The Risk Factors, Detection and Classification of Esophageal Cancer Using Ensemble Machine Learning Models

This study presents a robust ensemble machine learning framework utilizing a multi-seed strategy and Random Forest-based feature ranking to achieve near-perfect accuracy (98.3%) and zero false negatives in detecting esophageal cancer in Ethiopia, demonstrating that reduced feature sets focusing on dietary and environmental risk factors can effectively support early diagnosis in resource-limited settings.

The Gist

Imagine you are trying to find a single, specific needle in a massive, messy haystack. That needle is Esophageal Cancer (a dangerous cancer of the food pipe). Usually, doctors only find this needle when it's already huge and hard to remove, which is why so many people sadly pass away.

This paper is like a team of brilliant detectives who built a super-smart computer assistant to find that needle before it grows too big. They didn't just use one detective; they built a whole team of detectives (an "Ensemble") who work together to make sure they don't miss anything.

Here is the story of how they did it, broken down into simple parts:

1. The Clues (The Data)

The researchers gathered information from 312 people in Ethiopia (104 with cancer and 208 without). They looked at everything:

• What they ate: Did they eat too much hot porridge? Too much sugary candy?
• Their habits: Do they smoke? Chew khat? Drink coffee?
• Their environment: Do they live near factories?
• Their background: Age, job, and family history.

Think of this as gathering 52 different "clues" for every person.

2. The Problem: Too Many Clues!

Having 52 clues is a lot. Some clues are obvious (like "eating very hot food"), but others are just noise (like "what religion they follow" or "what color their house is"). If you try to solve a puzzle with too many useless pieces, you might get confused.

The Solution: The team used a special "Filter Machine" (called Random Forest) to sort the clues.

• Analogy: Imagine you have a bag of 52 marbles. Some are gold (important clues), and some are just pebbles (useless noise). The Filter Machine shook the bag and said, "Keep the gold marbles; throw away the pebbles."
• Result: They found that diet, hot food, and environmental exposure were the "gold marbles." They could throw away about 8 of the least important clues without losing any accuracy.

3. The Team of Detectives (The Ensemble Models)

Instead of trusting just one computer program to make the diagnosis, they built a team of five different AI detectives:

1. HGBC (The Speedster): A fast, efficient detective that groups clues together to make quick decisions.
2. XGBoost (The Strategist): A very smart detective that learns from its past mistakes to get better.
3. AdaBoost (The Coach): A detective that focuses extra hard on the cases it got wrong before.
4. Random Forest (The Committee): A group of trees (decisions) voting on the answer.
5. KNN (The Neighbor): A detective that looks at the people most similar to the patient to guess their condition.

The "Multi-Seed" Trick:
To make sure the team wasn't just getting lucky, they ran the experiment many times with different random starting points (like rolling dice to shuffle the deck). They averaged the results.

• Analogy: If you ask one person for directions, they might be wrong. If you ask 100 people and take the most common answer, you are almost certainly right. That's what this "Ensemble" did.

4. The Results: A Near-Perfect Score

When they tested their best team (the HGBC team), the results were incredible:

• Accuracy: They got it right 98.3% of the time.
• The Most Important Stat: They had ZERO missed cases.
  • Why this matters: In cancer detection, it is okay to have a few "false alarms" (telling a healthy person to get checked again), but it is disastrous to tell a sick person they are healthy. This team never missed a sick person. They caught every single cancer case in the test group.
• Speed: Even after throwing away the "pebble" clues, the team was just as good as when they had all the clues. This means the system is fast and doesn't need expensive equipment to work.

5. What Did They Learn?

The computer confirmed what doctors suspected but gave it hard numbers:

• Hot Food is Dangerous: Drinking scalding hot tea or eating very hot porridge is a major risk factor.
• Diet Matters: Eating lots of sugary foods, fatty foods, and preserved (salted) foods increases risk.
• Environment Counts: Exposure to certain chemicals and living conditions plays a big role.

The Bottom Line

This paper is like building a super-reliable metal detector for a beach.

• Old methods were like walking the beach with your eyes closed, hoping to step on the treasure.
• This new method is like a metal detector that ignores the seaweed and shells (the useless data) and only beeps loudly when it finds the gold (the cancer).

Why is this a big deal?
It offers a way to catch this deadly cancer early, especially in places where there aren't many expensive machines or specialist doctors. It's a cheap, fast, and incredibly accurate way to save lives by listening to the clues our bodies and habits give us.

Technical Summary

1. Problem Statement

Esophageal cancer (EC) is a highly lethal malignancy with a five-year survival rate below 15%, primarily due to late-stage diagnosis. While immunotherapy and targeted treatments are emerging, early detection remains the most critical factor for improving patient outcomes. The etiology of EC is multifactorial, involving complex interactions between socio-demographic, dietary, behavioral, environmental, and clinical risk factors. Traditional diagnostic methods often rely on invasive procedures (e.g., endoscopy) and may miss early-stage cases. There is a pressing need for robust, scalable, and interpretable predictive models that can utilize structured data to identify high-risk individuals, particularly in resource-constrained settings like the Arsi Zone of Ethiopia, where Esophageal Squamous Cell Carcinoma (ESCC) is prevalent.

2. Methodology

The study proposes a comprehensive machine learning framework designed to enhance model stability, interpretability, and predictive accuracy.

Data Acquisition and Pre-processing

• Dataset: The study utilized data from 312 individuals (104 confirmed EC cases and 208 controls) in the Arsi Zone, Ethiopia.
• Features: The initial dataset included 52 features spanning socio-demographics, diet (e.g., hot food consumption, sugary foods), behavior (e.g., smoking, khat chewing), environmental exposures, and clinical variables (e.g., tumor stage, dysphagia severity).
• Pre-processing: Unique identifiers were removed, and missing values were eliminated. The data was partitioned into training (80%) and validation (20%) sets.

Ensemble Feature Ranking

To address the instability of feature selection in small biomedical datasets, the authors developed an Ensemble Feature Ranking Framework:

• Multiple Random Forest (RF) models were trained using different random seeds and data splits.
• Feature importance scores were averaged across these models to generate a robust, stable ranking.
• This approach minimized the risk of identifying spurious biomarkers and ensured that selected features had consistent predictive power.

Multi-Seed Ensemble Classification

To reduce model variance and improve generalization, a Multi-Seed Ensemble Strategy was implemented:

• The dataset was repeatedly split using different random seeds ($S_1, S_2, \dots, S_n$).
• Five distinct base classifiers were trained on these splits:
  1. Histogram-based Gradient Boosting Classifier (HGBC) (Model I)
  2. Extreme Gradient Boosting (XGBoost) (Model II)
  3. AdaBoost (Model III)
  4. Random Forest (RF) (Model IV)
  5. k-Nearest Neighbors (KNN) (Model V)
• Final predictions were aggregated using majority voting to stabilize the output and mitigate the impact of outliers or specific data splits.

Experimental Design

Two experiments were conducted to evaluate the impact of dimensionality reduction:

• Experiment 1: Removed the bottom 8 lowest-ranked features.
• Experiment 2: Removed only the bottom 2 lowest-ranked features.
• Models were evaluated on both the full feature set and the reduced feature sets.

3. Key Contributions

1. Novel Multi-Seed Ensemble Framework: The paper introduces a robust strategy that aggregates predictions across multiple randomized training/testing splits, significantly reducing model variance and improving reliability in small-sample medical datasets.
2. Stable Feature Selection: By using an ensemble of Random Forests for feature ranking, the study ensures that identified biomarkers are statistically stable and biologically relevant, rather than artifacts of a single model run.
3. High-Performance Reduced-Feature Models: The study demonstrates that a reduced subset of features (focusing on diet and environment) can achieve performance equal to or better than models using the full feature set, offering a more efficient solution for clinical deployment.
4. Zero False-Negative Achievement: The proposed framework achieved perfect sensitivity (100%) in the evaluation cohort, a critical metric for cancer screening where missing a diagnosis is unacceptable.

4. Results

The study evaluated the models using Accuracy, Precision, Recall, F1-Score, and Area Under the Curve (AUC).

• Top Performer (Model I - HGBC):
  • Achieved the highest performance with 98.3% accuracy, 98.2% precision, 98.0% recall, and 98.1% F1-score using the reduced feature set (Experiment 2).
  • AUC: 0.994, indicating near-perfect discrimination between cancer and non-cancer cases.
  • Sensitivity: Achieved zero false negatives (100% sensitivity), resulting in a beta risk ($\beta$) of 0 and statistical power of 100%.
• Runner-up (Model II - XGBoost):
  • Demonstrated stable performance with accuracies of ~96.3% (full set) and ~96.1% (reduced set).
• Feature Reduction Impact:
  • Contrary to the typical expectation that reducing features lowers accuracy, the reduced feature sets often maintained or slightly improved performance. This suggests the removed features were redundant or noisy.
  • The reduced feature set retained the most critical predictors: dietary habits (hot food, sugary/fatty foods), environmental exposures, and behavioral factors.
• Comparison: Boosting-based ensembles (HGBC, XGBoost) consistently outperformed traditional classifiers (KNN, RF, AdaBoost) in terms of stability and accuracy.

5. Significance and Implications

• Clinical Decision Support: The model's ability to achieve 100% sensitivity with zero false negatives makes it an ideal tool for early screening, ensuring that no high-risk patients are missed. The trade-off of a single false positive is considered clinically acceptable given the severity of the disease.
• Resource-Constrained Settings: By proving that a reduced set of non-invasive, easily collectible features (diet, behavior, environment) is sufficient for high-accuracy prediction, the framework is highly suitable for healthcare systems with limited access to advanced imaging or genomic testing.
• Risk Factor Identification: The study reinforces the critical role of thermal injury (consuming very hot foods/beverages), dietary quality (low fruit/vegetable intake, high salt/sugar), and environmental exposures in the etiology of esophageal cancer in the studied region.
• Scalability: The ensemble approach provides a reproducible, interpretable, and scalable solution for precision medicine in oncology, bridging the gap between complex machine learning algorithms and practical clinical application.

In conclusion, this research validates that ensemble machine learning, specifically when combined with multi-seed strategies and rigorous feature selection, can provide a highly accurate, stable, and clinically viable tool for the early detection of esophageal cancer.