Risk Assessment of Liver Metastasis in Pancreatic Cancer Patients Using Multiple Models Based on Machine Learning: A Large Population-Based Study

Qinggang Li, Lu Bai, Jiyuan Xing, Xiaorui Liu, Dan Liu, Xiaobo HuDisease 2022 · DOI

TL;DR

This study leverages machine learning to predict liver metastasis (LM) in pancreatic cancer (PC) patients, a critical factor influencing survival. By analyzing a large SEER database, they developed and compared several models, finding that a Random Forest (RF) model demonstrated the best predictive performance, offering a promising tool for personalized clinical decision-making.

What Problem Does This Paper Address?

Pancreatic cancer (PC) is a highly lethal disease with a poor 5-year survival rate. A significant challenge in managing PC is the high incidence of liver metastasis (LM), which often occurs at diagnosis or shortly after treatment. The presence of LM drastically reduces the effectiveness of surgical interventions and overall patient prognosis. Current diagnostic methods for LM, such as imaging techniques, can be limited by cost, accessibility, and clinician expertise, leading to potential inaccuracies. Therefore, there's a pressing need for more accurate and reliable methods to predict the risk of LM, enabling better patient stratification and tailored treatment strategies.

Methods

The researchers employed a retrospective study design, utilizing data from the Surveillance, Epidemiology, and End Results (SEER) database. They focused on developing predictive models for liver metastasis using several machine learning algorithms: Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Random Forest (RF), and Deep Neural Network (DNN).

Key methodological aspects include:

  • Feature Engineering: Continuous variables like age and tumor size were converted into categorical variables to enhance model applicability.
  • Feature Selection: Recursive feature elimination combined with cross-validation (CV) was used to identify the most important features for predicting LM.
  • Model Training and Validation: The dataset was split into training (80%) and testing (20%) sets. Models were trained using 10-fold CV on the training set and then evaluated on the independent test set.
  • Model Evaluation: Performance was assessed using metrics such as specificity, sensitivity, Gini index, and Receiver Operating Characteristic (ROC) curves. The DeLong test was used for comparing ROC curves.
  • Risk Stratification: The best-performing model (RF) was used to generate risk scores, categorizing patients into high, intermediate, and low-risk groups for LM.
  • Survival Analysis: Kaplan-Meier curves and log-rank tests were used to compare cancer-specific survival (CSS) among the risk groups. Propensity score matching (PSM) was employed to balance patient characteristics when evaluating the impact of treatments (surgery, chemotherapy, radiotherapy) on survival across different risk groups.

Datasets

  • SEER Database: The study utilized data from the SEER database, a comprehensive source of cancer statistics in the United States. The dataset included 47,919 patients diagnosed with pancreatic cancer between 2010 and 2018. This is a large, population-based dataset, providing a robust foundation for the study. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Machine Learning for Prognosis: The study demonstrates the power of machine learning algorithms in accurately predicting a complex clinical outcome like liver metastasis in pancreatic cancer, moving beyond traditional statistical methods.
  • Random Forest Dominance: Among the tested models, the Random Forest algorithm consistently outperformed others in predicting LM, highlighting its suitability for this type of predictive task.
  • Key Predictors Identified: Several clinical and demographic factors, including primary tumor site (body and tail being higher risk), surgery, chemotherapy, radiotherapy, and tumor stage, were identified as significant predictors of LM. This reinforces the importance of these factors in clinical assessment.
  • Risk Stratification for Tailored Care: The ability to stratify patients into distinct risk groups (high, intermediate, low) based on their predicted LM risk is a crucial takeaway. This stratification can guide more personalized monitoring and treatment decisions.
  • Treatment Impact Varies by Risk: The study suggests that the effectiveness of treatments like surgery, chemotherapy, and radiotherapy can differ significantly based on a patient's risk group for LM, emphasizing the need for individualized treatment planning.

Key Results

  • Prevalence of LM: Out of 47,919 PC patients, 15,909 (33.2%) developed liver metastasis.
  • Model Performance: The Random Forest (RF) model achieved the highest performance, with an ROC of 0.871 for the training set and 0.832 for the test set.
  • Risk Group Distribution: Patients were divided into three equal risk groups. The high-risk group showed a significantly higher proportion of liver metastasis (74.5%) compared to the middle-risk (24.4%) and low-risk (0.7%) groups.
  • Survival Outcomes: Cancer-specific survival (CSS) was significantly worse in the high-risk group (5-year CSS: 2.6%) compared to the middle-risk (4.8%) and low-risk (26.2%) groups.
  • Treatment Efficacy: In the high-risk group, surgery, chemotherapy, and radiotherapy were associated with improved CSS. In the middle-risk group, surgery and chemotherapy showed significant benefits. In the low-risk group, surgery improved survival, but chemotherapy and radiotherapy did not show a significant benefit.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Design: The study is retrospective, which inherently limits the ability to establish causality.
  • Database Limitations: The SEER database has intrinsic limitations, including non-unified selection criteria for patients and a lack of detailed treatment information (e.g., specific surgery details, chemotherapy regimens, radiation therapy plans).
  • Missing Variables: Important variables such as time-to-treatment, specific surgical details, patient status, and tumor burden at the surgical margin were not available.
  • External Validation Needed: The authors suggest that further validation of the developed models on large-scale external cohorts is necessary.

Future work could focus on incorporating more granular treatment data, exploring additional predictive features, and conducting prospective studies to validate these findings and further refine personalized treatment strategies for pancreatic cancer patients at risk of liver metastasis.

Summary updated

Abstract

In this study, the RF model constructed could accurately predict the risk of LM in PC patients, which has the potential to provide clinicians with more personalized clinical decision-making recommendations.
BibTeX
@inproceedings{li2022risk,
  author = {Qinggang Li and Lu Bai and Jiyuan Xing and Xiaorui Liu and Dan Liu and Xiaobo Hu},
  title = {{Risk Assessment of Liver Metastasis in Pancreatic Cancer Patients Using Multiple Models Based on Machine Learning: A Large Population-Based Study}},
  booktitle = {Disease},
  year = {2022},
  doi = {10.1155/2022/1586074},
}
Figure