Toward an Optimized Staging System for Pancreatic Ductal Adenocarcinoma: A Clinically Interpretable, Artificial Intelligence-Based Model

JCO Clin Cancer Inform 2021 AI 5 Explanations View Original
Original Paper (PDF)

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

Plain-English Explanations
Page [1, 2]
Why the Current Pancreatic Cancer Staging System Falls Short

The American Joint Committee on Cancer (AJCC) staging system is the standard tool oncologists use to describe how advanced a cancer is and predict prognosis. For pancreatic ductal adenocarcinoma, the AJCC 8th edition uses tumor size (T stage) and the number of positive lymph nodes (N stage) as independent factors — but research has suggested these factors interact in important ways that the current system ignores.

A long-standing controversy in pancreatic cancer staging concerns how to best represent lymph node involvement: should doctors count the total number of positive lymph nodes (PLN), as the AJCC currently does, or use the lymph node ratio (LNR — positive nodes divided by total nodes examined)? LNR accounts for how thoroughly the surgeon sampled the lymph nodes, potentially providing more accurate risk assessment.

This study used artificial intelligence techniques — specifically game theory mathematics and interpretable decision trees — to address both questions at once: which lymph node metric is better, and whether T and N stages truly interact in a way that improves staging accuracy.

TL;DR: The AJCC pancreatic cancer staging system may be improved by using lymph node ratio instead of raw positive lymph node counts and by accounting for interactions between tumor size and nodal status.
Pages 2-2
AI Meets Game Theory: A Novel Approach to Cancer Staging

The study included 4,105 patients who underwent pancreatic cancer resection between 2000 and 2017 at six major cancer centers. The derivation cohort (2,874 patients) was used to build the models, while an independent validation cohort (1,231 patients) tested how well the models generalized to new patients.

The team used Shapley Additive Explanations (SHAP) — a mathematical technique borrowed from game theory (originally used to fairly divide credit among players in a cooperative game) — to determine the relative importance of LNR versus PLN in predicting survival at one and three years. SHAP values reveal not just which variable matters more, but by exactly how much.

Using the best-performing lymph node metric, the team then trained Optimal Classification Trees (OCTs) — a type of decision tree that can be displayed visually and interpreted by clinicians without statistical expertise. OCTs were compared against the AJCC staging system and XGBoost (a powerful but less interpretable AI model) on both 1-year and 3-year survival prediction.

TL;DR: Using SHAP game-theory analysis and interpretable Optimal Classification Trees, researchers systematically compared lymph node metrics and tested whether tumor-nodal interactions improve pancreatic cancer staging.
Page [3, 4]
Lymph Node Ratio Outperforms Raw Positive Node Count

SHAP analysis identified lymph node ratio (LNR) as a superior predictor of survival compared to positive lymph node count (PLN). This finding supports a shift away from the current AJCC approach and aligns with growing evidence that accounting for how many lymph nodes were examined — not just how many were positive — provides a more accurate picture of cancer spread.

The Optimal Classification Trees trained using LNR outperformed the AJCC 8th edition staging system at both 1-year and 3-year survival prediction in the validation cohort. The OCTs slightly underperformed the more complex XGBoost models, but offered the critical advantage of being visually interpretable — clinicians can follow the tree's branches without needing to understand the underlying math.

Subgroup analyses in patients with fewer than 15 lymph nodes examined (where LNR is particularly important to account for sampling bias) showed that OCTs performed even better, reinforcing that LNR is especially valuable when surgical sampling is limited.

TL;DR: Lymph node ratio proved superior to raw positive node count, and AI-built interpretable decision trees outperformed the current AJCC staging system for predicting 1-year and 3-year survival after pancreatic cancer surgery.
Page [2, 4]
Tumor Size and Lymph Node Status Interact in Clinically Important Ways

A key finding of the SHAP interaction analysis is that T stage and N stage are not independent — they interact with each other in ways that the current AJCC system (which treats them as separate additive factors) fails to capture. Specifically, the survival benefit of having low T stage can be overridden by coexisting high N stage, and vice versa, to a greater degree than simple addition would predict.

This means a patient with a small tumor but many positive lymph nodes may have a worse prognosis than the AJCC system suggests, while a patient with a large tumor but very few positive nodes may do better. These interaction effects are clinically meaningful and help explain cases where AJCC staging seems inconsistent with actual outcomes.

Incorporating these interactions into the OCT models increases their discriminatory power without requiring additional clinical variables — an elegant demonstration that smarter use of existing data can improve staging.

TL;DR: Tumor size and lymph node status interact in ways the current staging system misses, and incorporating these interactions into AI decision trees improves survival prediction without needing additional data.
Page [2]
Building a More Accurate Staging System That Doctors Can Actually Use

The Optimal Classification Tree format is designed to be clinically usable: a doctor can look at a printed tree, answer simple binary questions about tumor size and lymph node ratio, and arrive at a risk category. This is in contrast to many AI models that produce a number from a black-box algorithm that clinicians cannot audit.

This is the first study to use AI and game theory to successfully modify an AJCC staging classification not just in pancreatic cancer but in any cancer type — a methodological milestone. The approach could be applied to other cancers where current staging systems are debated.

The authors note that patients who received neoadjuvant therapy were excluded because tumor size measurement after treatment is unreliable. As neoadjuvant therapy becomes standard for more pancreatic cancer patients, future staging systems will need to accommodate post-treatment specimens.

TL;DR: The interpretable AI-based staging model provides a clinically usable decision tree that doctors can follow without statistical expertise, representing a first-of-its-kind AI modification of an AJCC staging system.
Citation: Open Access, 2021. Available at: PMC9848537.