← Lung Cancer

Machine learning-based prediction of recurrence-free survival after microwave ablation in patients with lung cancer: a retrospective study

IPBronch Review

🎯 Background & Rationale

  • Thermal ablation gap: Microwave ablation (MWA) is increasingly used for early-stage lung cancer, yet reliable tools for predicting recurrence-free survival (RFS) after MWA remain lacking.
  • Binary limitation: Prior predictive models treated local progression as a binary outcome rather than accounting for the specific timing of disease recurrence over time.
  • Machine learning potential: This retrospective study aimed to develop and evaluate multiple machine learning algorithms to generate individualized RFS predictions following MWA.

👥 Study Design & Population

  • Retrospective cohort: Single-center retrospective analysis of patients who underwent MWA at the Affiliated Hospital of Nantong University between January 2013 and January 2023.
  • Patient selection: 382 patients were initially screened, leaving 100 histopathologically confirmed primary lung malignancies included in the final analysis.
  • Cohort split: Patients were divided into a training set (n = 70) for model development and an internal test set (n = 30) for model validation.

📈 Methodology & Rigor

  • Multimodal data extraction: Utilized both structured preoperative/intraoperative data and unstructured electronic medical record notes processed via named entity recognition (LERT-BiLSTM-CRF architecture).
  • Candidate predictors: Evaluated 26 variables, selecting five candidate predictors via univariable Cox regression (P < 0.05) in the training set: sex, histologic subtype, CEA status, M stage, and overall TNM stage.
  • Model comparison: Developed and compared five machine learning models: Cox proportional hazards regression, eXtreme Gradient Boosting (XGBoost), random survival forest, gradient boosting machine, and support vector machine.

🔬 Key Findings [or Planned Endpoints]

  • Recurrence rate: A total of 46 out of 100 patients (46.0%) experienced disease recurrence during the follow-up period, with a median RFS of 26 months.
  • Best-performing model: XGBoost achieved the highest concordance index (C-index) in the test set at 0.856, outperforming traditional Cox regression.
  • Discriminative AUCs: The XGBoost model yielded test set areas under the curve (AUCs) of 0.933, 0.855, and 0.891 for predicting 1-, 2-, and 3-year RFS, respectively.
  • Key drivers: SHapley Additive exPlanations (SHAP) identified post-MWA carcinoembryonic antigen (CEA) status and TNM stage as the most influential predictors of recurrence.

⚖️ Critical Appraisal

  • Retrospective design: Single-center retrospective methodology inherently risks selection bias and limits broader clinical generalizability.
  • Validation limits: The modest sample size (n = 100 total, test set n = 30) provided internal validation only, requiring independent multicenter external validation before clinical adoption.
  • Unmeasured factors: Potentially crucial prognostic variables—such as tumor differentiation, EGFR mutation status, and EML4-ALK fusion status—were unavailable for model inclusion.

💡 The Clinical Bottom Line

  • Post-ablation risk: Integrating routine clinical variables like post-MWA CEA status and TNM stage into machine learning models can help stratify recurrence risk after lung cancer ablation.
  • Surveillance guidance: High-risk predicted scores can help clinicians tailor individualized surveillance schedules and identify patients who warrant closer clinical and imaging follow-up.

Abstract not available.
Read the full paper ↗

Keep reading

N opens the next one