Explainable Machine Learning Predicts Survival of Retroperitoneal Liposarcoma: A Study Based on the SEER Database and External Validation in China

Cancer Medicine 2024 AI 8 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Retroperitoneal Liposarcoma Needs Better Prognostic Tools

Retroperitoneal liposarcoma (RLPS) is the dominant pathological subtype of retroperitoneal sarcoma (RPS), accounting for approximately 45% of all RPS cases. RPS itself is a rare malignancy with an incidence of roughly 0.5 to 1 per 100,000 people. RLPS typically presents as a large mass in the retroperitoneal space, discovered incidentally on imaging obtained for unrelated complaints, which means most patients are diagnosed when the tumor has already grown substantially.

Histopathological heterogeneity: RLPS is not a single disease. Pathologically, it is classified into five distinct subtypes: well-differentiated liposarcoma (WDLPS), dedifferentiated liposarcoma (DDLPS), myxoid liposarcoma, pleomorphic liposarcoma, and round cell liposarcoma. These subtypes differ markedly in their biology, recurrence patterns, and response to treatment, with DDLPS carrying a substantially worse prognosis than WDLPS. This heterogeneity complicates both clinical trial design and prognostic assessment when treating patients as a single group.

Limitations of prior prognostic approaches: Previous studies on RLPS prognosis have largely relied on conventional Cox proportional hazards regression, which is constrained by assumptions of linearity and proportional hazards. Many prior analyses were also limited by small sample sizes from single-center case series or reports, which reduces the reliability and generalizability of their conclusions. The SEER (Surveillance, Epidemiology, and End Results) database, maintained by the National Cancer Institute, provides access to a large population-based dataset spanning multiple decades and institutions, enabling more robust modeling.

This 2024 study, published in Cancer Medicine, addresses this gap by developing and validating six machine learning (ML) models to predict overall survival (OS) in RLPS patients using SEER data, with additional external validation from a Chinese tertiary center. A key contribution is the integration of explainability tools, SurvSHAP(t) and SurvLIME, to make the ML predictions interpretable at both the population and individual patient levels.

TL;DR: RLPS accounts for 45% of retroperitoneal sarcomas (incidence 0.5 to 1 per 100,000). Prior Cox regression studies were limited by small samples and linearity assumptions. This study builds six ML survival models from the SEER database (n = 1206), with external validation from China (n = 159), and adds explainability via SurvSHAP(t) and SurvLIME.
Pages 2-4
Study Design: SEER Cohort, External Validation, and Feature Selection

The investigators queried the SEER database for patients diagnosed with RLPS between 2004 and 2015, using ICD-O-3 codes 8850 to 8858 to identify cases. A series of exclusion criteria were applied to ensure data quality, including removing cases with incomplete survival information, unknown histological grade, or missing key covariates. The final SEER cohort included 1,206 patients, who were randomly split into a training set (n = 845, 70%) and an internal validation set (n = 361, 30%). The 7:3 split ratio is a standard approach in ML-based survival analysis studies. A chi-squared test confirmed no statistically significant differences between the two groups across all baseline variables (all p values above 0.05).

External validation cohort: An independent external validation cohort of 159 patients was obtained from The First Affiliated Hospital of Naval Medical University in Shanghai, China, covering cases treated between 2001 and 2019. External validation from a geographically and ethnically distinct population is critical for assessing model generalizability beyond the American patient population from which SEER data is drawn.

Feature selection with LASSO regression: A total of 14 clinical parameters were initially considered, including age at diagnosis, sex, race, marital status, tumor size, histological grade, SEER stage, AJCC stage, AJCC 7th TNM stage, histopathological classification, and treatment modality (surgery and radiation). LASSO (Least Absolute Shrinkage and Selection Operator) regression with 10-fold cross-validation was applied to reduce this feature set and prevent overfitting. The optimal regularization parameter lambda was determined to be 0.0237, the point at which partial likelihood deviance was minimized. LASSO retained 10 variables with nonzero coefficients, which were then entered into multivariate Cox proportional hazards analysis. Seven factors emerged as independently significant prognostic predictors and were used to build the final ML models.

X-tile software was used to determine optimal cutoff points for continuous variables such as age, converting them into categorical variables. This data-driven cutpoint optimization ensures that the thresholds used in the models are derived from the outcome data itself rather than arbitrarily assigned.

TL;DR: SEER cohort: 1,206 patients (2004 to 2015), split 7:3 into training and validation. External validation: 159 patients from Shanghai (2001 to 2019). LASSO with lambda = 0.0237 narrowed 14 variables to 10 nonzero coefficients; multivariate Cox analysis refined these to 7 independent prognostic factors used in final models.
Pages 4-6
Six Algorithms for Survival Prediction: Architecture and Rationale

The study developed six prognostic models using different machine learning and statistical frameworks, all evaluated against each other. The baseline comparator was the conventional Cox proportional hazards model (Coxph), which serves as a reference because it remains the standard tool for survival analysis in clinical oncology. The five ML models introduced alongside it each bring distinct computational properties suited to handling the complexity and nonlinearity of cancer survival data.

Random Survival Forest (RSF): RSF integrates random forest methodology with survival analysis, handling right-censored data through novel survival splitting rules. It estimates the overall mortality rate across an ensemble of survival trees and uses an event retention principle to generate predictions. RSF was implemented using the randomForestSRC R package. Ensemble methods like RSF generally outperform single-tree approaches by averaging across many trees, which reduces variance without substantially increasing bias.

Ranger: Ranger is a decision tree-based method using random forests or recursive partitioning, particularly well-suited to high-dimensional datasets due to its computational speed. It trains a large number of survival trees and generates predictions through a weighted voting system based on each tree's contribution. Ranger is implemented in R through the ranger package and is known for rapid computation without sacrificing predictive accuracy. In this study, ranger emerged as the best-performing model across all metrics.

GLMboost, Boosting Trees, and Decision Trees: Gradient boosting with component-wise linear models (GLMboost) uses a progressive gradient boosting approach to iteratively improve predictive capability while regularizing model complexity, with the generalized linear model as its base learner. Boosting trees, also an ensemble approach, combine predictions from multiple models to improve generalization. Standard decision trees, the simplest approach in the study, predict outcomes through sequential decision rules derived from the training data and served as the weakest performer overall. All analyses were conducted in R version 4.2.1 using packages including survex, parsnip, mlr3proba, ranger, randomForestSRC, and survival.

TL;DR: Six models were compared: Cox proportional hazards (baseline), Random Survival Forest, Ranger, GLMboost, Boosting Trees, and Decision Trees. Ranger, using weighted voting across many survival trees, outperformed all others. All models were implemented in R with standard survival analysis packages.
Pages 6-8
Independent Predictors of Overall Survival in RLPS

The final seven prognostic factors retained after LASSO regression and multivariate Cox analysis were: age at diagnosis (cutoff 65 years), sex, marital status, surgical history, histopathological classification, histological grade, and SEER stage. Each of these variables was independently associated with overall survival after adjusting for all others, and together they form the feature set for all six prognostic models.

Age and sex: Patients aged 65 years or older had significantly worse OS than younger patients. Age was identified as a relatively important variable across multiple explainability analyses. Male sex was also associated with inferior survival, consistent with prior reports in the liposarcoma literature documenting sex-based differences in outcomes, possibly related to differences in tumor biology, treatment patterns, or both.

Marital status: Unmarried patients (including single, divorced, and widowed) had worse OS compared to married patients. This finding aligns with a broader literature demonstrating the prognostic influence of marital status across multiple cancer types, attributed variously to differences in social support, treatment adherence, early diagnosis, and systemic stress responses. The authors note that this association has been reported in retroperitoneal liposarcoma specifically, as well as in bladder cancer and prostate cancer cohorts using SEER data.

Surgery and histological factors: Patients who underwent surgical resection had significantly better OS than those who did not, reinforcing the established primacy of surgery in RLPS management. Complete resection is the primary determinant of long-term survival in RLPS, and the absence of surgery carries a substantial prognostic penalty. Among histopathological subtypes, DDLPS was associated with the worst prognosis, consistent with its aggressive biology. Higher histological grade (grade III or IV, corresponding to poorly differentiated or undifferentiated tumors) and more advanced SEER stage (local or distant disease) were both independently associated with inferior survival.

TL;DR: Seven independent prognostic factors: age 65 or older, male sex, unmarried status, no surgery, DDLPS histopathological subtype, grade III/IV histology, and local or distant SEER stage. Surgery was strongly protective. DDLPS and high-grade tumors carried the worst histology-based prognosis.
Pages 8-10
Ranger Outperforms Across All Metrics and Validation Sets

Model performance was assessed using three complementary metrics: the concordance index (C-index), the integrated cumulative/dynamic area under the curve (C/D AUC), and the integrated Brier score. The C-index measures how well the model discriminates between patients with different survival times, with values above 0.7 considered practically useful. The integrated C/D AUC extends the AUC concept to time-to-event data across multiple time points. The integrated Brier score quantifies calibration, measuring the average squared difference between predicted survival probabilities and actual outcomes, with lower values indicating better calibration.

Training set performance: Ranger achieved the best training set results with a C-index of 0.742, integrated C/D AUC of 0.777, and integrated Brier score of 0.147. RSF performed second-best with C-index 0.732 and AUC 0.764. The conventional Coxph model had a C-index of 0.709 and AUC of 0.705. Decision trees showed the weakest discrimination with C-index 0.635 and AUC 0.640, suggesting that a single tree without ensemble averaging is insufficient for this clinical prediction task.

Validation and external validation: Ranger's advantage was most apparent in the internal validation set, where it reached a C-index of 0.775 and AUC of 0.849, with an integrated Brier score of 0.129. This represents a meaningful improvement over Coxph (C-index 0.717, AUC 0.753). In the external Chinese validation cohort, all models showed some performance decline reflecting the challenges of cross-population generalization, but ranger maintained relative superiority. Across all three datasets and all time points, every model maintained an integrated Brier score below 0.18, indicating satisfactory calibration throughout.

ROC curve analysis: Time-specific ROC analysis for 1-year, 3-year, and 5-year survival predictions confirmed that the ranger model consistently achieved AUC values above 0.83 across both training and validation sets. Cox-Snell residual plots were used to additionally evaluate model fit. The consistent performance of ranger across all evaluation frameworks established it as the model selected for further explainability analysis.

TL;DR: Ranger: C-index 0.742 (training), 0.775 (validation); AUC 0.777/0.849; Brier score 0.147/0.129. Coxph baseline: C-index 0.709/0.717; AUC 0.705/0.753. AUC for 1, 3, and 5 year OS predictions all exceeded 0.83. All models kept integrated Brier score below 0.18 across all cohorts.
Pages 10-12
Time-Dependent Feature Importance and Population-Level Interpretations

One of the central innovations of this study is its explicit use of explainable AI (XAI) techniques to move beyond black-box predictions. The authors conducted both global and local interpretability analyses of the ranger model, using tools from the survex R package. This dual-level explanation framework is particularly important in clinical oncology, where physicians need to understand not just what a model predicts, but why it makes that prediction for a specific patient.

Time-dependent variable importance: Standard variable importance metrics in survival models provide a single importance score per variable. This study used a time-dependent approach, computing variable importance at each follow-up time point using two separate loss functions: Brier score loss after permutation and C/D AUC loss after permutation. The importance of a variable is measured by how much the model's performance degrades when that variable's values are randomly shuffled (permuted), which breaks any relationship between the feature and the outcome. The two loss functions produced largely consistent rankings, with histological grade, SEER stage, and histopathological classification consistently among the most influential variables.

Partial Dependence Profiles (PDP): PDP plots illustrate how the predicted survival function changes across the full range of a single variable while all other variables are held constant at their population averages. If the bands across different values of a feature are thin and overlapping, the variable has limited marginal effect on predicted survival. Conversely, wide separation between bands indicates a strong effect. For variables like histological grade and surgical history, PDP analysis showed large separations between levels, visually confirming their clinical importance and the direction of their effect on survival probability over time.

SurvSHAP(t) global summary: SurvSHAP(t) extends the SHAP (SHapley Additive exPlanations) framework to survival data, computing time-dependent SHAP values that quantify each variable's contribution to the survival prediction at each time point. The global SurvSHAP summary ranked all seven features by their overall impact on OS and displayed their temporal variability using cumulative importance curves. Bar chart heights represent total significance, while curve trajectories reveal whether a variable's influence peaks early or persists throughout follow-up.

TL;DR: Global explainability used three tools: time-dependent variable importance by permutation (Brier score and C/D AUC loss), Partial Dependence Profiles, and SurvSHAP(t) summaries. Histological grade, SEER stage, and histopathological classification were consistently the most influential features. All tools confirmed that surgical history and tumor biology dominate OS prediction.
Pages 12-14
Patient-Level Predictions with SurvSHAP(t) and SurvLIME

Beyond population-level explanations, the study demonstrates how the ranger model can generate individualized survival explanations for single patients using two complementary local interpretability methods: SurvSHAP(t) and SurvLIME. This patient-level interpretability is clinically valuable because a physician's actual need is to understand the prognosis and key risk drivers for the specific patient sitting in front of them, not a statistical average across a population.

SurvSHAP(t) for individual patients: For any given patient, SurvSHAP(t) computes time-dependent SHAP values for each feature, indicating at each follow-up time point whether that feature is pushing the predicted survival probability higher (positive SHAP, plotted above zero) or lower (negative SHAP, plotted below zero). The study illustrates this with a case example: a divorced male patient over 65 years old with a history of surgery for RLPS, a poorly differentiated tumor, and round cell histopathological classification at regional SEER stage. The SurvSHAP(t) plot for this individual clearly shows which features dominated his predicted survival trajectory and how their contributions shifted across time.

SurvLIME for individual patients: SurvLIME (Survival Local Interpretable Model-agnostic Explanations) is closely related to SurvSHAP(t) but uses a different underlying approach: it generates a set of synthetic neighborhood observations around the index patient, obtains model predictions for each, and fits a simpler interpretable local model to approximate the black-box model's behavior in that patient's vicinity. The local importance scores from SurvLIME can differ from SurvSHAP(t) values for the same patient, providing a second perspective on which features are most influential. SurvLIME also generates a visualization comparing the black-box ranger model's predictions against the fitted local alternative model, showing how closely the simplified model tracks the complex one in the patient's local neighborhood.

This dual approach to local explainability is methodologically rigorous because no single explainability technique is universally optimal. Consistency between SurvSHAP(t) and SurvLIME results for the same patient increases confidence that the identified risk factors are genuinely influential rather than artifacts of a particular algorithm's assumptions.

TL;DR: SurvSHAP(t) plots time-dependent SHAP values per patient, showing which features push survival up or down at each follow-up point. SurvLIME builds a local interpretable model around each patient using synthetic neighborhood predictions. Using both tools together provides cross-validated local explanations, with consistency between methods increasing confidence in the identified risk drivers.
Pages 14-16
Study Constraints and the Path Toward Clinical Deployment

Retrospective design and selection bias: As a retrospective study, the analysis is inherently subject to selection bias. Patients included in the SEER database and in the Chinese hospital cohort are not random samples of all RLPS patients, and unmeasured confounders that influence both treatment assignment and survival outcomes cannot be controlled for. While the SEER database is large by sarcoma standards, RLPS remains a rare disease, meaning the absolute sample size (1,206 SEER patients) is still modest compared to datasets available for common cancers like breast or colorectal cancer.

Missing molecular data: A significant limitation of SEER-based studies is the absence of molecular diagnostic information. For RLPS specifically, clinically relevant markers such as PD-1 expression, Ki-67 proliferation index, MDM2 amplification status (relevant to WDLPS and DDLPS), and CDK4 amplification are not captured in SEER. These molecular features are increasingly recognized as important determinants of prognosis and treatment selection in liposarcoma, and their absence limits the completeness of the prognostic model.

Racial generalizability: The SEER database is predominantly composed of White patients (84% of this cohort), with Black and other racial groups underrepresented. Combined with the Chinese external validation cohort, the study captures some cross-ethnic diversity, but the performance of these models in racially diverse patient populations outside these two groups remains unknown. Racial differences in tumor biology, access to surgery, and follow-up intensity could all influence model accuracy in underrepresented groups.

Future directions: The authors highlight several priorities for advancing this work. First, prospective validation in larger multicenter cohorts would be needed before any clinical deployment. Second, integration of molecular biomarkers such as MDM2 and CDK4 status, PD-L1 expression, and mutational profiles into the feature set could meaningfully improve predictive accuracy. Third, building online tools or nomograms based on the ranger model would facilitate practical use by clinicians without requiring R programming expertise. The explainability framework demonstrated here, combining global PDP and SurvSHAP(t) with patient-level SurvLIME, could serve as a template for similar studies in other rare sarcoma subtypes where SEER data are available.

TL;DR: Key limitations: retrospective design, SEER lacks molecular data (MDM2, Ki-67, CDK4), limited racial diversity (84% White), and modest sample size for a rare cancer. Future work should integrate molecular biomarkers, build online clinical decision tools from the ranger model, and pursue prospective multicenter validation before clinical adoption.