [23]. The
  Senility 4 (0.06%) CV provides an unbiased estimation; however, CV presents
  Trauma 6 (0.08%) high variance with small samples in some studies [16]. To
  Miscellaneous 34 (0.48%) overcome this limitation, a fivefold CV was chosen over the
  Unknown 11 (0.15%) commonly used tenfold CV due to the relatively small sam-
Survival outcome (%) ple size, leaving more instances for validation and lower
 5-year variance [16].
  Cancer-specific survival 89.4% Contemporary guidelines for PCa treatment recommend
  Overall survival 84.9% shared decision-making in which the patient and health-care
 10-year provider participate actively in selecting the treatment option
  Cancer-specific survival 76.4% that best fits the preference of the patient. Active involve-
  Overall survival 68.1% ment of patients in decision-making and provision of bal-
Data are number (%) anced information on treatment options have been positively
associated with patient-reported outcomes, including per-
ceived quality of care and decision satisfaction [24, 25]. Sev-
The advantages of utilizing ANN in survival prediction eral decision-aid trials have shown modest positive effects
have been shown in several malignancies [9, 10, 17–19]. [26, 27]. However, decision-aids are underused for the lack
In these studies, various ANN models were utilized to of specificity for individuals. Indeed, a comprehensive tool
develop nomograms that outperformed conventional lin- that provides comparative data for different treatment modal-
ear models such as the Cox-proportional hazard model. ities for a specific patient may help create a more success-
Within the disease landscape of PCa, there is a complex ful shared decision-making. Given this unmet clinical need,
non-linear relationship between cancer and host in which our decision-support system was developed and is provided
certain prognosticators show excellent prognostic capa- online at no cost. In addition to the ease of accessibility and
bility in some populations but are suboptimal in others. usability, the prognostic indicators used are readily avail-
Thus, it would be reasonable to utilize ANNs employing able in everyday clinical practice, which supports the general
interconnected groups of nodes that resemble the extensive applicability of our system. We hope that our system may
network of neurons within a human brain. ANNs are par- be utilized during patient counseling to determine the opti-
ticularly suitable for solving complex non-linear datasets, mal initial therapy. Moreover, our system may also be used
such as PCa survival prediction, as they provide accurate to reaffirm the health-care provider’s treatment plan and to
performance in the presence of unreliability, wrong data, reduce future patient regrets regarding treatment decisions.
or measurement errors [17–20]. The strength of this study was the incorporation of
Several models have been developed to predict CSS out- detailed clinicopathological data and wide ranges of poten-
comes for PCa [6, 7, 21, 22]. D’Amico et al. and Cooperberg tial prognostic indicators which may have contributed to the
et al. developed risk stratification tools using Cox-regression favorable results. Of note, the patient and tumor character-
analyses, with 80% discrimination [6, 7]. Stephenson et al. istics of our cohort were similar to those found in West-
developed a nomogram based on competing-risk regression ern PCa patients, with an estimated median CSS consistent
analyses that outperformed previous models with 82% dis- with contemporary survival estimates. Most importantly,
crimination [21]. Korets et al. utilized the Kattan nomogram our models provided higher survival predictive perfor-
and obtained a C-index of 0.67 [22]. The current study was mance than previously reported tools, obviating the need for
the first to develop an ANN model to predict PCa survival patient risk grouping according to arbitrary classifications
endpoints, and the model surpassed the predictive accuracies formulated for average-risk patients. Indeed, such average
of all the models mentioned above. survival estimates according to risk category would likely

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


2474 World Journal of Urology (2020) 38:2469–2476

Table 3  Prognostic accuracies of the artificial neural network and the conventional linear models
Survival endpoint Time point Model C-index (95% CI) AUC (95% CI) Accuracy Sensitivity Specificity PPV NPV

CRPC progression 5-year Cox 0.885 (0.853–0.898) 0.905 (0.866–0.923) 0.847 0.829 0.849 0.425 0.974
MLP 0.912 (0.888–0.927) 0.933 (0.902–0.948) 0.858 0.882 0.854 0.445 0.983
MLP-5 0.910 (0.889–0.923) 0.932 (0.906–0.954) 0.856 0.878 0.853 0.453 0.982
LSTM 0.914 (0.890–0.928) 0.936 (0.906–0.949) 0.855 0.887 0.851 0.435 0.983
10-year Cox 0.885 (0.853–0.898) 0.894 (0.866–0.912) 0.834 0.779 0.872 0.803 0.856
MLP 0.913 (0.888–0.928) 0.917 (0.894–0.934) 0.844 0.886 0.816 0.760 0.914
MLP-10 0.904 (0.878–0.919) 0.915 (0.895–0.932) 0.848 0.856 0.843 0.782 0.898
LSTM 0.914 (0.890–0.928) 0.920 (0.899–0.936) 0.846 0.878 0.827 0.770 0.910
CSM 5-year Cox 0.817 (0.812–0.822) 0.852 (0.837–0.871) 0.771 0.787 0.770 0.323 0.963
MLP 0.849 (0.845–0.861) 0.857 (0.831–0.887) 0.811 0.755 0.823 0.475 0.943
MLP-5 0.853 (0.843–0.865) 0.890 (0.877–0.909) 0.817 0.835 0.815 0.385 0.973
LSTM 0.857 (0.846–0.865) 0.893 (0.876–0.911) 0.802 0.856 0.795 0.373 0.975
10-year Cox 0.817 (0.812–0.822) 0.825 (0.813–0.839) 0.758 0.759 0.760 0.733 0.785
MLP 0.850 (0.845–0.861) 0.822 (0.806–0.842) 0.752 0.714 0.803 0.820 0.693
MLP-10 0.846 (0.832–0.858) 0.858 (0.839–0.873) 0.788 0.792 0.783 0.762 0.814
LSTM 0.857 (0.846–0.865) 0.860 (0.846–0.874) 0.795 0.758 0.827 0.792 0.798
OM 5-year Cox 0.779 (0.764–0.799) 0.815 (0.794–0.846) 0.772 0.721 0.783 0.407 0.932
MLP 0.804 (0.789–0.824) 0.843 (0.818–0.873) 0.808 0.732 0.823 0.464 0.938
MLP-5 0.809 (0.791–0.831) 0.849 (0.820–0.881) 0.805 0.740 0.817 0.455 0.940
LSTM 0.815 (0.796–0.834) 0.856 (0.826–0.885) 0.795 0.768 0.811 0.442 0.945
10-year Cox 0.779 (0.764–0.799) 0.793 (0.774–0.809) 0.727 0.694 0.771 0.794 0.672
MLP 0.804 (0.789–0.824) 0.807 (0.788–0.826) 0.738 0.696 0.792 0.808 0.678
MLP-10 0.798 (0.785–0.819) 0.820 (0.805–0.840) 0.749 0.724 0.779 0.804 0.694
LSTM 0.815 (0.796–0.834) 0.830 (0.811–0.848) 0.964 0.751 0.779 0.811 0.715

Data are median (interquartile range)

AUC area under the curve, CRPC  castration-resistant prostate cancer, CSM cancer-specific mortality, LSTM long short-term memory, MLP  multi-
layer perceptron, NPV negative predictive value, OM overall mortality, PPV positive predictive value

not be appropriate for all patients, especially for those with
a complex disease presentation. In this regard, our system
could be most useful for patients with atypical or unusual
disease features, such as advanced low-grade or localized
high-grade diseases.
Our study had several limitations. First, all patients
included in this study were of Asian ethnicity. An external
validation using multi-ethnic population data with “ethnic-
ity” as an additional “node” would be warranted to ensure
similar sensitivity and specificity. Moreover, a population-
based database with a larger number of subjects may provide
better generalizability. Albeit, we utilized our institutional
data, which provided a comprehensive and high-quality data-
set, to maximize the predictive power. Second, information
on long-term treatment toxicity and side effects were not
accounted for in the analysis. Preserving the quality of life
and accommodating patient treatment preferences are as
important as identifying the treatment that would maximize
the oncological outcome. Third, the main scope of this study
was to determine the optimal initial treatment modality at
the time of PCa diagnosis; therefore, predictions of survival
during the later stages of treatment were unavailable. The
inclusion of comprehensive data regarding subsequent treat-
ments may allow for real-time survival prediction in which
the prognosis can be reassessed according to previous treat-
ment, comorbidities, and nutritional and performance sta-
tuses, as a result of which the patient could be allocated
to certain multidisciplinary treatments. Lastly, data were
collected over a long time span, during which treatment
modalities and systemic agents had considerably improved.
Therefore, survival outcomes estimated by our ANN model
may not represent outcomes that would be expected in the
current era.
Conclusions
An LSTM ANN model may provide individualized long-
term survival prognoses for patients diagnosed with PCa
according to initial treatment strategy. Our online decision-
making support system can be utilized in everyday clinical
practice by patients and health-care providers to counsel

13

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

 World Journal of Urology (2020) 38:2469–2476

Content courtesy of Springer Nature, terms of use apply. Rights reserved.

Fig. 2  Schematic diagram of the online decision-making support system: (A) data input; (B) analytic service for survival endpoints; and (C) visualizations

13
2475

2476
patients regarding available treatment options and to deter-
mine the optimal treatment strategy for the best survival
outcome.
Acknowledgments  This study was supported through a Young
Researcher Program Grant from the National Research Foundation of
Korea (NRF-2017R1C1B5017516).
Author contributions  Protocol/project development: Koo, KS Lee,
Han, Rha, Hong, Yang and Chung. Data collection and management:
Koo, KS Lee, YH Lee, and GR Min. Data analysis: Kim and C Min.
Manuscript writing/editing: Koo and Chung.
Compliance with ethical standards  14. Chi CL, Street WN, Wolberg WH (2007) Application of artificial
Conflict of interest  All of the authors declare that they have no con-
flicts of interest to declare.
Ethical approval  All procedures performed in studies involving human
participants were in accordance with the ethical standards of the insti-
tutional and/or national research committee and with the 1964 Helsinki
Declaration and its later amendments or comparable ethical standards.
Informed consent  Informed consent was not required for the purposes
of this study as it was based upon retrospective anonymous patient
data and did not involve patient intervention or the use of human tis-
sue samples.
References
1. Koo KC, Cho JS, Bang WJ et al (2018) Cancer-specific mortality
among Korean men with localized or locally advanced prostate
cancer treated with radical prostatectomy versus radiotherapy: a
multi-center study using propensity scoring and competing risk
regression analyses. Cancer Res Treat 50:129–137
2. Mohler JL, Armstrong AJ, Bahnson RR et al (2016) Prostate can-
cer, version 1.2016. J Natl Compr Canc Netw 14:19–30
3. Mottet N, Bellmunt J, Bolla M et al (2017) EAU-ESTRO-SIOG
guidelines on prostate cancer. Part 1: screening, diagnosis, and
local treatment with curative intent. Eur Urol 71:618–629
4. Sanda MG, Cadeddu JA, Kirkby E et al (2018) Clinically local-
ized prostate cancer: AUA/ASTRO/SUO guideline. Part I: risk
stratification, shared decision making, and care options. J Urol
199:683–690
5. Cuypers M, Lamers RED, Cornel EB et al (2018) The impact
of prostate cancer diagnosis and treatment decision-making on
health-related quality of life before treatment onset. Support Care
Cancer 26:1297–1304
6. D’Amico AV, Cote K, Loffredo M et al (2002) Determinants
of prostate cancer-specific survival after radiation therapy for
patients with clinically localized prostate cancer. J Clin Oncol
20:4567–4573
7. Cooperberg MR, Broering JM, Carroll PR (2009) Risk assess-
ment for prostate cancer metastasis and mortality at the time of
diagnosis. J Natl Cancer Inst 101:878–887
8. Kourou K, Exarchos TP, Exarchos KP et al (2015) Machine learn-
ing applications in cancer prognosis and prediction. Comput
Struct Biotechnol J 13:8–17
9. Spelt L, Nilsson J, Andersson R et al (2013) Artificial neural
networks–a method for prediction of survival following liver
resection for colorectal cancer metastases. Eur J Surg Oncol
39:648–654
13
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
World Journal of Urology (2020) 38:2469–2476
10. Peng JH, Fang YJ, Li CX et al (2016) A scoring system based
on artificial neural network for predicting 10-year survival in
stage II a colon cancer patients after radical surgery. Oncotarget
7:22939–22947
11. Litwin MS, Tan HJ (2017) The diagnosis and treatment of prostate
cancer: a review. JAMA 317:2532–2542
12. Campbell JM, Raymond E, O’Callaghan ME et al (2017) Opti-
mum tools for predicting clinical outcomes in prostate cancer
patients undergoing radical prostatectomy: systematic review
of prognostic accuracy and validity. Clin Genitourin Cancer
15:e827–e834
13. Chang SW, Abdul-Kareem S, Merican AF et al (2013) Oral cancer
prognosis based on clinicopathologic and genomic markers using
a hybrid of feature selection and machine learning methods. BMC
Bioinform 14:170
neural network-based survival analysis on two breast cancer data-
sets. AMIA Annu Symp Proc 130–134.
15. Rumelhart D, Hinton G, Williams R et al (1986) Learning internal
representations by error propagation. In: Rumelhart DE, MCclel-
land JL (eds) Parallel distributed processing 1986. MIT Press,
Cambridge, pp 318–362
16. Hochreiter S, Schmidhuber J (1997) Long short-term memory.
Neural Comput 9:1735–1780
17. Oh SE, Seo SW, Choi MG et al (2018) Prediction of overall sur-
vival and novel classification of patients with gastric cancer using
the survival recurrent network. Ann Surg Oncol 25:1153–1159
18. Obrzut B, Kusy M, Semczuk A et al (2017) Prediction of 5-year
overall survival in cervical cancer patients treated with radical
hysterectomy using computational intelligence methods. BMC
Cancer 17:840
19. Qiao G, Li J, Huang A et al (2014) Artificial neural networking
model for the prediction of post-hepatectomy survival of patients
with early hepatocellular carcinoma. J Gastroenterol Hepatol
29:2014–2020
20. Snow PB, Rodvold DM, Brandt JM (1999) Artificial neural net-
works in clinical urology. Urology 54:787–790
21. Stephenson AJ, Kattan MW, Eastham JA et  al (2009) Pros-
tate cancer-specific mortality after radical prostatectomy for
patients treated in the prostate-specific antigen era. J Clin Oncol
27:4300–4305
22. Korets R, Motamedinia P, Yeshchina O et al (2011) Accuracy of
the Kattan nomogram across prostate cancer risk-groups. BJU Int
108:56–60
23. Listgarten J, Damaraju S, Poulin B et al (2004) Predictive models
for breast cancer susceptibility from multiple single nucleotide
polymorphisms. Clin Cancer Res 10:2725–2737
24. van Stam MA, Pieterse AH, van der Poel HG et al (2018) Shared
decision making in prostate cancer care-encouraging every patient
to be actively involved in decision making or ensuring the patient
preferred level of involvement? J Urol 200:582–589
25. Stacey D, Legare F, Lewis KB (2017) Patient decision aids
to engage adults in treatment or screening decisions. JAMA
318:657–658
26. Feldman-Stewart D, Tong C, Brundage M et al (2018) Making
their decisions for prostate cancer treatment: patients’ experiences
and preferences related to process. Can Urol Assoc J 12:337–343
27. Harter M, Moumjid N, Cornuz J et al (2017) Shared decision mak-
ing in 2017: International accomplishments in policy, research and
implementation. Z Evid Fortbild Qual Gesundhwes 123–124:1–5
Publisher’s Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

onlineservice@springernature.com