A race against time: planning postoperative critical care

Prediction modeling of postoperative pulmonary complications following lung resection based on random forest algorithm

Postoperative pulmonary complications (PPCs) are a significant concern following lung resection due to prolonged hospital stays and increased morbidity and mortality among patients. This study aims to develop and validate a risk prediction model for PPCs after lung resection using the random forest (RF) algorithm to enhance early detection and intervention. Data from 180 patients who underwent lung resections at the Third Affiliated Hospital of the Naval Medical University between September 2022 and February 2024 were retrospectively analyzed. The patients were randomly allocated into a training set and a test set in an 8:2 ratio. An RF model was constructed using Python, with feature importance ranked based on the mean Gini index. The predictive performance of the model was evaluated through analyses of the receiver operating characteristic curve, calibration curve, and decision curve. Among the 180 patients included, 47 (26.1%) developed PPCs. The top 5 predictive factors identified by the RF model were blood loss, maximal length of resection, number of lymph nodes removed, forced expiratory volume in the first second as a percentage of predicted value, and age. The receiver operating characteristic curve and calibration curve analyses demonstrated favorable discrimination and calibration capabilities of the model, while decision curve analysis indicated its clinical applicability. The RF algorithm is effective in predicting PPCs following lung resection and holds promise for clinical application.

The LAS VEGAS risk score for prediction of postoperative pulmonary complications: An observational study

BACKGROUND

Currently used pre-operative prediction scores for postoperative pulmonary complications (PPCs) use patient data and expected surgery characteristics exclusively. However, intra-operative events are also associated with the development of PPCs.

OBJECTIVE

We aimed to develop a new prediction score for PPCs that uses both pre-operative and intra-operative data.

DESIGN

This is a secondary analysis of the LAS VEGAS study, a large international, multicentre, prospective study.

SETTINGS

A total of 146 hospitals across 29 countries.

PATIENTS

Adult patients requiring intra-operative ventilation during general anaesthesia for surgery.

INTERVENTIONS

The cohort was randomly divided into a development subsample to construct a predictive model, and a subsample for validation.

MAIN OUTCOME MEASURES

Prediction performance of developed models for PPCs.

RESULTS

Of the 6063 patients analysed, 10.9% developed at least one PPC. Regression modelling identified 13 independent risk factors for PPCs: six patient characteristics [higher age, higher American Society of Anesthesiology (ASA) physical score, pre-operative anaemia, pre-operative lower SpO2 and a history of active cancer or obstructive sleep apnoea], two procedure-related features (urgent or emergency surgery and surgery lasting ≥ 1 h), and five intra-operative events [use of an airway other than a supraglottic device, the use of intravenous anaesthetic agents along with volatile agents (balanced anaesthesia), intra-operative desaturation, higher levels of positive end-expiratory pressures > 3 cmH2O and use of vasopressors]. The area under the receiver operating characteristic curve of the LAS VEGAS risk score for prediction of PPCs was 0.78 [95% confidence interval (95% CI), 0.76 to 0.80] for the development subsample and 0.72 (95% CI, 0.69 to 0.76) for the validation subsample.

CONCLUSION

The LAS VEGAS risk score including 13 peri-operative characteristics has a moderate discriminative ability for prediction of PPCs. External validation is needed before use in clinical practice.

TRIAL REGISTRATION

The study was registered at Clinicaltrials.gov, number NCT01601223.