Predicting Symptom Trajectories After Thoracoscopic Lung Cancer Surgery Using an Interpretable Machine Learning Model

Patients suffer from a variety of symptoms after thoracoscopic surgery. However, there is a lack of validated predictive tools to identify potentially high-risk patients. This study is anticipated to include approximately 1,500 lung cancer patients who undergo thoracoscopic surgery. Latent class mixed modeling (LCMM) will be used to dentify subgroups of patients with similar symptom trajectories. Machine learning models were developed to predict postoperative symptom trajectories based on collected information. Effective prediction of postoperative symptoms can help identify high-risk patients and take preventive measures.

Thoracoscopic lung cancer surgery is a widely utilized approach for treating early and locally advanced lung cancer. Despite the advantages of thoracoscopic surgery, such as minimal invasion and rapid recovery, patients still suffer from a variety of symptoms such as pain, shortness of breath, sleep disorders or fatigue after surgery, which seriously affects the quality of life. However, there is a lack of validated predictive tools to identify potentially high-risk patients. This study is anticipated to include approximately 1,500 lung cancer patients who undergo thoracoscopic surgery. Patients are invited to fill out the MD Anderson Symptom Inventory-Lung Cancer module after thoracoscopic surgery. Symptoms of interest include pain, shortness of breath, sleep disturbance, and fatigue. Moderate to severe symptoms were defined as a score of ≥ 4. Latent class mixed modeling (LCMM), a clustering technique, can identify subgroups of patients with similar symptom trajectories based on longitudinal patient-reported outcome (PRO) data. Machine learning models were developed to predict postoperative symptom trajectories based on collected information including demographic and clinical information, and operative data. The machine learning models mainly include Random Forest, Support Vector Machines, Neural Networks, XGBoost, etc. The most appropriate model is selected, and model interpretation is performed using the SHAP method. Effective prediction of postoperative symptoms can help identify high-risk patients and take preventive measures.