Objective To evaluate the efficiency of different machine learning algorithms in constructing prediction models for postoperative complications in thoracic surgery and compare their performance. Methods This study involved secondary analysis of data from two previous randomized controlled trials, collecting preoperative, intraoperative, and postoperative data from 868 thoracic surgery patients. The patients were divided into two groups based on the presence of postoperative complications: a complication group (n=255) and a non‑complication group (n=613). Primary endpoints were postoperative complications during hospitalization, including pulmonary complications, cardiovascular complications, acute kidney injury, postoperative stroke, delirium, and surgical complications. Machine learning algorithms such as K⁃Nearest Neighbors (KNN), Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XG⁃Boost), regularized regression (Glmnet), and Logistic Regression were used to establish prediction models for postoperative complications during hospitalization. The models were internally validated through 5⁃fold cross‑validation, and their area under the receiver operating characteristic curve (AUC), calibration, and accuracy were calculated to select the optimal model for feature importance analysis. Results Statistical differences were observed between the two groups in preoperative data, such as gender, age, albumin levels, smoking history, alcohol consumption, hypertension, American Society of Anesthesiologists (ASA) classification, forced expiratory volume in 1 second/forced vital capacity (FEV1/FVC) ratio, and maximal voluntary ventilation (MVV) (all P<0.05). There were also statistical differences in intraoperative data, such as prophylactic inhalation of penehyclidine hydrochloride, use of sevoflurane/dexmedetomidine, equivalent doses of intraoperative sufentanil, intraoperative sufentanil dosage, intraoperative blood loss, fluid balance, intraoperative transfusion, intraoperative hypotension/hypertension, duration of surgery, surgical approach, surgical type, postoperative intensive care unit (ICU) admission, and extubation at the end of surgery (all P<0.05). Postoperative data, including the Numerical Rating Scale (NRS) scores at rest on postoperative day 3 and the scores during movement on postoperative days 1−3, pathological diagnosis of malignancy, and length of hospitalization stay, were statistically different (all P<0.05). There were no statistical differences in other indicators (all P>0.05). Among the 868 patients, 255 (approximately 29.4%) experienced postoperative complications, with pulmonary complications (n=154, approximately 17.7%) being the most common, followed by cardiovascular complications (n=92, approximately 10.6%). The Glmnet algorithm showed the optimal performance, with an AUC of 0.773 [95% confidence interval (CI) 0.746, 0.779], an accuracy of 0.765 (95%CI 0.721, 0.778), and a calibration score of 0.165 (95%CI 0.158, 0.182). The top 10 variables most associated with an increased risk of postoperative complications were: longer surgery duration, lack of prophylactic inhalation of penehyclidine hydrochloride, male gender, higher NRS pain scores during movement on postoperative day 3, postoperative ICU admission, gastric/esophageal surgery, lower MVV, higher intraoperative sufentanil dosage, older age, and greater intraoperative blood loss. Conclusions Among the machine learning algorithms, the model constructed using Glmnet algorithm has the optimal performance for predicting postoperative complications in thoracic surgery. Constructing an optimal model and performing feature importance analysis is of significant value for clinical decision⁃making and perioperative management.