Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Deep Learning-Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer

Deep Learning-Based Survival Analysis Identified Associations Between Molecular Subtype and... abstract original reports Deep Learning–Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer Purpose Gastric cancer (GC) is the third-leading cause of cancer-related deaths. Several pivotal clinical trials of adjuvant treatments were performed during the previous decade; however, the optimal regimen for adjuvant treatment of GC remains controversial. Jeeyun Lee Patients and Methods We developed a novel deep learning–based survival model (survival recur- rent network [SRN]) in patients with GC by including all available clinical and pathologic data Ji Yeong An and treatment regimens. This model uses time-sequential data only in the training step, and upon Min Gew Choi being trained, it receives the initial data from the first visit and then sequentially predicts the Se Hoon Park outcome at each time point until it reaches 5 years. In total, 1,190 patients from three cohorts Seung Tae Kim (the Asian Cancer Research Group cohort, n = 300; the fluorouracil, leucovorin, and radiotherapy cohort, n = 432; and the Adjuvant Chemoradiation Therapy in Stomach Cancer cohort, n = 458) Jun Ho Lee were included in the analysis. In addition, we added Asian Cancer Research Group molecular Tae Sung Sohn classifications into the prediction model. SRN simulated the sequential learning process of clini- Jae Moon Bae cians in the outpatient clinic using a recurrent neural network and time-sequential outcome data. Sung Kim Results The mean area under the receiver operating characteristics curve was 0.92 ± 0.049 at Hyuk Lee the fifth year. The SRN demonstrated that GC with a mesenchymal subtype should elicit a more Byung-Hoon Min risk-adapted postoperative treatment strategy as a result of its high recurrence rate. In addition, the SRN found that GCs with microsatellite instability and GCs of the papillary type exhibited sig- Jae J. Kim nificantly more favorable survival outcomes after capecitabine plus cisplatin chemotherapy alone. Woo Kyoung Jeong Conclusion Our SRN predicted survival at a high rate, reaching 92% at postoperative year 5. Our Dong-Il Choi findings suggest that SRN-based clinical trials or risk-adapted adjuvant trials could be considered Kyoung-Mee Kim for patients with GC to investigate more individualized adjuvant treatments after curative gastrec- Won Ki Kang tomy. Mijung Kim Clin Cancer Inform. © 2018 by American Society of Clinical Oncology Licensed under the Creative Commons Attribution 4.0 License Sung Wook Seo INTRODUCTION chemoradiotherapy (CRT) with fluorouracil (FU) Author affiliations and and leucovorin (LV). The Adjuvant Chemoradi- support information (if Gastric cancer (GC) is one of the most fre- applicable) appear at the ation Therapy in Stomach Cancer (ARTIST) trial quently occurring malignancies worldwide and end of this article. was a phase III trial that compared postoperative the third-leading cause of cancer-related deaths Licensed under the treatment with capecitabine plus cisplatin (XP) Creative Commons Attri- worldwide. Most patients with GC present with versus XP plus radiotherapy (RT) in patients bution 4.0 License metastatic disease at recurrence, and the over- 3,4 with extended D2 lymph node dissection. The all prognosis remains poor, with an expected Capecitabine and Oxaliplatin Adjuvant Study in Corresponding author: survival of < 1 year upon recurrence. Several Sung Wook Seo, MD, De- Stomach Cancer trial compared capecitabine pivotal clinical trials were performed in the pre- partment of Orthopaedic plus oxaliplatin treatment with observation in vious decade and aimed at reducing the recur- Surgery, Samsung Med- completely resected GCs and demonstrated an rence rate after curative surgery in patients with ical Center, Sungkyunk- wan University School additional survival benefit with adjuvant capecit- GC. First, the Intergroup 0116 (INT-0116) trial of Medicine, Irwonro80, 5 published in 2001 demonstrated significant abine plus oxaliplatin chemotherapy. The Adju- Gangnamgu, Seoul, Ko- vant Chemotherapy Trial of Titanium Silicate improvement in survival when patients with rea; e-mail: sungwseo@ completely resected GC received postoperative for GC trial compared titanium silicate (TS-1) skku.edu. © 2018 by American Society of Clinical Oncology ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 1 with observation in patients with D2-resected available for the three cohorts: pathology, type of GC and also showed prolonged survival in the surgery, lymphatic invasion, perineural invasion, TS-1 chemotherapy group. Hence, there are at histologic Lauren type, depth of invasion, num- least three to four postoperative chemotherapy ber of dissected lymph nodes, number of positive regimens available for patients with completely lymph nodes (pathologically), age at diagnosis, resected GC. sex, Epstein-Barr virus positivity, human epider- mal growth factor receptor 2 positivity, adjuvant To additionally complicate clinical decision mak- treatment modality, first sites of recurrence at the ing, two recent molecular landscape studies time of diagnosis for recurrence, date of recur- demonstrated the existence of different molec- rence, vital status at last follow-up, and molec- 7,8 ular GC subtypes. The Asian Cancer Research ular subtypes. Detailed information about these Group (ACRG) categories were determined on 8,9 cohorts has been published previously. Briefly, the basis of gene expression profiling of tumors 141 patients in the ACRG cohort received adju- with microsatellite instability (MSI), tumors with vant chemotherapy or CRT. The ARTIST trial was an epithelial-to-mesenchymal transition phe- a prospective adjuvant phase III trial that com- notype, tumors with a p53 signature (CDKN1A pared patients who received six cycles of XP and MDM2 expressing), or tumors without the (n = 228) with patients who received two cycles p53 signature. Notably, the ACRG study iden- of XP followed by RT with capecitabine and then tified four distinct molecular subtypes highly two more cycles of XP (n = 230). In the FU/LV/RT associated with GC recurrence rate and, thus, cohort, all patients (n = 432) received the INT- survival after surgery. To the best of our knowl- 0116 regimen (FU/LV followed by CRT with the edge, no model currently exists that is capable of same agents and then FU/LV again separately) integrating ACRG molecular subtypes and clin- as an adjuvant postoperative treatment. All of icopathologic information, such as stage, Lau- the patients in the FU/LV/RT and ARTIST cohorts ren classification, type of surgery, demographic received adjuvant treatment after D2 resection. data, and molecular subtypes, to predict survival The overall survival rates after 5 and 10 years after surgery. were 63.9% and 56.9%, respectively. Tumor recurrence was observed in approximately 40% In this study, we developed a deep learning– of the patients for each cohort. The protocol based prediction algorithm for survival predic- was approved by the Samsung Medical Center tion in patients with GC on the basis of data from Institutional Review Board (IRB; ACRG: IRB 1,190 patients with GC. The aims of this study No. 2010-12-088), ARTIST (ClinicalTrials.gov were to develop a deep learning–based predic- identifier: NCT0176146), and the DASL (cDNA- tion model to predict survival after surgery on mediated Annealing, S Selection, extension and the basis of sequential prediction of the outcome Ligation) cohort (IRB No. SMC 2010-10-025). at each time point until 5 years after operation and to predict the most optimal postoperative regimen after surgery using a recurrent neural Molecular Subtypes network (RNN) on the basis of available clinical variables. We used a previously published data set (accessed via https://www.ncbi.nlm.nih.gov/geo/ query/acc.cgi?acc=GSE62254). RNA was extracted PATIENTS AND METHODS from 300 tumors according to manufacturer Patients 8 protocol (Affymetrix, Santa Clara, CA), and we used the Affymetrix human genome U133 Plus We included the following three cohorts from 8 2.0 array (Affymetrix) for gene expression profil- our previous study: the ACRG cohort (n = 296), ing. The ACRG subtypes were the same as those the postoperative FU/LV/RT cohort (n = 432; published previously. Gene Expression Omnibus database identifier: 9 3 GSE26253), and the ARTIST cohort (n = 458). In total, 1,186 patients were included. We pro- Data Preprocessing cured all tissue specimens that were chemother- apy naïve during the primary resection of GC. Primary data were specified according to their No patients received neoadjuvant chemotherapy variable types. Categorical variables, such as or preoperative CRT. The following data were sex, tumor type and location, and molecular 2 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics expression, were converted using one hot encod- Basic Concept of Survival Recurrent Network Model ing, which transforms categorical features to a binary (0 or 1) format group. Ordinal variables A survival recurrent network (SRN) model was and quantitative variables were preprocessed constructed on the basis of logistic regression with normalization. All variables were then trans- function σ at discrete time point (t) and long formed by a standard scaler to standardize the 10,11 short-term memory (LSTM) neural network, intervals of values between the variables. To save which can be represented as follows: the numbers down to the third decimal place, all values were multiplied by 10 and then trans- Survival probability at time  t :    ( ) formed into integers (int32). As such, the value ft X = σ  Wt ∗ Xt =   ( ) ( ) − ω X t t of each variable was transformed into a standard 1 + e score, which was compatible with embedding. Missing data were imputed using the k-nearest neighbor algorithm after separating the training W ∗ X t t HRt(X ) = e and test sets. The event cases during the time interval were ranked by months, and the rank scores were inserted in the censored cases. θ X+ θ t 1 2 HRt(X ) = e Data Separation and Cross-Validation Patients were randomly sorted into a training set (80%) and a test set (20%). The test set was where ft X is the survival probability,   ( ) separated for the final test. Using the training HRt X is the hazard ratio function. ( ) set, bootstrap training (80% of the training set) and validation patients (20% of the training set) Once the parameter vector, Wt = θ1X + θ2t, is were generated by randomly selecting patients optimized with the patient group (features Xt and and repeating the selection 100 times to find target Yt) at the first time point (t), LSTM cells an optimal condition of the neural networks. memorize Wt and the model is retrained with the The validation error reached to the minimum at target value (Yt+1) of the next time point yielding epoch 7. The optimized model was tested over Wt+1. For example, if a patient (X) died of dis- the separated test patients. ease 2 years after the first visit, the model should learn the target value (Y1 = 1) at the first year and learn the target value (Y2 = 0) at the second Performance Evaluation and Statistics year. Because LSTM memorizes and optimizes W for each target value, our RNN-based model Receiver operating characteristics (ROC) curves, is able to infer a target value at a certain time areas under the receiver operating character- point. istic curve (AUCs), and concordance index (c-index) were compared using a nonparamet- Patient factors (X) were hardly updated at every time point because we could not collect all those ric Mann-Whitney U test using the MedCalc program (MedCalc Software, Seoul, Korea). All data without any loss. Moreover, the purpose of the survival model is to predict long-term survival neural networks were constructed using the with the information from the first visit. Thus, Keras with Theano backend in Python (https:// we generated Xt with the following assumption: keras.io/). The scikit-learn library (http://scikit- patients’ features should be constant during the learn.org/) was used for data management and observation time. Instead, there should be latent preprocessing. The Mann-Whitney U test was features that are dependent to the sequential performed to compare AUCs between models, time and indicate the patients’ status at a dis- whereas the Pearson χ test was performed to crete time point. We defined those latent fea- identify factors related to a specific subgroup. tures as time-dependent life value: The current study was developed and written according to Transparent Reporting of a Multi- variable Prediction Model for Individual Progno- Time‐dependent life value = θ t + sis or Diagnosis model development guidelines. ∂ S ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 3 Time‐dependent hazard function   represents live or dead probability. Forty-seven WtXt θ X+ θ t+ θ ∂S 1 2 3 e   = e features were preprocessed before accessing the input layer, with each feature preprocessed The life value dimensions were embedded in using a standard scaler and each value encoded the constant patient features (X) to generate as an integer within 10,000 scores. The clinical time-dependent features (Xt). The life value fea- variables of an individual were embedded in tures consisted of time (t) and prior life expec- a 47 × 32 vector for dimensionality reduction. tancy (St), which is updated using gradient The other two layers comprised fully connected descent equation ( ∂ S). nodes implementing the rectified linear-unit function. Gaussian dropout was performed to The model is retrained at the following sequen- prevent overfitting problems. The number of tial time with Xt+1 where St+1 is embedded. nodes was gradually reduced across the hidden layers (Appendix Fig A2, online only). The SRN St + 1 = St  +   ∂ S was trained every year sequentially with the 47 clinical features and the previous survival proba- bility. Through this time-sequential training, the ^f t ( x) ^ _____ probability differences among patients became St + 1 = St  +  α Yt −   Y t ∗ Y t ( ) | ft ( x) | more distinct, and the accuracy improved (Appen- dix Fig A3, online only). We randomly sorted 1,186 patients into a train- −wx dft(x) e ing group (80%; n = 950) and a test group (20%; ____ ________ = −  W* ≈  α ∗ ft(x ) (1 − ft ( x) )   −wx 2 dx 1  +   e ( ) n = 236). In the training group, bootstrap train- ing (80%; n = 760) and validation populations where α is a step size and  Y t  = ft(x). Thus, the (20%; n = 190) were generated by randomly function can be rewritten as follows: selecting patients and repeating the selection ^ ^ ^ 100 times. At each time point, the predicted sur- St + 1 = St  +  α(Yt −   Y t ) (1 − Y t ) Y t vival probability of the model was compared with the actual survival data. RESULTS The mean AUC of the 100 training group patients was 0.79 ± 0.052 at the first year, 0.839 ± 0.045 Predictive Accuracy of the SRN Model at the second year, 0.89 ± 0.049 at the third We simulated the sequential learning process year, 0.915 ± 0.05 at the fourth year, and 0.92 ± of clinicians in the outpatient clinic using RNN 0.049 at the fifth year (Fig 1A). The AUC of each and time-sequential outcome data (Appendix time point was compared using the Mann- Fig A1A, online only). The SRN was composed Whitney U test, showing that the AUC improved of the following three learning system parts: significantly in sequential years. The AUCs at the the first included information on patient status fifth and fourth years were significantly higher (covariates, X), the second included the time- than those at the first or second years according dependent life value ( θ t+ θ ∂ S), and the third 2 3 to the SRN model (Fig 1B). included nonparametric rank scores (R) of event that occurred during the interval (0 < R < 1). The R was inserted into the target value instead Performance of the Final Model for Predicting of binary values. The Xt were input into the SRN, Survival of Sample Test Group and the SRN was sequentially retrained with the Using the separated test group, the performance updated Xn (Appendix Fig A1B). The network of our final model was evaluated. The AUC of architecture of the SRN comprised five hidden each time point was 0.858 at the first year, 0.869 layers with two RNN layers (LSTM). The input at the second year, 0.879 at the third year, 0.912 layer comprised 49 nodes that represented 47 at the fourth year, and 0.923 at the fifth year input features and two life value features (phase (Fig 2A). [time] feature and prior survival probability fea- ture). The output layer comprised two nodes The c-index of the final model was evaluated implementing the Softmax function, which with the test group. The c-index was 0.951. 4 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Fig 1. The mean receiver A B operating characteristic (ROC) curve for 100 valida- Receiver Operating Characteristic 1.0 tion data sets. (A) The 100 1.0 trained survival recurrent networks (SRNs) were 0.8 0.9 tested iteratively using the corresponding test patients, 0.6 and ROC curves were 0.8 generated to evaluate the 0.4 average predictive power at each time point. (B) The 0.2 61 0.7 mean area under the ROC curve (AUC) of the test data sets. The AUC of each time 0 0.2 0.4 0.6 0.8 1.0 12345 point was compared using False-Positive Rate Time (years) the Mann-Whitney U test. The AUC of the fifth year was significantly higher than The calibration curve from the SRN was evalu- After a sequential learning process, two clus- that of the second year (P = .00) or the first year (P = ated using the sample test group (Fig 2B). In ters of patients were separated in terms of sur- .00). The AUC of the fourth this curve, the actual survival proportion was vival curves around the postoperative fifth year year was significantly higher compared with the SRN-predicted survival prob- (Appendix Fig A4A, online only). To identify sig- than that of the second year ability, with the actual survival rate calculated nificant factors differentiating the two prognostic (P = .00) or the first year (P = .00). using the Kaplan-Meier method. Our results clusters, we performed the Pearson χ test. Pos- showed that the actual survival rate was closely itive perineural invasion, the presence of lym- correlated with the predicted survival probability phovascular invasion, high tumor or node stage, within 15% margin of error. pathologic stage, a primary tumor located in the antrum or cardia, and recurrence were more The decision curves showed that the SRN model frequently observed to be statistically significant will be useful clinically for predicting the survival in the poor prognostic group in terms of survival of patients with GC. The SRN model has a pos- (Table 1). itive net benefit at all five different time points, rather than assuming all patients or none of the patients will survive at each year (Fig 3). SRN Guidance of the Optimal Adjuvant Treatment Regimen Fig 2. The receiver oper- ating characteristic (ROC) Cumulative Survival Probability of Individual To determine the optimal postoperative regimen curves and the calibration Patients in the Test Group plot of the survival recurrent according to the available data for each patient, network (SRN) for predict- we generated a new simulation data set, in The cumulative survival probability of each ing survival of a sample patient was visualized as a survival graph. which four different treatment options were test group. (A) The ROC and area under the ROC curve of the final model were evaluated using the A B separated test group at each 5-year point. (B) Calibration Receiver Operating Characteristic 1.0 curve of the SRN in the 1.2 test data set. The x-axis represents the probability 1.0 0.8 of 5-year survival, and the y-axis represents the actual 0.8 0.6 survival proportion. The thick solid line represents 0.6 1 year the calibration plot of the 0.4 2 years SRN. The actual survival 0.4 3 years proportion and 95% CIs 0.2 4 years were calculated using the 0.2 5 years Kaplan-Meier method. The prediction scores were 0.0 0.2 0.4 0.6 0.810 0 0.2 0.4 0.6 0.81 1.2 within a 15% margin of the perfect prediction line False-Positive Rate Predicted Survival Probability (dashed line). ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 5 True-Positive Rate True-Positive Rate Actual Survival Proportion ROC_AUC Fig 3. The decision curve A B analysis (dashed line) of the survival recurrent network 0.5 None 0.5 None (SRN) model for predicting All All SRN SRN the survival of sample test 0.4 0.4 group. The prediction model has a positive net benefit 0.3 0.3 if the line is above the lines assuming none of the 0.2 0.2 patients or all of the patients survive at each year. (A) 0.1 0.1 First year. (B) Second year. (C) Third year. (D) Fourth 0.0 0.0 year. (E) Fifth year. The decision curves of the model 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.1 0.2 0.3 0.40.5 show positive net benefit Threshold Probability Threshold Probability rather than assuming that all patients or none of the patients will survive at each CD year. 0.5 None 0.5 None All All SRN SRN 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.60.8 Threshold Probability Threshold Probability 0.5 None All 0.4 SRN 0.3 0.2 0.1 0.0 0.0 0.2 0.4 0.6 0.8 Threshold Probability applied to all patients. Five subgroups were this population from additional analysis. For all adjuvant treatments, significant predictors identified according to the survival probability at for poor responders in terms of survival after the fifth year (Appendix Fig A4B). The treatment adjuvant treatment were mesenchymal subtype, options included in the analysis were as follows: the presence of perineural invasion, advanced XP chemotherapy; XP followed by capecitabine stage, location of the primary tumor in the cardia, plus RT followed by XP; the INT-0116 regimen; and others, such as oral TS-1 chemotherapy. and a greater number of positive lymph nodes Because oral chemotherapy or other regimens (Table 2). Notably, we found that patients with composed < 5% of the options, we excluded GCs with MSI (odds ratio, 4.2; P = .042) and 6 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Net Benefit Net Benefit Net Benefit Net Benefit Net Benefit one of the most popular methods for predicting Table 1. Poor Prognostic Factors for Survival cancer prognoses. A recent Cox model for sur- Variable Relative Risk for Death P vival prediction of patients with GC showed an PNI positive 44.947 — average c-statistic of 0.822. The Cox model Lymphovascular invasion 30.367 — uses the hazard rate of the covariates as model Tumor stage 12.865 — coefficients and, therefore, can encompass fac- Node stage 70.416 — tors that change over time. However, the model assumes that the hazard function is constant Pathologic stage 23.962 — through patient life span; therefore, it cannot Tumor located in cardia 16.567 — accurately predict the risk of death at a certain Recurrence 10.735 .001 time point because it does not represent the Abbreviation: PNI, perineural invasion. changing weight of covariates during the time intervals. A parametric model, such as a linear with GCs of the papillary type (OR, 5.9; P = .015) regression model, measures a direct relation- had significantly better survival outcomes after ship between the covariates and the survival XP chemotherapy alone (Table 3). Subgroup time. If the data from the survival distribution categorization according to SRN resulted in the at each time point are not large enough or if following five major groups (Fig A4B): subgroup there are large numbers of missing values, the I, good prognosis regardless of the adjuvant survival distribution at each time point becomes regimen; subgroup II, better prognosis after unreliable. This model often produces mislead- adjuvant chemotherapy, except for the others ing outcomes in the event of a violation of the group; subgroup III, better prognosis after adju- proportionality of the hazard assumption. Signifi- vant XP plus RT and XP alone; subgroup IV, cant risk factors that are not time dependent are better prognosis after XP alone; and subgroup V, often ignored in these models, thereby poten- poor prognosis regardless of the type of postop- tially resulting in false inferences. Another cause erative regimen. of misleading results is that the residual, the dif- ference between the real and expected values, DISCUSSION is not properly reflected in the model. Another recent model, the least absolute shrinkage and Studies have recently been performed to formal- selection operator, became a popular regres- ize risk prediction in cancer care. Kaplan-Meier sion method as a result of its enhanced predic- curves represent a nonparametric method for tive accuracy acquired by shrin king important estimating survival, with one of the strengths of covariates from thousands of variables, espe- this method being its ability to consider cen- cially when using genetic data. sored data, particularly right-censoring data using the log-rank test. Parametric survival Our model simulated the physician learning models and the Cox proportional hazards model process. Physicians learn about the prognoses might be useful to estimate covariate-adjusted of their patients through serial observations in survival. The Cox proportional hazards model is outpatient clinics. At the first visit, a clinician predicts the patient condition at the next visit on the basis of their current medical status Table 2. Predictive Factors for Poor Responders to Adjuvant Treatment and confirms that prediction during the sub- Variable Relative Risk for Death P sequent visit. We suggest that LSTM, an RNN proposed in 1997 by Sepp Hochreiter and Mesenchymal subtype 10.465 .001 10,11 Jürgen Schmidhuber, represents the optimal PNI positive 10.639 .001 choice for a serial learning system. LSTM was Signet ring 13.235 — designed to avoid the long-term dependency Location of cardia 14.516 — problem of RNNs and comprises input, out- Diffuse type by Lauren classification 16.436 — put, and forget gates, which determine whether Tumor stage 35.908 — a value should be remembered or forgotten. No. of positive nodes 54.943 — Because LSTM can remember a value for a lon- Node stage 59.483 — ger time period (ie, it is iterative), it is useful Pathologic stage 97.700 — for training using clinical data with various time Abbreviation: PNI, perineural invasion. series. However, LSTM exhibits a limited ability ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 7 SRN model was based on data from > 1,000 Table 3. Favorable Factors for Adjuvant XP Over Adjuvant Chemoradiation Therapy patients with GC and demonstrated that GC OR for XP Chemotherapy with a mesenchymal subtype located in the Variables Alone P cardia should elicit a more risk-adapted post- MSI subtype 0.630 .427 operative treatment strategy. Interestingly, we p53 active, MSS subtype 0.006 .938 found that GCs with MSI and GCs of the papil- p53 inactive, MSS subtype 5.533 .019 lary type have significantly more favorable sur - Mesenchymal subtype 3.429 .064 vival outcomes after XP chemotherapy alone. Sex 0.051 .821 These factors will be validated in the ARTIST-II Age 3.645 .056 trial currently recruiting patients to compare TS-1 alone versus TS-1 and oxaliplatin versus Well differentiated adenocarcinoma 0.436 .509 TS-1, oxaliplatin, and RT in patients with D2- Moderately differentiated adenocarcinoma 0.024 .876 resected GC. Poorly differentiated adenocarcinoma 0.388 .533 In this study, the available input data used Signet ring cell carcinoma 0.034 .854 for survival prediction were mainly pathologic Mucinous adenocarcinoma 0.075 .784 results, molecular subtypes, adjuvant treat- Papillary adenocarcinoma 5.901 .015 ments, and recurrence information. However, Adenosquamous carcinoma 0.700 .403 there are other factors affecting patient sur- Hepatoid adenocarcinoma 2.912 .088 vival that could not be evaluated in our study. Tubular adenocarcinoma 0.037 .848 These included accompanying comorbidities Intestinal subtype by Lauren classification 3.243 .072 or nutritional status; operative factors, such as Diffuse subtype by Lauren classification 0.178 .673 the amount of intraoperative blood loss, trans- Mixed subtype 0.553 .457 fusion, or postoperative complications; and Presence of PNI 0.009 .926 postoperative recovery factors, such as nutri- 10,11,15-17 tional status, weight loss, or anemia. Presence of lymphovascular invasion 0.057 .812 Therefore, although pathologic and molecular Tumor stage 1.409 .235 data are strong prognostic factors, they are not Node stage 0.081 .776 modifiable, but are merely predictive of survival Pathologic stage 0.758 .384 and act as reference factors for deciding adju- Abbreviations: MSI, microsatellite instability; MSS, microsatellite stability; PNI, perineural invasion; vant treatment modality. Notably, some patient XP, cisplatin. and operative factors can be modified and improved by doctor-patient deliberations. If our to estimate survival involving censored events. model can include these modifiable factors for In our SRN model, the survival distribution at each patient, more individualized treatments the discrete time points was determined by the and accurate survival prediction would be pos- logistic hazard function, in which death events sible. Another possible limitation of this study were also scored relative to censored data. In was that our model did not combine imaging every sequential time, the model was updated data, which would be an excellent and natu- by the time-specific Xt, which contains the first ral extension of the current deep learning SRN visit information, phase dimension, and survival approach. probability dimension. In conclusion, our SRN predicted survival at a On the basis of the use of this algorithm, strong high rate, reaching 92% at postoperative year 5. predictors for poor responders in terms of SRN-based clinical trials or risk-adapted adju- survival after adjuvant treatment were mesen- vant trials should be considered for patients with chymal subtype, the presence of perineural GC to investigate more individualized adjuvant invasion, advanced stage, location of the pri- treatments. mary tumor in the cardia, and a greater num- ber of positive lymph nodes. These findings supported previous studies that observed that mesenchymal subtype produces continuous DOI: https://doi.org/10.1200/CCI.17.00065 recurrence and, thus, earlier death, after a Published online on ascopubs.org/journal/cci on March 5-year surveillance program after surgery. Our 14, 2018. 8 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics AUTHOR CONTRIBUTIONS Jun Ho Lee No relationship to disclose Conception and design: Jeeyun Lee, Woo Kyoung Jeong, Won Ki Kang Tae Sung Sohn Provision of study material or patients: Min Gew Choi, Se No relationship to disclose Hoon Park, Jun Ho Lee, Tae Sung Sohn, Jae Moon Bae, Sung Kim, Sung Wook Seo Jae Moon Bae Collection and assembly of data: Jeeyun Lee, Min Gew Choi, No relationship to disclose Se Hoon Park, Seung Tae Kim, Jun Ho Lee, Jae Moon Bae, Sung Kim, Hyuk Lee, Byung-Hoon Min, Jae J. Kim, Sung Kim Dong-Il Choi, Kyoung-Mee Kim No relationship to disclose Data analysis and interpretation: Jeeyun Lee, Ji Yeong An, Tae Sung Sohn, Kyoung-Mee Kim, Mijung Kim, Sung Wook Hyuk Lee Seo No relationship to disclose Manuscript writing: All authors Byung-Hoon Min Final approval of manuscript: All authors No relationship to disclose Accountable for all aspects of the work: All authors Jae J. Kim AUTHORS' DISCLOSURES OF No relationship to disclose POTENTIAL CONFLICTS OF INTEREST Woo Kyoung Jeong The following represents disclosure information provided Consulting or Advisory Role: Samsung Medison by authors of this manuscript. All relationships are considered compensated. Relationships are self-held Dong-Il Choi unless noted. I = Immediate Family Member, Inst = My No relationship to disclose Institution. Relationships may not relate to the subject matter of this manuscript. For more information about Kyoung-Mee Kim ASCO's conflict of interest policy, please refer to No relationship to disclose www.asco.org/rwc or ascopubs.org/jco/site/ifc. Won Ki Kang Jeeyun Lee No relationship to disclose No relationship to disclose Mijung Kim Ji Yeong An No relationship to disclose No relationship to disclose Sung Wook Seo Min Gew Choi No relationship to disclose No relationship to disclose Se Hoon Park ACKNOWLEDGMENT No relationship to disclose We thank Kyunga Kim, PhD, and Hye Seung Kim in the Statistics and Data Center at Samsung Medical Center for Seung Tae Kim their statistical support. No relationship to disclose J.L. and J.Y.A. contributed equally to this work. Affiliations Jeeyun Lee, Ji Yeong An, Min Gew Choi, Se Hoon Park, Seung Tae Kim, Jun Ho Lee, Tae Sung Sohn, Jae Moon Bae, Sung Kim, Hyuk Lee, Byung-Hoon Min, Jae J. Kim, Woo Kyoung Jeong, Dong-Il Choi, Kyoung-Mee Kim, Won Ki Kang, and Sung Wook Seo, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea; and Mijung Kim, Ghent University, Ghent, Belgium. Support Supported by funding from the Korean Health Technology R&D Project, Ministry of Health & Welfare, Republic of Korea (Grants No. HI14C3418 and HI16C1990). REFERENCES 1. Torre LA, Bray F, Siegel RL, et al: Global cancer statistics, 2012. CA Cancer J Clin 65:87-108, 2. Macdonald JS, Smalley SR, Benedetti J, et al: Chemoradiotherapy after surgery compared with surgery alone for adenocarcinoma of the stomach or gastroesophageal junction. N Engl J Med 345:725-730, 2001 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 9 3. Lee J, Lim DH, Kim S, et al: Phase III trial comparing capecitabine plus cisplatin versus capecitabine plus cisplatin with concurrent capecitabine radiotherapy in completely resected gastric cancer with D2 lymph node dissection: The ARTIST trial. J Clin Oncol 30:268-273, 2012 4. Park SH, Sohn TS, Lee J, et al: Phase III trial to compare adjuvant chemotherapy with capecitabine and cisplatin versus concurrent chemoradiotherapy in gastric cancer: Final report of the adjuvant chemoradiotherapy in stomach tumors trial, including survival and subset analyses. J Clin Oncol 33:3130-3136, 2015 5. Bang YJ, Kim YW, Yang HK, et al: Adjuvant capecitabine and oxaliplatin for gastric cancer after D2 gastrectomy (CLASSIC): A phase 3 open-label, randomised controlled trial. Lancet 379:315- 321, 2012 6. Sakuramoto S, Sasako M, Yamaguchi T, et al: Adjuvant chemotherapy for gastric cancer with S-1, an oral fluoropyrimidine. N Engl J Med 357:1810-1820, 2007 7. Cancer Genome Atlas Research Network: Comprehensive molecular characterization of gastric adenocarcinoma. Nature 513:202-209, 2014 8. Cristescu R, Lee J, Nebozhyn M, et al: Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes. Nat Med 21:449-456, 2015 9. Lee J, Sohn I, Do IG, et al: Nanostring-based multigene assay to predict recurrence for gastric cancer patients after surgery. PLoS One 9:e90133, 2014 10. Kanda M, Mizuno A, Tanaka C, et al: Nutritional predictors for postoperative short-term and long- term outcomes of patients with gastric cancer. Medicine (Baltimore) 95:e3781, 2016 11. Lee JY, Kim HI, Kim YN, et al: Clinical significance of the Prognostic Nutritional Index for predicting short- and long-term surgical outcomes after gastrectomy: A retrospective analysis of 7781 gastric cancer patients. Medicine (Baltimore) 95:e3539, 2016 12. Woo Y, Son T, Song K, et al: A novel prediction model of prognosis after gastrectomy for gastric carcinoma: Development and validation using Asian databases. Ann Surg 264:114-120, 2016 13. Hochreiter S, Schmidhuber J: Long short-term memory. Neural Comput 9:1735-1780, 1997 14. Gers FA, Schmidhuber J, Cummins F: Learning to forget: Continual prediction with LSTM. Neural Comput 12:2451-2471, 2000 15. Mizuno A, Kanda M, Kobayashi D, et al: Adverse effects of intraoperative blood loss on long-term outcomes after curative gastrectomy of patients with stage II/III gastric cancer. Dig Surg 33:121- 128, 2016 16. Li L, Zhu D, Chen X, et al: Perioperative allogenenic blood transfusion is associated with worse clinical outcome for patients undergoing gastric carcinoma surgery: A meta-analysis. Medicine (Baltimore) 94:e1574, 2015 17. An JY, Kim KM, Kim YM, et al: Surgical complications in gastric cancer patients preoperatively treated with chemotherapy: Their risk factors and clinical relevance. Ann Surg Oncol 19:2452- 2458, 2012 10 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Appendix First Hospital Visit First year Second year Third year Fourth year Fifth year Personal RNN RNN RNN RNN RNN Features 0 1 1-p p 1-p p 1-p p 1-p p 1-p p Nonparametric Nonparametric Nonparametric Nonparametric Nonparametric Estimate Estimate Estimate Estimate Estimate (death, (death, (death, (death, (death, censored) censored) censored) censored) censored) Survival Recurrent Network [X ]+[S ][X ]+[S ] S RNN 1 0 t t–1 t wS Fig A1. Survival recurrent network (SRN) schema. (A) SRN simulated the sequential learning process of clinicians in the outpatient clinic using a recursive neural network (RNN) and time-sequential outcome data. The basic learning unit was composed of the RNN system and analyzes patient information at the first visit. At each time point, the unit takes this prediction and trains itself according to the actual survival data. If patients were censored during the time interval, the ranking scores estimated among the censored patients are used for training instead of survival data. The probability of survival is input and learned for the prediction of the survival probability for the following year. This sequential loop ends at the 5-year visit to yield the final survival probability. (B) Schematic SRN equation. The life value features consisted of time (t) and prior life expectancy (St), which is updated using the gradient descent equation (∂S). The model is retrained at the following sequential time with Xt+1 where St is embedded. ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 11 Test Data Embedding Fourth layer Seventh layer Output 120 120 100 100 100 80 80 80 60 60 0 10 20 30 40 50 40 40 40 10 20 30 40 50 20 20 20 10 20 30 40 50 01 20 40 60 80 00 120 0 53 10 15 20 25 30 5 08 1 2 3 4 5 67 0 0.5 1.0 1.5 2.0 MinimumMaximum Scale of each node Fig A2. The architec- ture of the basic survival recurrent network learning unit. The network archi- tecture of the basic unit comprised eight hidden layers with a single recursive neural network layer (long short-term memory). The input layer comprised 49 nodes that represented 47 input features and two survival features. The output layer comprised two nodes implementing the Softmax function, which represents live or dead probability. The individual features were embedded in a 49 × 32 vector. The other four layers comprised fully connected nodes implementing the rectified linear-unit function. The number of nodes was gradually reduced across the hidden layers. 12 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Alive Death Fourth layer Seventh layer Softmax 120 120 100 100 100 80 80 80 First-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 0 18 2 3 4 5 67 0 0.5 1.0 1.5 2.0 ∂S 120 120 100 100 100 80 80 80 Third-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 0 18 2 3 4 5 67 0 0.5 1.0 1.5 2.0 ∂S 120 120 120 100 100 100 80 80 80 Fifth-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 06 18 2 3 4 5 7 0 0.5 1.0 1.5 2.0 Fig A3. The architecture of the survival recurrent net- work (SRN) time-sequential learning unit. The SRN was trained every year sequen- tially with the 47 clinical features and the previous survival probability. Through this time-sequential training, the probability difference among patients became more distinct, and the accuracy improved. RNN, recursive neural network. ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 13 Sequential Training Scale of each node Minimum Maximum A B Personalized Survival Prediction XP XP+RT 0.9 LF+RT 0.9 Others 0.8 0.8 0.7 0.6 0.7 0.5 0.6 0.4 0.3 0.5 0.2 0.4 0.1 II III IV V VI 0.3 01 2 345 Subgroups Time (years) Fig A4. (A) Cumulative survival probability of a sample test group. The survival recurrent network (SRN) survival prediction curve was unique for each individual according to their unique clinical inputs. Two clusters were distinctly divided until the fifth year (blue lines indicate poor prognostic group; red lines indicate good prognostic group). (B) Five-year survival probability graph after simulation of four different adjuvant treatments by the trained SRN. The trained SRN was used to predict the survival probability associated with a virtual data set, in which four different adjuvant options were applied to all patients. According to the probability, the following distinct subgroups were identified: subgroup I, good prognostic subgroup regard- less of adjuvant therapy; subgroup II, good responder to three adjuvant options (cisplatin [XP], XP plus radiotherapy [RT], and leucovorin-fluorouracil [LF] plus RT); other options, including no treatment, abruptly decreased the survival probability < 50%; subgroup III, good responder to XP and XP plus RT, although LF plus RT should not be recommended because the response to LF plus RT was poor in this subgroup; subgroup IV, good responder only to XP, with XP without RT highly recommended in this subgroup; and subgroups V and VI, poor prognostic subgroups, where conventional adju- vant options will not be effective. In subgroups V and VI, trial regimens should be initiated from the beginning. 14 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Cumulative Survival Probability 5-Year Survival Probability http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png JCO Clinical Cancer Informatics Wolters Kluwer Health

Deep Learning-Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer

Loading next page...
 
/lp/wolters-kluwer-health/deep-learning-based-survival-analysis-identified-associations-between-ojBDi3P56m
Publisher
Wolters Kluwer Health
Copyright
(C) 2018 by Lippincott Williams & Wilkins, Inc.
ISSN
2473-4276
DOI
10.1200/CCI.17.00065
Publisher site
See Article on Publisher Site

Abstract

abstract original reports Deep Learning–Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer Purpose Gastric cancer (GC) is the third-leading cause of cancer-related deaths. Several pivotal clinical trials of adjuvant treatments were performed during the previous decade; however, the optimal regimen for adjuvant treatment of GC remains controversial. Jeeyun Lee Patients and Methods We developed a novel deep learning–based survival model (survival recur- rent network [SRN]) in patients with GC by including all available clinical and pathologic data Ji Yeong An and treatment regimens. This model uses time-sequential data only in the training step, and upon Min Gew Choi being trained, it receives the initial data from the first visit and then sequentially predicts the Se Hoon Park outcome at each time point until it reaches 5 years. In total, 1,190 patients from three cohorts Seung Tae Kim (the Asian Cancer Research Group cohort, n = 300; the fluorouracil, leucovorin, and radiotherapy cohort, n = 432; and the Adjuvant Chemoradiation Therapy in Stomach Cancer cohort, n = 458) Jun Ho Lee were included in the analysis. In addition, we added Asian Cancer Research Group molecular Tae Sung Sohn classifications into the prediction model. SRN simulated the sequential learning process of clini- Jae Moon Bae cians in the outpatient clinic using a recurrent neural network and time-sequential outcome data. Sung Kim Results The mean area under the receiver operating characteristics curve was 0.92 ± 0.049 at Hyuk Lee the fifth year. The SRN demonstrated that GC with a mesenchymal subtype should elicit a more Byung-Hoon Min risk-adapted postoperative treatment strategy as a result of its high recurrence rate. In addition, the SRN found that GCs with microsatellite instability and GCs of the papillary type exhibited sig- Jae J. Kim nificantly more favorable survival outcomes after capecitabine plus cisplatin chemotherapy alone. Woo Kyoung Jeong Conclusion Our SRN predicted survival at a high rate, reaching 92% at postoperative year 5. Our Dong-Il Choi findings suggest that SRN-based clinical trials or risk-adapted adjuvant trials could be considered Kyoung-Mee Kim for patients with GC to investigate more individualized adjuvant treatments after curative gastrec- Won Ki Kang tomy. Mijung Kim Clin Cancer Inform. © 2018 by American Society of Clinical Oncology Licensed under the Creative Commons Attribution 4.0 License Sung Wook Seo INTRODUCTION chemoradiotherapy (CRT) with fluorouracil (FU) Author affiliations and and leucovorin (LV). The Adjuvant Chemoradi- support information (if Gastric cancer (GC) is one of the most fre- applicable) appear at the ation Therapy in Stomach Cancer (ARTIST) trial quently occurring malignancies worldwide and end of this article. was a phase III trial that compared postoperative the third-leading cause of cancer-related deaths Licensed under the treatment with capecitabine plus cisplatin (XP) Creative Commons Attri- worldwide. Most patients with GC present with versus XP plus radiotherapy (RT) in patients bution 4.0 License metastatic disease at recurrence, and the over- 3,4 with extended D2 lymph node dissection. The all prognosis remains poor, with an expected Capecitabine and Oxaliplatin Adjuvant Study in Corresponding author: survival of < 1 year upon recurrence. Several Sung Wook Seo, MD, De- Stomach Cancer trial compared capecitabine pivotal clinical trials were performed in the pre- partment of Orthopaedic plus oxaliplatin treatment with observation in vious decade and aimed at reducing the recur- Surgery, Samsung Med- completely resected GCs and demonstrated an rence rate after curative surgery in patients with ical Center, Sungkyunk- wan University School additional survival benefit with adjuvant capecit- GC. First, the Intergroup 0116 (INT-0116) trial of Medicine, Irwonro80, 5 published in 2001 demonstrated significant abine plus oxaliplatin chemotherapy. The Adju- Gangnamgu, Seoul, Ko- vant Chemotherapy Trial of Titanium Silicate improvement in survival when patients with rea; e-mail: sungwseo@ completely resected GC received postoperative for GC trial compared titanium silicate (TS-1) skku.edu. © 2018 by American Society of Clinical Oncology ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 1 with observation in patients with D2-resected available for the three cohorts: pathology, type of GC and also showed prolonged survival in the surgery, lymphatic invasion, perineural invasion, TS-1 chemotherapy group. Hence, there are at histologic Lauren type, depth of invasion, num- least three to four postoperative chemotherapy ber of dissected lymph nodes, number of positive regimens available for patients with completely lymph nodes (pathologically), age at diagnosis, resected GC. sex, Epstein-Barr virus positivity, human epider- mal growth factor receptor 2 positivity, adjuvant To additionally complicate clinical decision mak- treatment modality, first sites of recurrence at the ing, two recent molecular landscape studies time of diagnosis for recurrence, date of recur- demonstrated the existence of different molec- rence, vital status at last follow-up, and molec- 7,8 ular GC subtypes. The Asian Cancer Research ular subtypes. Detailed information about these Group (ACRG) categories were determined on 8,9 cohorts has been published previously. Briefly, the basis of gene expression profiling of tumors 141 patients in the ACRG cohort received adju- with microsatellite instability (MSI), tumors with vant chemotherapy or CRT. The ARTIST trial was an epithelial-to-mesenchymal transition phe- a prospective adjuvant phase III trial that com- notype, tumors with a p53 signature (CDKN1A pared patients who received six cycles of XP and MDM2 expressing), or tumors without the (n = 228) with patients who received two cycles p53 signature. Notably, the ACRG study iden- of XP followed by RT with capecitabine and then tified four distinct molecular subtypes highly two more cycles of XP (n = 230). In the FU/LV/RT associated with GC recurrence rate and, thus, cohort, all patients (n = 432) received the INT- survival after surgery. To the best of our knowl- 0116 regimen (FU/LV followed by CRT with the edge, no model currently exists that is capable of same agents and then FU/LV again separately) integrating ACRG molecular subtypes and clin- as an adjuvant postoperative treatment. All of icopathologic information, such as stage, Lau- the patients in the FU/LV/RT and ARTIST cohorts ren classification, type of surgery, demographic received adjuvant treatment after D2 resection. data, and molecular subtypes, to predict survival The overall survival rates after 5 and 10 years after surgery. were 63.9% and 56.9%, respectively. Tumor recurrence was observed in approximately 40% In this study, we developed a deep learning– of the patients for each cohort. The protocol based prediction algorithm for survival predic- was approved by the Samsung Medical Center tion in patients with GC on the basis of data from Institutional Review Board (IRB; ACRG: IRB 1,190 patients with GC. The aims of this study No. 2010-12-088), ARTIST (ClinicalTrials.gov were to develop a deep learning–based predic- identifier: NCT0176146), and the DASL (cDNA- tion model to predict survival after surgery on mediated Annealing, S Selection, extension and the basis of sequential prediction of the outcome Ligation) cohort (IRB No. SMC 2010-10-025). at each time point until 5 years after operation and to predict the most optimal postoperative regimen after surgery using a recurrent neural Molecular Subtypes network (RNN) on the basis of available clinical variables. We used a previously published data set (accessed via https://www.ncbi.nlm.nih.gov/geo/ query/acc.cgi?acc=GSE62254). RNA was extracted PATIENTS AND METHODS from 300 tumors according to manufacturer Patients 8 protocol (Affymetrix, Santa Clara, CA), and we used the Affymetrix human genome U133 Plus We included the following three cohorts from 8 2.0 array (Affymetrix) for gene expression profil- our previous study: the ACRG cohort (n = 296), ing. The ACRG subtypes were the same as those the postoperative FU/LV/RT cohort (n = 432; published previously. Gene Expression Omnibus database identifier: 9 3 GSE26253), and the ARTIST cohort (n = 458). In total, 1,186 patients were included. We pro- Data Preprocessing cured all tissue specimens that were chemother- apy naïve during the primary resection of GC. Primary data were specified according to their No patients received neoadjuvant chemotherapy variable types. Categorical variables, such as or preoperative CRT. The following data were sex, tumor type and location, and molecular 2 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics expression, were converted using one hot encod- Basic Concept of Survival Recurrent Network Model ing, which transforms categorical features to a binary (0 or 1) format group. Ordinal variables A survival recurrent network (SRN) model was and quantitative variables were preprocessed constructed on the basis of logistic regression with normalization. All variables were then trans- function σ at discrete time point (t) and long formed by a standard scaler to standardize the 10,11 short-term memory (LSTM) neural network, intervals of values between the variables. To save which can be represented as follows: the numbers down to the third decimal place, all values were multiplied by 10 and then trans- Survival probability at time  t :    ( ) formed into integers (int32). As such, the value ft X = σ  Wt ∗ Xt =   ( ) ( ) − ω X t t of each variable was transformed into a standard 1 + e score, which was compatible with embedding. Missing data were imputed using the k-nearest neighbor algorithm after separating the training W ∗ X t t HRt(X ) = e and test sets. The event cases during the time interval were ranked by months, and the rank scores were inserted in the censored cases. θ X+ θ t 1 2 HRt(X ) = e Data Separation and Cross-Validation Patients were randomly sorted into a training set (80%) and a test set (20%). The test set was where ft X is the survival probability,   ( ) separated for the final test. Using the training HRt X is the hazard ratio function. ( ) set, bootstrap training (80% of the training set) and validation patients (20% of the training set) Once the parameter vector, Wt = θ1X + θ2t, is were generated by randomly selecting patients optimized with the patient group (features Xt and and repeating the selection 100 times to find target Yt) at the first time point (t), LSTM cells an optimal condition of the neural networks. memorize Wt and the model is retrained with the The validation error reached to the minimum at target value (Yt+1) of the next time point yielding epoch 7. The optimized model was tested over Wt+1. For example, if a patient (X) died of dis- the separated test patients. ease 2 years after the first visit, the model should learn the target value (Y1 = 1) at the first year and learn the target value (Y2 = 0) at the second Performance Evaluation and Statistics year. Because LSTM memorizes and optimizes W for each target value, our RNN-based model Receiver operating characteristics (ROC) curves, is able to infer a target value at a certain time areas under the receiver operating character- point. istic curve (AUCs), and concordance index (c-index) were compared using a nonparamet- Patient factors (X) were hardly updated at every time point because we could not collect all those ric Mann-Whitney U test using the MedCalc program (MedCalc Software, Seoul, Korea). All data without any loss. Moreover, the purpose of the survival model is to predict long-term survival neural networks were constructed using the with the information from the first visit. Thus, Keras with Theano backend in Python (https:// we generated Xt with the following assumption: keras.io/). The scikit-learn library (http://scikit- patients’ features should be constant during the learn.org/) was used for data management and observation time. Instead, there should be latent preprocessing. The Mann-Whitney U test was features that are dependent to the sequential performed to compare AUCs between models, time and indicate the patients’ status at a dis- whereas the Pearson χ test was performed to crete time point. We defined those latent fea- identify factors related to a specific subgroup. tures as time-dependent life value: The current study was developed and written according to Transparent Reporting of a Multi- variable Prediction Model for Individual Progno- Time‐dependent life value = θ t + sis or Diagnosis model development guidelines. ∂ S ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 3 Time‐dependent hazard function   represents live or dead probability. Forty-seven WtXt θ X+ θ t+ θ ∂S 1 2 3 e   = e features were preprocessed before accessing the input layer, with each feature preprocessed The life value dimensions were embedded in using a standard scaler and each value encoded the constant patient features (X) to generate as an integer within 10,000 scores. The clinical time-dependent features (Xt). The life value fea- variables of an individual were embedded in tures consisted of time (t) and prior life expec- a 47 × 32 vector for dimensionality reduction. tancy (St), which is updated using gradient The other two layers comprised fully connected descent equation ( ∂ S). nodes implementing the rectified linear-unit function. Gaussian dropout was performed to The model is retrained at the following sequen- prevent overfitting problems. The number of tial time with Xt+1 where St+1 is embedded. nodes was gradually reduced across the hidden layers (Appendix Fig A2, online only). The SRN St + 1 = St  +   ∂ S was trained every year sequentially with the 47 clinical features and the previous survival proba- bility. Through this time-sequential training, the ^f t ( x) ^ _____ probability differences among patients became St + 1 = St  +  α Yt −   Y t ∗ Y t ( ) | ft ( x) | more distinct, and the accuracy improved (Appen- dix Fig A3, online only). We randomly sorted 1,186 patients into a train- −wx dft(x) e ing group (80%; n = 950) and a test group (20%; ____ ________ = −  W* ≈  α ∗ ft(x ) (1 − ft ( x) )   −wx 2 dx 1  +   e ( ) n = 236). In the training group, bootstrap train- ing (80%; n = 760) and validation populations where α is a step size and  Y t  = ft(x). Thus, the (20%; n = 190) were generated by randomly function can be rewritten as follows: selecting patients and repeating the selection ^ ^ ^ 100 times. At each time point, the predicted sur- St + 1 = St  +  α(Yt −   Y t ) (1 − Y t ) Y t vival probability of the model was compared with the actual survival data. RESULTS The mean AUC of the 100 training group patients was 0.79 ± 0.052 at the first year, 0.839 ± 0.045 Predictive Accuracy of the SRN Model at the second year, 0.89 ± 0.049 at the third We simulated the sequential learning process year, 0.915 ± 0.05 at the fourth year, and 0.92 ± of clinicians in the outpatient clinic using RNN 0.049 at the fifth year (Fig 1A). The AUC of each and time-sequential outcome data (Appendix time point was compared using the Mann- Fig A1A, online only). The SRN was composed Whitney U test, showing that the AUC improved of the following three learning system parts: significantly in sequential years. The AUCs at the the first included information on patient status fifth and fourth years were significantly higher (covariates, X), the second included the time- than those at the first or second years according dependent life value ( θ t+ θ ∂ S), and the third 2 3 to the SRN model (Fig 1B). included nonparametric rank scores (R) of event that occurred during the interval (0 < R < 1). The R was inserted into the target value instead Performance of the Final Model for Predicting of binary values. The Xt were input into the SRN, Survival of Sample Test Group and the SRN was sequentially retrained with the Using the separated test group, the performance updated Xn (Appendix Fig A1B). The network of our final model was evaluated. The AUC of architecture of the SRN comprised five hidden each time point was 0.858 at the first year, 0.869 layers with two RNN layers (LSTM). The input at the second year, 0.879 at the third year, 0.912 layer comprised 49 nodes that represented 47 at the fourth year, and 0.923 at the fifth year input features and two life value features (phase (Fig 2A). [time] feature and prior survival probability fea- ture). The output layer comprised two nodes The c-index of the final model was evaluated implementing the Softmax function, which with the test group. The c-index was 0.951. 4 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Fig 1. The mean receiver A B operating characteristic (ROC) curve for 100 valida- Receiver Operating Characteristic 1.0 tion data sets. (A) The 100 1.0 trained survival recurrent networks (SRNs) were 0.8 0.9 tested iteratively using the corresponding test patients, 0.6 and ROC curves were 0.8 generated to evaluate the 0.4 average predictive power at each time point. (B) The 0.2 61 0.7 mean area under the ROC curve (AUC) of the test data sets. The AUC of each time 0 0.2 0.4 0.6 0.8 1.0 12345 point was compared using False-Positive Rate Time (years) the Mann-Whitney U test. The AUC of the fifth year was significantly higher than The calibration curve from the SRN was evalu- After a sequential learning process, two clus- that of the second year (P = .00) or the first year (P = ated using the sample test group (Fig 2B). In ters of patients were separated in terms of sur- .00). The AUC of the fourth this curve, the actual survival proportion was vival curves around the postoperative fifth year year was significantly higher compared with the SRN-predicted survival prob- (Appendix Fig A4A, online only). To identify sig- than that of the second year ability, with the actual survival rate calculated nificant factors differentiating the two prognostic (P = .00) or the first year (P = .00). using the Kaplan-Meier method. Our results clusters, we performed the Pearson χ test. Pos- showed that the actual survival rate was closely itive perineural invasion, the presence of lym- correlated with the predicted survival probability phovascular invasion, high tumor or node stage, within 15% margin of error. pathologic stage, a primary tumor located in the antrum or cardia, and recurrence were more The decision curves showed that the SRN model frequently observed to be statistically significant will be useful clinically for predicting the survival in the poor prognostic group in terms of survival of patients with GC. The SRN model has a pos- (Table 1). itive net benefit at all five different time points, rather than assuming all patients or none of the patients will survive at each year (Fig 3). SRN Guidance of the Optimal Adjuvant Treatment Regimen Fig 2. The receiver oper- ating characteristic (ROC) Cumulative Survival Probability of Individual To determine the optimal postoperative regimen curves and the calibration Patients in the Test Group plot of the survival recurrent according to the available data for each patient, network (SRN) for predict- we generated a new simulation data set, in The cumulative survival probability of each ing survival of a sample patient was visualized as a survival graph. which four different treatment options were test group. (A) The ROC and area under the ROC curve of the final model were evaluated using the A B separated test group at each 5-year point. (B) Calibration Receiver Operating Characteristic 1.0 curve of the SRN in the 1.2 test data set. The x-axis represents the probability 1.0 0.8 of 5-year survival, and the y-axis represents the actual 0.8 0.6 survival proportion. The thick solid line represents 0.6 1 year the calibration plot of the 0.4 2 years SRN. The actual survival 0.4 3 years proportion and 95% CIs 0.2 4 years were calculated using the 0.2 5 years Kaplan-Meier method. The prediction scores were 0.0 0.2 0.4 0.6 0.810 0 0.2 0.4 0.6 0.81 1.2 within a 15% margin of the perfect prediction line False-Positive Rate Predicted Survival Probability (dashed line). ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 5 True-Positive Rate True-Positive Rate Actual Survival Proportion ROC_AUC Fig 3. The decision curve A B analysis (dashed line) of the survival recurrent network 0.5 None 0.5 None (SRN) model for predicting All All SRN SRN the survival of sample test 0.4 0.4 group. The prediction model has a positive net benefit 0.3 0.3 if the line is above the lines assuming none of the 0.2 0.2 patients or all of the patients survive at each year. (A) 0.1 0.1 First year. (B) Second year. (C) Third year. (D) Fourth 0.0 0.0 year. (E) Fifth year. The decision curves of the model 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.0 0.1 0.2 0.3 0.40.5 show positive net benefit Threshold Probability Threshold Probability rather than assuming that all patients or none of the patients will survive at each CD year. 0.5 None 0.5 None All All SRN SRN 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0.0 0.0 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.60.8 Threshold Probability Threshold Probability 0.5 None All 0.4 SRN 0.3 0.2 0.1 0.0 0.0 0.2 0.4 0.6 0.8 Threshold Probability applied to all patients. Five subgroups were this population from additional analysis. For all adjuvant treatments, significant predictors identified according to the survival probability at for poor responders in terms of survival after the fifth year (Appendix Fig A4B). The treatment adjuvant treatment were mesenchymal subtype, options included in the analysis were as follows: the presence of perineural invasion, advanced XP chemotherapy; XP followed by capecitabine stage, location of the primary tumor in the cardia, plus RT followed by XP; the INT-0116 regimen; and others, such as oral TS-1 chemotherapy. and a greater number of positive lymph nodes Because oral chemotherapy or other regimens (Table 2). Notably, we found that patients with composed < 5% of the options, we excluded GCs with MSI (odds ratio, 4.2; P = .042) and 6 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Net Benefit Net Benefit Net Benefit Net Benefit Net Benefit one of the most popular methods for predicting Table 1. Poor Prognostic Factors for Survival cancer prognoses. A recent Cox model for sur- Variable Relative Risk for Death P vival prediction of patients with GC showed an PNI positive 44.947 — average c-statistic of 0.822. The Cox model Lymphovascular invasion 30.367 — uses the hazard rate of the covariates as model Tumor stage 12.865 — coefficients and, therefore, can encompass fac- Node stage 70.416 — tors that change over time. However, the model assumes that the hazard function is constant Pathologic stage 23.962 — through patient life span; therefore, it cannot Tumor located in cardia 16.567 — accurately predict the risk of death at a certain Recurrence 10.735 .001 time point because it does not represent the Abbreviation: PNI, perineural invasion. changing weight of covariates during the time intervals. A parametric model, such as a linear with GCs of the papillary type (OR, 5.9; P = .015) regression model, measures a direct relation- had significantly better survival outcomes after ship between the covariates and the survival XP chemotherapy alone (Table 3). Subgroup time. If the data from the survival distribution categorization according to SRN resulted in the at each time point are not large enough or if following five major groups (Fig A4B): subgroup there are large numbers of missing values, the I, good prognosis regardless of the adjuvant survival distribution at each time point becomes regimen; subgroup II, better prognosis after unreliable. This model often produces mislead- adjuvant chemotherapy, except for the others ing outcomes in the event of a violation of the group; subgroup III, better prognosis after adju- proportionality of the hazard assumption. Signifi- vant XP plus RT and XP alone; subgroup IV, cant risk factors that are not time dependent are better prognosis after XP alone; and subgroup V, often ignored in these models, thereby poten- poor prognosis regardless of the type of postop- tially resulting in false inferences. Another cause erative regimen. of misleading results is that the residual, the dif- ference between the real and expected values, DISCUSSION is not properly reflected in the model. Another recent model, the least absolute shrinkage and Studies have recently been performed to formal- selection operator, became a popular regres- ize risk prediction in cancer care. Kaplan-Meier sion method as a result of its enhanced predic- curves represent a nonparametric method for tive accuracy acquired by shrin king important estimating survival, with one of the strengths of covariates from thousands of variables, espe- this method being its ability to consider cen- cially when using genetic data. sored data, particularly right-censoring data using the log-rank test. Parametric survival Our model simulated the physician learning models and the Cox proportional hazards model process. Physicians learn about the prognoses might be useful to estimate covariate-adjusted of their patients through serial observations in survival. The Cox proportional hazards model is outpatient clinics. At the first visit, a clinician predicts the patient condition at the next visit on the basis of their current medical status Table 2. Predictive Factors for Poor Responders to Adjuvant Treatment and confirms that prediction during the sub- Variable Relative Risk for Death P sequent visit. We suggest that LSTM, an RNN proposed in 1997 by Sepp Hochreiter and Mesenchymal subtype 10.465 .001 10,11 Jürgen Schmidhuber, represents the optimal PNI positive 10.639 .001 choice for a serial learning system. LSTM was Signet ring 13.235 — designed to avoid the long-term dependency Location of cardia 14.516 — problem of RNNs and comprises input, out- Diffuse type by Lauren classification 16.436 — put, and forget gates, which determine whether Tumor stage 35.908 — a value should be remembered or forgotten. No. of positive nodes 54.943 — Because LSTM can remember a value for a lon- Node stage 59.483 — ger time period (ie, it is iterative), it is useful Pathologic stage 97.700 — for training using clinical data with various time Abbreviation: PNI, perineural invasion. series. However, LSTM exhibits a limited ability ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 7 SRN model was based on data from > 1,000 Table 3. Favorable Factors for Adjuvant XP Over Adjuvant Chemoradiation Therapy patients with GC and demonstrated that GC OR for XP Chemotherapy with a mesenchymal subtype located in the Variables Alone P cardia should elicit a more risk-adapted post- MSI subtype 0.630 .427 operative treatment strategy. Interestingly, we p53 active, MSS subtype 0.006 .938 found that GCs with MSI and GCs of the papil- p53 inactive, MSS subtype 5.533 .019 lary type have significantly more favorable sur - Mesenchymal subtype 3.429 .064 vival outcomes after XP chemotherapy alone. Sex 0.051 .821 These factors will be validated in the ARTIST-II Age 3.645 .056 trial currently recruiting patients to compare TS-1 alone versus TS-1 and oxaliplatin versus Well differentiated adenocarcinoma 0.436 .509 TS-1, oxaliplatin, and RT in patients with D2- Moderately differentiated adenocarcinoma 0.024 .876 resected GC. Poorly differentiated adenocarcinoma 0.388 .533 In this study, the available input data used Signet ring cell carcinoma 0.034 .854 for survival prediction were mainly pathologic Mucinous adenocarcinoma 0.075 .784 results, molecular subtypes, adjuvant treat- Papillary adenocarcinoma 5.901 .015 ments, and recurrence information. However, Adenosquamous carcinoma 0.700 .403 there are other factors affecting patient sur- Hepatoid adenocarcinoma 2.912 .088 vival that could not be evaluated in our study. Tubular adenocarcinoma 0.037 .848 These included accompanying comorbidities Intestinal subtype by Lauren classification 3.243 .072 or nutritional status; operative factors, such as Diffuse subtype by Lauren classification 0.178 .673 the amount of intraoperative blood loss, trans- Mixed subtype 0.553 .457 fusion, or postoperative complications; and Presence of PNI 0.009 .926 postoperative recovery factors, such as nutri- 10,11,15-17 tional status, weight loss, or anemia. Presence of lymphovascular invasion 0.057 .812 Therefore, although pathologic and molecular Tumor stage 1.409 .235 data are strong prognostic factors, they are not Node stage 0.081 .776 modifiable, but are merely predictive of survival Pathologic stage 0.758 .384 and act as reference factors for deciding adju- Abbreviations: MSI, microsatellite instability; MSS, microsatellite stability; PNI, perineural invasion; vant treatment modality. Notably, some patient XP, cisplatin. and operative factors can be modified and improved by doctor-patient deliberations. If our to estimate survival involving censored events. model can include these modifiable factors for In our SRN model, the survival distribution at each patient, more individualized treatments the discrete time points was determined by the and accurate survival prediction would be pos- logistic hazard function, in which death events sible. Another possible limitation of this study were also scored relative to censored data. In was that our model did not combine imaging every sequential time, the model was updated data, which would be an excellent and natu- by the time-specific Xt, which contains the first ral extension of the current deep learning SRN visit information, phase dimension, and survival approach. probability dimension. In conclusion, our SRN predicted survival at a On the basis of the use of this algorithm, strong high rate, reaching 92% at postoperative year 5. predictors for poor responders in terms of SRN-based clinical trials or risk-adapted adju- survival after adjuvant treatment were mesen- vant trials should be considered for patients with chymal subtype, the presence of perineural GC to investigate more individualized adjuvant invasion, advanced stage, location of the pri- treatments. mary tumor in the cardia, and a greater num- ber of positive lymph nodes. These findings supported previous studies that observed that mesenchymal subtype produces continuous DOI: https://doi.org/10.1200/CCI.17.00065 recurrence and, thus, earlier death, after a Published online on ascopubs.org/journal/cci on March 5-year surveillance program after surgery. Our 14, 2018. 8 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics AUTHOR CONTRIBUTIONS Jun Ho Lee No relationship to disclose Conception and design: Jeeyun Lee, Woo Kyoung Jeong, Won Ki Kang Tae Sung Sohn Provision of study material or patients: Min Gew Choi, Se No relationship to disclose Hoon Park, Jun Ho Lee, Tae Sung Sohn, Jae Moon Bae, Sung Kim, Sung Wook Seo Jae Moon Bae Collection and assembly of data: Jeeyun Lee, Min Gew Choi, No relationship to disclose Se Hoon Park, Seung Tae Kim, Jun Ho Lee, Jae Moon Bae, Sung Kim, Hyuk Lee, Byung-Hoon Min, Jae J. Kim, Sung Kim Dong-Il Choi, Kyoung-Mee Kim No relationship to disclose Data analysis and interpretation: Jeeyun Lee, Ji Yeong An, Tae Sung Sohn, Kyoung-Mee Kim, Mijung Kim, Sung Wook Hyuk Lee Seo No relationship to disclose Manuscript writing: All authors Byung-Hoon Min Final approval of manuscript: All authors No relationship to disclose Accountable for all aspects of the work: All authors Jae J. Kim AUTHORS' DISCLOSURES OF No relationship to disclose POTENTIAL CONFLICTS OF INTEREST Woo Kyoung Jeong The following represents disclosure information provided Consulting or Advisory Role: Samsung Medison by authors of this manuscript. All relationships are considered compensated. Relationships are self-held Dong-Il Choi unless noted. I = Immediate Family Member, Inst = My No relationship to disclose Institution. Relationships may not relate to the subject matter of this manuscript. For more information about Kyoung-Mee Kim ASCO's conflict of interest policy, please refer to No relationship to disclose www.asco.org/rwc or ascopubs.org/jco/site/ifc. Won Ki Kang Jeeyun Lee No relationship to disclose No relationship to disclose Mijung Kim Ji Yeong An No relationship to disclose No relationship to disclose Sung Wook Seo Min Gew Choi No relationship to disclose No relationship to disclose Se Hoon Park ACKNOWLEDGMENT No relationship to disclose We thank Kyunga Kim, PhD, and Hye Seung Kim in the Statistics and Data Center at Samsung Medical Center for Seung Tae Kim their statistical support. No relationship to disclose J.L. and J.Y.A. contributed equally to this work. Affiliations Jeeyun Lee, Ji Yeong An, Min Gew Choi, Se Hoon Park, Seung Tae Kim, Jun Ho Lee, Tae Sung Sohn, Jae Moon Bae, Sung Kim, Hyuk Lee, Byung-Hoon Min, Jae J. Kim, Woo Kyoung Jeong, Dong-Il Choi, Kyoung-Mee Kim, Won Ki Kang, and Sung Wook Seo, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea; and Mijung Kim, Ghent University, Ghent, Belgium. Support Supported by funding from the Korean Health Technology R&D Project, Ministry of Health & Welfare, Republic of Korea (Grants No. HI14C3418 and HI16C1990). REFERENCES 1. Torre LA, Bray F, Siegel RL, et al: Global cancer statistics, 2012. CA Cancer J Clin 65:87-108, 2. Macdonald JS, Smalley SR, Benedetti J, et al: Chemoradiotherapy after surgery compared with surgery alone for adenocarcinoma of the stomach or gastroesophageal junction. N Engl J Med 345:725-730, 2001 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 9 3. Lee J, Lim DH, Kim S, et al: Phase III trial comparing capecitabine plus cisplatin versus capecitabine plus cisplatin with concurrent capecitabine radiotherapy in completely resected gastric cancer with D2 lymph node dissection: The ARTIST trial. J Clin Oncol 30:268-273, 2012 4. Park SH, Sohn TS, Lee J, et al: Phase III trial to compare adjuvant chemotherapy with capecitabine and cisplatin versus concurrent chemoradiotherapy in gastric cancer: Final report of the adjuvant chemoradiotherapy in stomach tumors trial, including survival and subset analyses. J Clin Oncol 33:3130-3136, 2015 5. Bang YJ, Kim YW, Yang HK, et al: Adjuvant capecitabine and oxaliplatin for gastric cancer after D2 gastrectomy (CLASSIC): A phase 3 open-label, randomised controlled trial. Lancet 379:315- 321, 2012 6. Sakuramoto S, Sasako M, Yamaguchi T, et al: Adjuvant chemotherapy for gastric cancer with S-1, an oral fluoropyrimidine. N Engl J Med 357:1810-1820, 2007 7. Cancer Genome Atlas Research Network: Comprehensive molecular characterization of gastric adenocarcinoma. Nature 513:202-209, 2014 8. Cristescu R, Lee J, Nebozhyn M, et al: Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes. Nat Med 21:449-456, 2015 9. Lee J, Sohn I, Do IG, et al: Nanostring-based multigene assay to predict recurrence for gastric cancer patients after surgery. PLoS One 9:e90133, 2014 10. Kanda M, Mizuno A, Tanaka C, et al: Nutritional predictors for postoperative short-term and long- term outcomes of patients with gastric cancer. Medicine (Baltimore) 95:e3781, 2016 11. Lee JY, Kim HI, Kim YN, et al: Clinical significance of the Prognostic Nutritional Index for predicting short- and long-term surgical outcomes after gastrectomy: A retrospective analysis of 7781 gastric cancer patients. Medicine (Baltimore) 95:e3539, 2016 12. Woo Y, Son T, Song K, et al: A novel prediction model of prognosis after gastrectomy for gastric carcinoma: Development and validation using Asian databases. Ann Surg 264:114-120, 2016 13. Hochreiter S, Schmidhuber J: Long short-term memory. Neural Comput 9:1735-1780, 1997 14. Gers FA, Schmidhuber J, Cummins F: Learning to forget: Continual prediction with LSTM. Neural Comput 12:2451-2471, 2000 15. Mizuno A, Kanda M, Kobayashi D, et al: Adverse effects of intraoperative blood loss on long-term outcomes after curative gastrectomy of patients with stage II/III gastric cancer. Dig Surg 33:121- 128, 2016 16. Li L, Zhu D, Chen X, et al: Perioperative allogenenic blood transfusion is associated with worse clinical outcome for patients undergoing gastric carcinoma surgery: A meta-analysis. Medicine (Baltimore) 94:e1574, 2015 17. An JY, Kim KM, Kim YM, et al: Surgical complications in gastric cancer patients preoperatively treated with chemotherapy: Their risk factors and clinical relevance. Ann Surg Oncol 19:2452- 2458, 2012 10 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Appendix First Hospital Visit First year Second year Third year Fourth year Fifth year Personal RNN RNN RNN RNN RNN Features 0 1 1-p p 1-p p 1-p p 1-p p 1-p p Nonparametric Nonparametric Nonparametric Nonparametric Nonparametric Estimate Estimate Estimate Estimate Estimate (death, (death, (death, (death, (death, censored) censored) censored) censored) censored) Survival Recurrent Network [X ]+[S ][X ]+[S ] S RNN 1 0 t t–1 t wS Fig A1. Survival recurrent network (SRN) schema. (A) SRN simulated the sequential learning process of clinicians in the outpatient clinic using a recursive neural network (RNN) and time-sequential outcome data. The basic learning unit was composed of the RNN system and analyzes patient information at the first visit. At each time point, the unit takes this prediction and trains itself according to the actual survival data. If patients were censored during the time interval, the ranking scores estimated among the censored patients are used for training instead of survival data. The probability of survival is input and learned for the prediction of the survival probability for the following year. This sequential loop ends at the 5-year visit to yield the final survival probability. (B) Schematic SRN equation. The life value features consisted of time (t) and prior life expectancy (St), which is updated using the gradient descent equation (∂S). The model is retrained at the following sequential time with Xt+1 where St is embedded. ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 11 Test Data Embedding Fourth layer Seventh layer Output 120 120 100 100 100 80 80 80 60 60 0 10 20 30 40 50 40 40 40 10 20 30 40 50 20 20 20 10 20 30 40 50 01 20 40 60 80 00 120 0 53 10 15 20 25 30 5 08 1 2 3 4 5 67 0 0.5 1.0 1.5 2.0 MinimumMaximum Scale of each node Fig A2. The architec- ture of the basic survival recurrent network learning unit. The network archi- tecture of the basic unit comprised eight hidden layers with a single recursive neural network layer (long short-term memory). The input layer comprised 49 nodes that represented 47 input features and two survival features. The output layer comprised two nodes implementing the Softmax function, which represents live or dead probability. The individual features were embedded in a 49 × 32 vector. The other four layers comprised fully connected nodes implementing the rectified linear-unit function. The number of nodes was gradually reduced across the hidden layers. 12 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Alive Death Fourth layer Seventh layer Softmax 120 120 100 100 100 80 80 80 First-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 0 18 2 3 4 5 67 0 0.5 1.0 1.5 2.0 ∂S 120 120 100 100 100 80 80 80 Third-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 0 18 2 3 4 5 67 0 0.5 1.0 1.5 2.0 ∂S 120 120 120 100 100 100 80 80 80 Fifth-Year 60 60 60 RNN 40 40 40 20 20 20 03 53 10 15 20 25 0 5 06 18 2 3 4 5 7 0 0.5 1.0 1.5 2.0 Fig A3. The architecture of the survival recurrent net- work (SRN) time-sequential learning unit. The SRN was trained every year sequen- tially with the 47 clinical features and the previous survival probability. Through this time-sequential training, the probability difference among patients became more distinct, and the accuracy improved. RNN, recursive neural network. ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics 13 Sequential Training Scale of each node Minimum Maximum A B Personalized Survival Prediction XP XP+RT 0.9 LF+RT 0.9 Others 0.8 0.8 0.7 0.6 0.7 0.5 0.6 0.4 0.3 0.5 0.2 0.4 0.1 II III IV V VI 0.3 01 2 345 Subgroups Time (years) Fig A4. (A) Cumulative survival probability of a sample test group. The survival recurrent network (SRN) survival prediction curve was unique for each individual according to their unique clinical inputs. Two clusters were distinctly divided until the fifth year (blue lines indicate poor prognostic group; red lines indicate good prognostic group). (B) Five-year survival probability graph after simulation of four different adjuvant treatments by the trained SRN. The trained SRN was used to predict the survival probability associated with a virtual data set, in which four different adjuvant options were applied to all patients. According to the probability, the following distinct subgroups were identified: subgroup I, good prognostic subgroup regard- less of adjuvant therapy; subgroup II, good responder to three adjuvant options (cisplatin [XP], XP plus radiotherapy [RT], and leucovorin-fluorouracil [LF] plus RT); other options, including no treatment, abruptly decreased the survival probability < 50%; subgroup III, good responder to XP and XP plus RT, although LF plus RT should not be recommended because the response to LF plus RT was poor in this subgroup; subgroup IV, good responder only to XP, with XP without RT highly recommended in this subgroup; and subgroups V and VI, poor prognostic subgroups, where conventional adju- vant options will not be effective. In subgroups V and VI, trial regimens should be initiated from the beginning. 14 ascopubs.org/journal/cci JCO™ Clinical Cancer Informatics Cumulative Survival Probability 5-Year Survival Probability

Journal

JCO Clinical Cancer InformaticsWolters Kluwer Health

Published: Mar 14, 2018

References