Cardiovascular Adverse Events After Definitive Chemoradiotherapy for Lung Cancer in an Appalachian Population: Incidence and Machine Learning Based Prediction

This retrospective study of an Appalachian lung cancer cohort treated with chemoradiotherapy reveals a high incidence (59%) of cardiovascular adverse events, primarily driven by age and cardiac radiation dose, which machine learning models identified as key predictors despite achieving only modest overall predictive accuracy.

The Gist

🏥 The Big Picture: A "Double Trouble" Situation

Imagine you are fighting a fire in a house (the lung cancer). To put out the fire, you need to use a powerful hose (the radiation and chemotherapy).

However, the house is built right next to a very important, fragile water main (the heart). In many cases, especially in the Appalachian region where this study took place, the water main is already old and rusty because the residents have smoked a lot and have other health issues.

The big question this study asked was: "When we blast the fire with our hose, how often do we accidentally damage the old, rusty water main? And can we use a smart computer to predict which houses are at the highest risk of a burst pipe?"

---

🔍 What Did They Do? (The Investigation)

The researchers looked back at the medical records of 86 patients in West Virginia who were treated for lung cancer between 2013 and 2025. These patients received "definitive chemoradiotherapy," which means they got a heavy dose of radiation and drugs to try to cure the cancer.

They wanted to find out two things:

1. The Reality Check: How many of these patients ended up with heart problems (like heart attacks, irregular heartbeats, or fluid around the heart) after their treatment?
2. The Crystal Ball: Could they build a Machine Learning (ML) model (a type of smart computer program) to look at a patient's data before treatment and say, "Hey, this person is at high risk for heart trouble"?

---

📉 What Did They Find? (The Results)

1. The Heart Trouble Was Common

The results were a bit scary. Out of the 86 patients, 59% (more than half) developed a heart problem during or after their treatment.

• The most common issues: Non-ST-elevation heart attacks (NSTEMI) and inflammation of the sac around the heart (pericarditis).
• Why so high? The Appalachian population often has a "perfect storm" of risk factors: high rates of smoking, existing heart disease, and the fact that the heart is right next to the lungs, making it hard to avoid hitting it with radiation.

2. The "Smart Computer" (Machine Learning)

The researchers tried to teach four different types of computer models to predict who would get heart trouble. They fed the computers data like:

• Patient's age.
• Smoking history.
• Existing heart disease.
• The most important data: Exactly how much radiation hit the heart (measured in "Gy" or "dose").

The Outcome:

• The Prediction: The computers were okay at guessing, but not perfect. Think of it like a weather forecast that says "50% chance of rain." It's better than a guess, but not a guarantee.
• The Best Model: A model called Gradient Boosting (GBM) was the best at spotting people who would get sick (high sensitivity), but it sometimes cried "wolf" too often (low specificity).
• The Key Predictors: The computer learned that Age and How much radiation hit the heart were the two biggest clues. If an older patient got a lot of radiation to their heart, the computer knew they were in the danger zone.

3. Predicting Death (Mortality)

The computers were actually slightly better at predicting who would pass away (mortality) than predicting specific heart events. Again, Age and Heart Radiation Dose were the top factors. This suggests that if your heart gets hit hard by radiation, it doesn't just cause a heart attack; it might shorten your life overall.

---

💡 The "Aha!" Moments (Key Takeaways)

🎯 The "Dosimetry" Analogy

Imagine the radiation treatment is like a spotlight. The doctors need to shine the light on the tumor (the bad guy). But the heart is standing right in the beam.

• The study found that how bright the light hits the heart (the dose) is a massive factor.
• Even with modern, precise technology (like IMRT), you can't always avoid the heart completely.
• The Lesson: Doctors need to be extra careful to dim the light on the heart as much as possible, even if it means the tumor gets a slightly different angle.

🤖 The "Simple vs. Complex" Analogy

The researchers tried using a super-complex computer model with 20 different variables (like a 20-ingredient soup). Then, they tried a simple model with just 4 ingredients (Age, Smoking, Existing Heart Disease, and Radiation Dose).

• Surprise: The simple soup tasted almost as good as the complex one!
• Meaning: You don't need a super-complex algorithm to know the basics. If you know a patient is old, smokes, has heart issues, and gets a high radiation dose, you already know they are at high risk.

---

🚧 Why This Matters (The "So What?")

1. Appalachia is Unique: This region has a lot of people with heart risks. What works for a healthy patient in a clinical trial might not work here. We need to treat these patients with extra caution.
2. Cardio-Oncology is Essential: We can't just treat the cancer and ignore the heart. These patients need a "heart team" (cardiologists) working alongside the cancer doctors (oncologists).
3. Better Planning: By using these simple computer tools, doctors can identify the "high-risk" patients before they start treatment. They can then adjust the radiation plan to spare the heart or monitor the heart more closely during treatment.

🏁 The Conclusion

In short: Lung cancer treatment in this specific population is tough on the heart. While our "smart computers" aren't perfect crystal balls yet, they confirmed that Age and Radiation Dose are the two biggest villains.

The study is a call to action: We need to be smarter about how we aim our radiation beams to protect the heart, especially for patients who are already walking a tightrope with their cardiovascular health.

Technical Summary

1. Problem Statement

Lung cancer patients treated with definitive chemoradiotherapy (CRT) face significant risks of cardiovascular adverse events (CVAEs), a concern exacerbated in the Appalachian region due to high baseline rates of smoking, pre-existing cardiovascular disease (CVD), and socioeconomic health disparities. While radiation-induced cardiac toxicity is well-documented, existing literature often relies on clinical trial populations with lower comorbidity burdens, potentially underestimating risks in real-world, high-risk cohorts. Furthermore, traditional regression-based risk stratification models may fail to capture the complex, non-linear interactions between clinical comorbidities, radiation dosimetry, and survival outcomes. There is a lack of data characterizing CVAE incidence and predictive modeling specifically for Appalachian lung cancer populations.

2. Methodology

Study Design and Population:

• Type: Retrospective single-institution cohort study (2013–2025).
• Cohort: 86 adult patients with biopsy-confirmed lung cancer (80% NSCLC, 20% LS-SCLC) treated with definitive CRT using 3D-CRT, IMRT, or VMAT.
• Follow-up: Median follow-up of 30.5 months.
• Exclusions: Palliative intent, non-pulmonary primaries, early-stage surgery/SBRT, or prior thoracic RT.

Data Collection & Variables:

• Predictors: Demographics (age, sex, race), smoking status, pre-existing CVD, tumor characteristics (stage, histology), treatment variables (chemotherapy, immunotherapy, RT dose/fractions), and cardiac dosimetry (Mean Heart Dose [MHD], Heart V20, V30, V40, V50).
• Outcomes:
  1. Primary: Occurrence of any CVAE (dichotomized: Yes/No), including NSTEMI, STEMI, arrhythmia, CHF, pericardial disease, and valvular dysfunction.
  2. Secondary: Patient mortality (Alive/Deceased).

Machine Learning (ML) Pipeline:

• Preprocessing: Missing values imputed (median for numerical, mode for categorical); z-score normalization applied. Synthetic Minority Over-sampling Technique (SMOTE) used to address class imbalance.
• Models Evaluated: Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), and Logistic Regression (LR).
• Validation: 5-fold stratified cross-validation with shuffling (seed=42).
• Evaluation Metrics: Area Under the Curve (AUC), Sensitivity, Specificity, F1-score, and Calibration curves. Uncertainty quantified via nonparametric bootstrapping (2000 resamples).
• Interpretability: Permutation feature importance analysis to identify key predictors.

3. Key Contributions

• First Real-World Appalachian Characterization: Provides the first incidence data on post-CRT CVAEs in an Appalachian cohort, highlighting a 59% event rate, significantly higher than many clinical trial reports.
• ML Risk Stratification: Systematically evaluates four ML algorithms for predicting both CVAEs and mortality in a high-risk, comorbid population.
• Dosimetric vs. Clinical Drivers: Demonstrates that cardiac dosimetric parameters (specifically heart V20, V40, V50, and MHD) and patient age are the dominant predictors, often outweighing traditional clinical variables in tree-based models.
• Reduced-Feature Viability: Shows that simplified models using only age, smoking status, pre-existing CVD, and cardiac dose metrics perform comparably to complex models, suggesting potential for streamlined clinical risk tools.

4. Results

Incidence of CVAEs:

• Overall Rate: 59% (51/86 patients) developed a CVAE.
• Most Common Events: NSTEMI (29.4%) and Pericardial Disease (29.4%), followed by Arrhythmia (15.7%).
• Dosimetric Trend: The CVAE group had a higher mean heart dose (13.4 Gy) compared to the non-CVAE group (9.4 Gy), though this difference was not statistically significant (p=0.27) likely due to sample size.

ML Performance for CVAE Prediction:

• Best Model: Gradient Boosting (GBM) achieved the highest AUC (0.567, 95% CI 0.44–0.69) with high sensitivity (73%) but low specificity (29%).
• Random Forest (RF): Showed the highest sensitivity (80%) but lower AUC (0.514).
• Other Models: LR (AUC 0.524) and SVM (AUC 0.492) showed modest to chance-level discrimination.
• Feature Importance: For GBM, Heart V20 was the most critical predictor, followed by Age. For RF, Heart V50 and Age were top predictors.

ML Performance for Mortality Prediction:

• Improved Discrimination: Models performed better for mortality than for CVAE prediction.
• Best Model: Random Forest achieved the highest AUC (0.63, 95% CI 0.496–0.750).
• Key Predictors: Age, Mean Heart Dose (MHD), Heart V50/V20/V30, Disease Stage (N/M), and Pre-existing CVD.
• Observation: The occurrence of a CVAE itself was a modest predictor of mortality, suggesting a link between cardiac toxicity and survival.

Reduced-Feature Analysis:

• Models trained on a reduced set of variables (Age, Smoking, CVD, Cardiac Dose) maintained performance comparable to full-feature models, indicating that additional baseline variables did not significantly enhance predictive power.

5. Significance and Clinical Implications

• High-Risk Population Validation: Confirms that Appalachian patients, characterized by high baseline CVD and smoking, experience a very high burden of cardiac toxicity (59%) following CRT, necessitating aggressive cardio-oncology surveillance.
• Dose Optimization: The consistent identification of heart V20, V40, V50, and MHD as top predictors reinforces the critical need for cardiac dose constraints (e.g., V40 ≤ 20%, MHD ≤ 20 Gy) in treatment planning, even with modern techniques like IMRT/VMAT.
• Risk Stratification Tools: While ML models showed only "modest" discrimination (AUC < 0.7), their high sensitivity (especially RF) makes them potentially useful for screening to identify high-risk patients who require intensified monitoring, rather than for definitive diagnosis.
• Multifactorial Nature: The study highlights that CVAE risk is driven by a complex interplay of radiation dose, patient age, and pre-existing comorbidities, suggesting that future models should integrate cardiac substructure dosimetry and biological biomarkers to improve accuracy.

Limitations:

• Small sample size (n=86) limiting statistical power and model complexity.
• Retrospective, single-institution design.
• Aggregation of diverse cardiac events into a single outcome variable.
• Lack of external validation.

Conclusion:
The study establishes a high incidence of CVAEs in Appalachian lung cancer patients and demonstrates that machine learning models, particularly tree-based ensembles, can effectively utilize age and cardiac dosimetry to stratify risk. These findings support the integration of cardiac dose optimization and ML-guided surveillance into standard care for high-risk populations.