Prostate cancer prediction model: A retrospective analysis based on machine learning using the MIMIC-IV database
Wei Wang, Xin Jin
Abstract
Prostate cancer is a common malignant tumor in men and early diagnosis and treatment are crucial for the survival rate of patients. This study aimed to establish a machine learning-based prostate cancer prediction model to help physicians accurately identify high-risk patients. This study performed a retrospective analysis using prostate cancer patient data from the MIMIC-IV database. First, the data was cleaned and preprocessed, including imputing missing values, handling outliers, and feature selection. Then, prediction models were established using machine learning algorithms (including logistic regression, support vector machine, deep neural networks, XGBoost, LightGBM and CatBoost) and evaluated using cross-validation and ROC curve analysis. We screened out 1975 patients diagnosed with PC and 11,745 patients diagnosed with BPH based on ICD codes. However, among the BPH patients, 467 were also diagnosed with PC, so we excluded these patients. The LightGBM machine learning model outperformed the other models in distinguishing patients with PC [LightGBM vs. CatBoost vs. XGBoost vs. DNN vs. SVM vs. LR; area under the curve (AUC): 0.93 vs. 0.91 vs. 0.89 vs. 0.86 vs. 0.70 vs. 0.68, respectively]. The LightGBM model had a sensitivity of 86%, specificity of 85% at the best cut-off value. The model was capable of predicting whether a patient has prostate cancer based on their clinical features (including age, Laboratory test, etc.) and had a high level of accuracy and stability. The machine learning-based prostate cancer prediction model established in this study has some clinical application value and can help physicians accurately identify high-risk patients, providing more precise prevention and treatment plans for patients.