Laboratory blood parameters and machine learning for the prognosis of esophageal squamous cell carcinoma

Feng Lu, Yang Lin-lan, Zhenglian Luo, Qiao He, Lijuan Shangguan, Mingfei Cao, Lichun Wu

TL;DR

Researchers have developed a machine learning model that significantly improves survival predictions for patients with esophageal squamous cell carcinoma (ESCC). By combining traditional clinical staging with routine laboratory blood parameters, the Random Survival Forest (RSF) model outperforms the standard AJCC 8th edition staging system, offering a more precise tool for identifying high-risk patients.

What Problem Does This Paper Address?

Esophageal squamous cell carcinoma (ESCC) is a highly aggressive malignancy with a sobering five-year survival rate of less than 20%. Currently, clinicians rely heavily on the AJCC TNM (Tumor, Node, Metastasis) staging system to predict patient outcomes and decide on treatment paths.

However, TNM staging doesn't tell the whole story. While factors like tumor size and lymph node involvement are critical, they ignore the systemic physiological state of the patient. Routine blood tests—which measure everything from hemoglobin to electrolyte levels—contain a wealth of "hidden" prognostic information. The challenge lies in the fact that these variables are complex and non-linear. This paper addresses the need for a more holistic, data-driven prognostic tool that integrates clinical staging with these readily available laboratory parameters.

Methods

The researchers compared three distinct modeling approaches to determine which could best handle the complexity of the data:

  1. Gradient Boosting Machine (GBM): An ensemble technique that builds models sequentially to correct errors from previous iterations.
  2. Random Survival Forest (RSF): An extension of the Random Forest algorithm specifically designed for "right-censored" survival data (where the event of interest, like death, may not have occurred for all subjects by the end of the study).
  3. Classical Cox Proportional Hazards: The traditional statistical "gold standard" for survival analysis.

To refine the model, they used minimal depth and permutation importance metrics to whittle down 27 initial variables to the 14 most impactful features. This ensures the model remains interpretable and focuses on the most relevant biological signals.

Datasets

The study utilized a substantial real-world dataset of 2,521 consecutive ESCC patients who underwent surgery at the Sichuan Cancer Hospital in China between 2009 and 2017.

  • Training Set: 1,765 patients (used for model development and hyperparameter tuning).
  • Test Set: 756 patients (used for independent validation).

The dataset included a mix of clinicopathological characteristics (like T/N stage and tumor grade) and laboratory indicators (like Mean Corpuscular Hemoglobin, Sodium, and Fibrinogen).

Key Insights and Ideas

  • The Power of "Boring" Data: One of the most interesting takeaways is that routine blood parameters—often viewed as secondary to imaging or pathology—carry significant weight. Features like Sodium (Na), Chlorine (Cl), and Fibrinogen (FIB) were among the top 14 predictors.
  • Non-Linearity Wins: The Random Survival Forest (RSF) slightly outperformed the classical Cox model. This suggests that the relationship between blood markers and cancer progression is non-linear and involves complex interactions that traditional statistics might miss.
  • Risk Stratification: The model doesn't just provide a "score"; it effectively divides patients into high- and low-risk groups. The survival gap between these groups was massive: in the training set, the 5-year survival rate for the low-risk group was 69.8%, compared to a staggering 11.5% for the high-risk group.

Key Results

  • Model Performance: The RSF model achieved a C-index of 0.746. In the world of survival models, a C-index (concordance index) measures how well the model ranks the timing of events; 0.5 is random, and 1.0 is perfect.
  • Predictive Accuracy: The Area Under the Curve (AUC) for predicting survival was 0.761 at three years and 0.771 at five years, showing consistent performance over time.
  • Superiority to Standard Staging: When compared directly to the AJCC 8th edition staging system, the ML model showed a significantly greater discriminative ability, meaning it was better at correctly identifying which patients were at higher risk of earlier mortality.

Limitations and Future Work

The authors acknowledge a few key hurdles:

  • Single-Center Bias: The data comes exclusively from the Sichuan Cancer Hospital. Because ESCC characteristics can vary by geography and ethnicity, the model needs to be validated on more diverse, international populations.
  • Retrospective Nature: The study looked at historical data. A prospective study—where the model is used in real-time to guide treatment—would be the next step to prove its clinical utility.
  • Dynamic Changes: The blood parameters were measured at a single point in time (pre-surgery). Future research could look at how changes in these levels over the course of treatment might provide even deeper prognostic insights.

Summary updated

Abstract

We developed an ESCC prognostic model with good performance by clinical features and laboratory blood parameters.
BibTeX
@inproceedings{lu2024laboratory,
  author = {Feng Lu and Yang Lin-lan and Zhenglian Luo and Qiao He and Lijuan Shangguan and Mingfei Cao and Lichun Wu},
  title = {{Laboratory blood parameters and machine learning for the prognosis of esophageal squamous cell carcinoma}},
  booktitle = {Frontiers in Oncology},
  year = {2024},
  doi = {10.3389/fonc.2024.1367008},
}
Figure