Development and validation of a LightGBM based machine learning model for predicting postoperative erectile...

Area of Science:

• Urology
• Oncology
• Artificial Intelligence

Background:

• Postoperative erectile dysfunction is a significant complication after robot-assisted radical prostatectomy for prostate cancer.
• Traditional risk assessment models often fail to capture the complex factors influencing sexual recovery.

Purpose of the Study:

• To develop and validate an advanced machine learning model for predicting erectile dysfunction after robot-assisted radical prostatectomy.
• To integrate conventional surgical metrics with novel psychosocial and systemic immunological indices for improved prediction.

Main Methods:

• A multicenter retrospective cohort study of 824 patients undergoing robot-assisted radical prostatectomy.
• Dimensionality reduction using Least Absolute Shrinkage and Selection Operator regression and Random Forest.
• A LightGBM (Light Gradient Boosting Machine) framework with SHapley Additive exPlanations for prediction and interpretability.

Main Results:

• Seven key predictors were identified: Patient Health Questionnaire 9 score, age, HALP score, testosterone, clinical stage, Gleason score, and prostate volume.
• The LightGBM model achieved high predictive performance (AUC 0.946, accuracy 0.927).
• Patient Health Questionnaire 9 score was the dominant predictor; depressive symptoms increased risk, while good immune health and testosterone were protective.

Conclusions:

• A novel LightGBM prognostic framework incorporating psychological and physiological metrics can predict erectile dysfunction post-prostatectomy.
• This approach offers a more comprehensive risk assessment than traditional models.
• The framework may aid urologists in personalized risk stratification and rehabilitation planning.