2024 Latin America Conference on Lung Cancer (LALCA) - Posters-PP01.15: Machine Learning Models to Predict Operative Mortality in Thoracic Oncology Surgery for Non-Small Cell Lung Cancer

PP01.15: Machine Learning Models to Predict Operative Mortality in Thoracic Oncology Surgery for Non-Small Cell Lung Cancer

Back to course

Pdf Summary

This study addresses the need for better understanding of thoracic surgery outcomes in Latin America by utilizing machine learning models to predict operative mortality in patients undergoing thoracic oncologic surgery for non-small cell lung cancer (NSCLC). The research was conducted on patients who had undergone lung cancer resections at a hospital in Cali, Colombia, between 2009 and 2023. <br /><br />The primary objective was to predict operative mortality, defined as death during the initial hospitalization or within 30 days following surgery. Given the dataset's imbalance—175 negative cases and only 8 positive cases—the Adaptive Synthetic Sampling (ADASYN) technique was employed to generate synthetic examples for minority classes to address this issue. <br /><br />Two machine learning models were applied, and the Extra Trees model was highlighted for its performance, achieving an F1-Score of 66.67, Recall of 50, and an Area Under the Receiver Operating Characteristic (AUC ROC) of 0.79.<br /><br />The study concluded that data limitations, including small sample size and class imbalance, significantly impacted the model's performance, reducing its ability to capture robust statistical patterns. The researchers suggested that employing synthetic data generation techniques, such as generative AI, could enhance the model's stability and reliability. Furthermore, addressing these data constraints is crucial to improving the predictive accuracy and clinical applicability of machine learning models in assessing operative mortality risk in lung cancer surgeries.

Asset Subtitle

Nicolas Felipe Torres-España

Keywords

thoracic surgery

Latin America

machine learning

operative mortality

non-small cell lung cancer

NSCLC

ADASYN

Extra Trees model

synthetic data

predictive accuracy