American College of Surgeons Survival Calculator for Biliary Tract Cancers; using Machine Learning to Individualize Predictions

Abstract

BACKGROUND

Although cancer prognosis is most commonly estimated by tumor stage, survival is multifactorial. Our objective was to develop an American College of Surgeons “Biliary Tract Cancer Survival Calculator” prototype using machine learning to generate personalized survival estimates based on patient, tumor, and treatment factors.

METHODS

The National Cancer Database was used to identify all patients with biliary tract malignancies between 2010–2017 including intrahepatic bile duct, extrahepatic bile duct, and gallbladder cancers. Included variables were determined based on random forest algorithms and review by subject-matter experts. Data were split into 80% training and 20% test datasets. Extreme gradient boosting with survival embeddings, a machine learning class, generated 3-year survival curves. Internal 5-fold cross validation was evaluated through concordance statistics (c-index), Brier scores, distant calibration, and time-dependent area under the curve (AUC).

RESULTS

Overall, 62,877 patients were included. Metastatic disease, age at diagnosis, and lack of surgical treatment were identified as most influential on worse survival outcomes via random forest. The final model included patient (age, sex, race/ethnicity, comorbidities), tumor (clinical TNM stage, disease site, grade), and treatment (surgery, chemotherapy, radiation) factors. Accurate model discrimination, calibration, and performance was demonstrated on internal validation (c-index: 0.74, Brier score: 0.14, distant calibration: p<0.001, AUC: 0.83). These metrics were notably improved compared to a model based solely on stage (c-index: 0.64, Brier score: 0.18, distant calibration: p<0.001, time-dependent AUC: 0.68).

CONCLUSIONS

This “Biliary Tract Cancer Survival Calculator” represents a highly accurate and comprehensive prognostic tool to estimate individualized survival estimates in real time.

Introduction

Biliary tract cancers, including intrahepatic bile duct, extrahepatic bile duct, and gallbladder malignancies, represent a diverse yet highly morbid disease group of which the associated 5-year survival rate across all stages and pathologies ranges from 2–69%.^1–3 However, noteworthy progress in the safety of available surgical techniques as well as implementation of multimodal treatments have led to improvements in survival outcomes for patients with biliary tract cancers despite frequently being described as having an aggressive tumor biology.^4–8 Thus, prognosis is increasingly viewed among cancer providers as being multifactorial, and dependent on additional factors such patient characteristics, tumor markers beyond those captured by stage alone, and treatment factors.

Despite recognition that survival is multifactorial, prognosis is most commonly estimated for patients with cancer based on the American Joint Committee on Cancer (AJCC) staging system alone.⁹ However, as AJCC stage often only accounts for the anatomic extent of disease, patients categorized within the same overall stage group have varying risk of survival outcomes. Although numerous cancer survival calculators have previously been developed across different primary cancer sites,^10–12 including patients with biliary tract malignancies,^13–15 many of these prognostic models have significant limitations and do not meet recommended criteria for prognostic model development.^16,17 Specifically, several existing survival calculators for patients with cancer (i) utilize single institution or registry databases which capture less than 50% of the nation’s cancer population, (ii) have limited patient, tumor, and treatment specific information, and (iii) do not leverage machine learning techniques which have been shown to improve risk prediction accuracy compared to previously developed prognostic nomograms.¹⁸

The National Cancer Database (NCDB) is one of the most comprehensive tumor registries worldwide, collecting more than 450 data fields among 74% of the nation’s cancer population each year, of which 75–85% of all biliary tract cancers are captured.^19,20 However, despite being an ideal database through which to develop a cancer survival calculator, no widely accepted prognostic nomogram which incorporates the multiple prognostic factors within the NCDB is available for use in clinical practice. Therefore, the objective of this study was to develop an American College of Surgeons “Biliary Tract Cancer Survival Calculator” prototype which uses the NCDB and machine learning to enable clinicians to tailor survival estimates more accurately for patients with intrahepatic bile duct, extrahepatic bile duct, or gallbladder cancers.

Methods

Data Source and Patient Population

This retrospective cohort study utilized the NCDB, identifying all patients who were diagnosed with biliary tract malignancies between January 1^st 2010 and December 31^st 2017 which included intrahepatic bile duct, extrahepatic bile duct, and gallbladder cancers. All patients within the final cohort had a minimum of 3 years follow-up. As previously described, the NCDB is one of the largest and most detailed cancer registries worldwide, capturing approximately 1.5 million newly diagnosed cancer cases from over 1,500 hospitals across the United States each year.¹⁹ Additionally, the NCDB is jointly maintained by the American College of Surgeon’s Commission on Cancer (CoC) and American Cancer Society as an obligatory component of CoC-accreditation. Details with regards to variables collected as well as data abstraction and quality maintenance processes implemented by the NCDB have previously been described.^20–23

All patients aged 18 or older with newly diagnosed primary biliary tract cancers were included in this study. Patients with duplicated records, with missing data on the original diagnosis and/or first course of treatment, and with missing survival data were excluded from analysis (Figure 1). Tumor site and histology were classified by using the International Classification of Disease for Oncology, 3rd Edition codes.²⁴ In addition, this study is exempt from institutional review board review as data within the NCDB is completely de-identified and Health Insurance Portability and Accountability Act (HIPAA) compliant.

Figure 1.

Consort diagram for patients diagnosed with biliary tract cancers from 2010–2017 within the National Cancer Database.

*Class of Case code 00 defined as follows: Diagnosis at the reporting facility and all treatment or a decision not to treat was done elsewhere.

Statistical Analysis

The primary outcome for this study was estimated 3-year overall survival curves, calculated from the date of diagnosis to the date of death or last follow-up. All statistical analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC) for variable selection as well as model development and validation, and Python version 3.12.0 for user interface development and personalized survival calculations.

Variable Selection

A data-driven approach was implemented in order to select variables included in the biliary tract cancer survival calculator model. First, a random forest algorithm was performed which is a supervised, machine learning class, that rank orders the most highly influential variables within the NCDB with regards to overall survival through a decision tree regression. Additional information in terms of the algorithm and properties of random forests have previously been described.²⁵ In addition to random forest algorithms, all variables included within the final prognostic model were reviewed by subject matter experts based on clinical validity, previously published literature, and variable availability within the NCDB. Of note, although individual patient comorbidities are not publicly available within the NCDB participant user file, this analysis was conducted by American College of Surgeon’s Cancer Programs staff through which ICD-10 codes for specific patient comorbidities are available and thus were included in the model.

Model Development

To develop the biliary tract cancer survival calculator, the cohort of patients identified within the NCDB was first randomly split into 80% training and 20% test datasets. The 80% training dataset was then used to generate our extreme gradient boosting with survival embeddings (XGBoostSE) machine learning model, incorporating the final variables selected via our data-driven approach. In summary, XGBoostSE is a machine learning class which recognizes patters between large sets of data inputs and outputs thereby allowing the model to learn through a series of multiple additive regression trees and ultimately generate autonomous predictions which has previously been described at length.²⁶ XGBoostSE was utilized specifically given its ability to generate predicted survival curves rather than single point estimates in time which provides clinicians with additional prognostic information throughout the course of the patient’s follow-up. Furthermore, XGBoostSE is able to automatically handle missing data as it is incorporated into the decision branch points of generated regresion trees, rather than excluding cases with missing data, which permits the inclusion of larger sample sizes, increasing the diversity of our cohorts through which the survival estimates are generated. Lastly, XGBoostSE models have been shown to improve both model accuracy and speed compared with traditionally used linear regression models which are important considerations when developing a cancer survival calculator to be used by clinicians in real time.^18,26

Model Validation

The 20% test dataset was used to internally validate our XGBoostSE machine learning model through concordance statistics (c-index), Brier scores, distant calibration, and time-dependent area under the curve (AUC). Specifically, the c-index characterizes the probability of concordance between predicted and observed survival of which values closer to 1.0 designate improved discrimination.²⁷ Brier scores are the squared mean of differences between the forecasted probability and outcome on a scale of 0 to 1 with scores closer to 0 demonstrating better predictive accuracy.²⁸ Distant calibration was assessed with the Hosmer-Lemeshow test with statistical significance set a p<0.05. Lastly, time-dependent AUC were assessed with values closer to 1.0 indicating better performance. A c-index, Brier score, distant calibration, and time-dependent AUC were then similarly calculated using AJCC staging criteria alone and compared to the calculated metrics of the XGBoostSE machine learning model.

Sensitivity Analysis

Recognizing that some providers may wish to counsel their patients regarding 5-year survival as well, our XGBoostSE model and sample cases were repeated using 5-year survival as the outcome.

Results

Patient Characteristics

In total, 62,877 patients with biliary tract malignancies were included of which 19,345 (30.8%) were intrahepatic bile duct, 21,246 (33.8%) were extrahepatic bile duct, and 22,286 (35.4%) were gallbladder cancers (Table 1). The median age of patients included in the cohort was 69 years old [IQR 60–77] and the majority of patients were Non-Hispanic White (n=44,012, 70.0%). For tumor specific factors, patients most often presented with Clinical T2 (n=16,354, 26.0%), node negative (n=31,194, 49.6%), non-metastatic disease (n=35,505, 56.5%). In terms of treatment, 26,078 (41.5%) patients underwent curative-intent surgical resection, 12,522 (19.9%) received systemic therapy, and 5,237 (8.3%) received radiation therapy.

Table 1.

Demographics of patients diagnosed with biliary tract cancers from 2010–2017 within the National Cancer Database.

	N (%) of patients diagnosed with biliary tract cancer (n=62,877)
Patient Specific Factors a
Age (years), median [IQR]	69 [60–77]
Sex
Male	28,576 (45.4%)
Female	34,301 (54.6%)
Race/Ethnicity
Non Hispanic White	44,012 (70.0%)
Non Hispanic Black	7,150 (11.4%)
Hispanic	2,789 (4.4%)
Asian, Pacific Islander	5,651 (9.0%)
Other	3,275 (5.2%)
Tumor Specific Factors
Clinical T-category b
T1	9,216 (1.8%)
T2	16,354 (26.0%)
T3	15,279 (24.3%)
T4	4,158 (6.6%)
TX	17,870 (28.4%)
Clinical N-category b
N0	31,194 (49.6%)
N1	16,534 (26.3%)
N2	2,140 (3.4%)
NX Unknown	5,824 (9.3%) 7,185 (11.4%)
Clinical M-category b
M0	35,505 (56.5%)
M1	19,407 (30.9%)
Unknown	7,965 (12.7%)
Disease Site
Intrahepatic Bile Duct	19,345 (30.8%)
Extrahepatic Bile Duct	21,246 (33.8%)
Gallbladder	22,286 (35.4%)
Tumor Grade
Well-differentiated	4,103 (6.5%)
Moderately-differentiated	14,531 (23.1%)
Poorly-differentiated	13,516 (21.5%)
Unknown	30,727 (48.9%)
Treatment Specific Factors
Surgery
Surgery	26,078 (41.5%)
No Surgery	36,799 (58.5%)
Systemic Therapy
No Systemic Therapy	50,355 (80.1%)
Systemic Therapy Administered	12,522 (19.9%)
Neoadjuvantc	1,458 (5.6%)
Adjuvantc	11,064 (42.4%)
Radiation Therapy
No Radiation Therapy	57,640 (91.7%)
Radiation Therapy Administered	5,237 (8.3%)
Neoadjuvantc	434 (1.7%)
Adjuvantc	4,803 (18.4%)

Abbreviations: IQR Interquartile range; NH Non-Hispanic

^aIndividual patient comorbidities not listed in Table 1 due to extensive number included.

^bClinical T-Category, N-Category, and M-Category determined by the American Joint Committee on Cancer 8^th Edition staging system.

^cNumber and percentages include only among those who underwent curative-intent resection. All else included within the total number and percentage of patients who received systemic and/or radiation therapy did not undergo surgery.

Variable Selection

Random forest algorithm found that patient specific factors (age at diagnosis, male sex), tumor specific factors (metastatic disease, clinical T4 disease, positive lymph nodes, intrahepatic disease site) and treatment specific factors (lack of surgical treatment, no systemic therapy, neoadjuvant systemic therapy, lack of radiation therapy) were all highly influential to account for when estimating overall survival (Figure 2). After clinical review among subject-matter experts of the random forest algorithm, the final model included patient specific factors: age at diagnosis, sex, race/ethnicity, selected patient comorbidities listed below; tumor specific factors: T-category, N-category, M-category, disease site, tumor grade; and treatment specific factors: surgery, systemic therapy, radiation therapy. Selected comorbidities were based on those incorporated into the Charlson-Deyo comorbidity score in addition to clinical face validity and included chronic obstructive pulmonary disease, dementia, myocardial infarction, heart failure, peripheral arterial disease, cerebrovascular accident, connective tissue diseases, peptic ulcer disease, liver disease, kidney failure, hypothyroidism, obesity, diabetes, and hypertension.

Figure 2.

Random forest algorithm demonstrating rank order of variables most highly influential on survival among patients diagnosed with biliary tract cancers from 2010–2017 within the National Cancer Database.

*Includes the top 10 variables identified on random forest algorithm.

**Positive Lymph Nodes is representative of the clinical stage.

Model Development and Validation

The XGBoostSE machine learning model was then constructed using the training dataset which incorporated the final list of variables selected after random forest algorithms and clinical review. Using the test dataset, internal 5-fold cross validation was then performed of which the results are presented in Table 2, demonstrating that this calculator is highly reproducible for use among individual patients. The average c-index of the constructed XGBoostSE model was 0.74 which suggests the model correctly ranks predicted survival probabilities among pairs of individuals 74% of the time. The average Brier score was 0.14 which means the squared error between predicted survival probabilities and observed outcomes was 14%. The distant calibration was p<0.001 suggesting that the overall calibration of predicted survival probabilities accurately reflects those of observed outcomes over time, and lastly, the time-dependent AUC was 0.83 which means the model is highly accurate in differentiating time-dependent outcomes.

Table 2.

Internal 5-fold cross validation metrics of the extreme gradient boosting with survival embeddings including the final model of patient, tumor, and treatment factors and AJCC stage only model.

Comprehensive Prognostic Model (includes patient, tumor, and treatment factors)	Validation Result
Average c-index	0.74
Average Brier score	0.14
Average distant calibration	<0.001
Time-dependent AUC	0.83
AJCC Stage Only Model
Average c-index	0.64
Average Brier score	0.18
Average distant calibration	<0.001
Time-dependent AUC	0.68

Abbreviations: AJCC American Joint Committee on Cancer; C-index concordance statistics; AUC area under the curve.

We then compared the predictive probabilities of the comprehensive XGBoostSE prognostic model including patient, tumor, and treatment specific factors compared to a similarly constructed model which included AJCC staging criteria alone (T-category, N-category, M-category). Of note, improved model performance was demonstrated across nearly all validation metrics after including additional prognostic factors when compared to the model built using only TNM stage (Table 2). Specifically, the c-index based solely on a model of AJCC stage was 0.64, the Brier score was 0.18, the distant calibration remained p<0.001, and the time-dependent AUC was only 0.68.

Sample cases are shown in Figure 3 which demonstrate the generated 3-year survival curves as a real-world application of the XGBoostSE model. For example, the estimated 3-year survival for a 75 year old Non-Hispanic Black male patient with a history of chronic obstructive pulmonary disease who presents with T1N0M0 intrahepatic cholangiocarcinoma with moderately differentiated disease and undergoes surgical resection followed by adjuvant chemotherapy is 56.6% [95% CI 49.3–63.5] (Figure 3a). In comparison, the estimated 3-year survival among all patients with T1N0M0 disease is 41.4% [95% CI 40.3–42.6] (Supplemental Figure 1). Additionally, a 50 year old Non-Hispanic White female with no medical comorbidities who presents with well differentiated T2N1M0 distal bile duct cholangiocarcinoma and undergoes surgical resection followed by adjuvant chemotherapy and radiation has an estimated 3-year survival of 41.8% [95% CI 34.9–48.9] using our current prediction tool (Figure 3b) as compared to 29.0% [95% CI 27.5–30.6] among all patients with T2N1M0 disease (Supplemental Figure 2).

Figure 3.

Sample cases demonstrating the generated output and survival estimates using the extreme gradient boosting with survival embeddings model: a) a 75 year old Non-Hispanic Black male with chronic obstructive pulmonary disease and T1N0M0 intrahepatic cholangiocarcinoma with moderately differentiated disease, who undergoes surgical resection followed by adjuvant chemotherapy, and b) a 50 year old Non-Hispanic White female with no medical comorbidities with T2N1M0 distal bile duct cholangiocarcinoma and well differentiated disease, who undergoes surgical resection followed by adjuvant chemotherapy and radiation.

*Solid line demonstrates represented survival outcomes up to 3 years with the band including 95% confidence intervals.

Sensitivity Analysis

After repeating the XGBoostSE model with 5-year survival as the outcome, the predictive abilities of the model remained largely unchanged as the c-index was 0.74, the Brier score was 0.13, the distant calibration was p<0.001, and the time-dependent AUC was 0.83. Furthermore, the estimated 5-year survival for the sample case of a 75 year old Non-Hispanic Black male patient with a history of chronic obstructive pulmonary disease who presents with T1N0M0 intrahepatic cholangiocarcinoma with moderately differentiated disease and undergoes surgical resection followed by adjuvant chemotherapy is 39.8% [95% CI 36.7–45.7] whereas the estimated 5-year survival among all patients with T1N0M0 disease is 36.6% [95% CI 32.9–38.0] (Supplemental Figure 3). Additionally, the estimated 5-year survival for the sample case of a 50 year old Non-Hispanic White female with no medical comorbidities who presents with well differentiated T2N1M0 distal bile duct cholangiocarcinoma and undergoes surgical resection followed by adjuvant chemotherapy and radiation is 24.3% [95% CI 19.9–28.9] as compared to 19.6% [95% CI 17.1–22.5] among all patients with T2N1M0 disease (Supplemental Figure 4).

Discussion

Estimating survival for patients with cancer has traditionally been based solely upon tumor stage. However, this approach fails to take into consideration the multifactorial prognostic characteristics which may impact survival such as those related to the patient, their tumor, and the planned course of treatment. Shared decision-making and increased patient education is critical to counseling patients with cancer, especially those diagnosed with primary tumors associated with a high mortality rate such as biliary tract malignancies. By including data from more than 62,877 patients with biliary tract cancers, we not only demonstrated that patient, tumor, and treatment specific factors are all important to include when estimating a patient’s survival, but also developed a highly accurate, prognostic nomogram based on these specific characteristics. Moving forward, this cancer survival calculator can be utilized by clinicians in real time to generate individualized survival estimates for patients with biliary tract malignancies.

In this study, we demonstrated that patient, tumor and treatment specific factors are important to account for when developing a prognostic model for patients with biliary tract cancers. For example, we found that the presence of metastatic disease, age at diagnosis, lack of surgical treatment, and lack of systemic therapy were ranked among the most influential variables in terms of overall survival on random forest algorithms and were associated with worse outcomes. By including these multiple prognostic factors in our biliary tract survival calculator, the accuracy of our nomogram improved by approximately 15% compared to a model based solely on TNM stage alone. It is important to note that prognostic nomograms for patients with biliary tract malignancies have been developed before; however, the vast majority have significant limitations. For example, a recent systematic review identified that all but one prognostic model for patients with intrahepatic cholangiocarcinoma demonstrated poor to moderate performance with a c-index of less than 0.7, and furthermore, were validated among cohorts of less than 500 patients.¹⁷ Additionally, most survival calculators for biliary tract cancers are based on less comprehensive databases, such as single or multi-institutional datasets, which only incorporate data from highly specialized institutions,^13,14,29,30 and thus, are likely not representative of the outcomes among patients with biliary tract cancers nationwide. Moreover, while other prognostic nomograms for biliary tract malignancies include limited non-anatomic tumor biomarkers, the majority fail to incorporate additional patient and treatment factors in their model, both of which were determined to be highly influential when estimating overall survival in the present study. Lastly, most previously developed calculators are limited to select cohorts of patients such as those who present with intrahepatic cholangiocarcinoma or gallbladder cancer.^14,31 Yet, this quickly becomes impractical in a real world application for clinicians who likely treat patients with all types of biliary tract cancers, and thus, needing to access multiple different user-interfaces in the midst of a busy clinical workflow. The cancer survival calculator in the current study addresses many of the limitations of other prognostic models for biliary tract cancers by using a national cancer-specific registry such as the NCDB to create a more comprehensive, generalizable, and personalized prognostic tool across all biliary tract malignancies with increased accuracy as well.

We also demonstrated the feasibility and importance in using machine learning when developing models to estimate survival for patients with biliary tract cancers. For example, although regression analyses are most commonly used in the creation of prognostic models, these often rely on the inclusion of complete cases only, and thus, not only reduces the statistical power of the model but has the potential to introduce selection bias as well.^17,32 Moreover, including additional prognostic factors beyond just the anatomic stage of disease results in hundreds, if not thousands, of unique combinations among all included prognostic data, and thus, often becomes impractical to apply in real world practices. An important benefit of incorporating machine learning methodologies when developing cancer survival calculators is in a machine learning models ability to identify complex, nonlinear patterns between large sets of data inputs and outputs which are often not readily apparent when using standard regression.²⁶ Using a machine learning model specifically designed for survival analyses, such as XGBoostSE, also provides additional prognostic data through the generation of a personalized survival curve, rather than single point estimates in time.

The implications of this work are substantial as the clinical relevance of personalized medicine, and specifically prognostication, in the field of oncology is increasingly recognized. For example, prior work has demonstrated that clinician-based survival predictions tend to be overly optimistic and highly inaccurate.^33–35 Furthermore, as previously mentioned, the majority of survival prediction tools currently available for patients with biliary tract cancers have demonstrated moderate performance at best in terms of their predictive accuracy.¹⁷ The cancer survival calculator in the present study thus presents an opportunity to strengthen the patient-provider relationship by increasing patient education, specific to their presentation, as well as improve shared decision-making given the additional prognostic information and increased accuracy of our model. Additionally, although most prior calculators are built using static datasets, this calculator will dynamically incorporate new data over time as it not only uses the NCDB but also was directly developed through the American College of Surgeons Cancer Programs and will incorporate new data on an annual basis in order to maintain contemporary survival predictions. Moving forward, a web-based user-interface which is currently under development will be made available for use of this “Biliary Tract Cancer Survival Calculator” by clinicians at CoC-accredited programs where clinicians would enter information specific to their individual patients on a website protected behind a login to be used in their own clinics when treating patients.

This study has limitations worth noting in addition to those inherent with the use of any large retrospective database. First, the NCDB only captures all-cause mortality, not cancer-specific mortality; therefore, while the objective of using this calculator in clinical practice is provide patients with a personalized survival estimate based on the diagnosis of biliary tract cancers, it is possible that patients included in the study cohort died of causes other than malignancy. However, as biliary tract cancers are associated with a high mortality, it is likely that most patients included in this prognostic nomogram died of causes related to their cancer. Additionally, individual patient co-morbidities are included in the calculator to help control for non-cancer related death to a certain degree. Second, only variables captured within the NCDB can be included within this cancer survival calculator, and thus, additional factors which may impact survival, such as circulating tumor cells or specific mutations, cannot accounted for within our prognostic nomogram. However, our cancer survival calculator includes more prognostic factors than nearly all other nomograms for patients diagnosed with biliary tract malignancies, and thus, we believe helps to address many of the limitations of other survival calculators given the robust data collected within the NCDB. Furthermore, we were unable to include CA19–9 in the current calculator due to a high degree of missing data for this variable, but plan to incorporate it moving forward as its capture within the NCDB improves as we recognize the importance of including this prognostic factor in the model. Third, given that the NCDB is a hospital-based registry which only captures data from patients diagnosed at CoC-accredited facilities, the survival estimates from this prognostic nomogram may not be representative of outcomes among all patient care settings, as patients treated at CoC-accredited facilities are more likely to receive guideline concordant care and subsequently experience improved survival.^36,37 Nevertheless, this limitation is mitigated by the fact that our future user-interface will only be made available to clinicians at CoC-accredited programs as a part of the value obtained through CoC-accreditation. Lastly, this cancer survival calculator will need to be externally validated moving forward as well which will represent an important future study as external validation is a critical component to ensuring the legitimacy of the risk predictions of any prognostic nomogram.¹⁶ While the inclusion of numerous patient, tumor, and treatment specific factors makes external validation challenging, we suspect that our internal validations will be consistent with the results generated through a future external validation study as the NCDB is one of the largest and most comprehensive tumor registries worldwide.

Conclusion

In this study of 62,877 patients with biliary tract cancers, we found that patient, tumor, and treatment specific factors were highly influential to account for when estimating overall survival. Additionally, incorporating a machine learning model designed specifically for survival outcomes allowed for the generation of increased prognostic data. Overall, this national cancer survival calculator for biliary tract malignancies represents a comprehensive prognostic tool through which a future web-based platform will be enable clinicians to input specific patient, tumor, and treatment factors to calculate personalized survival estimates for their patients in real time.