IntroductionProstatic carcinoma (PCA) is one of the most common cancers and is a major burden for men, with an estimated 0.4 million deaths worldwide in 20201. In China, PCA had the 6th highest incidence and 7th highest mortality among male cancers in 2020. PCA is characterized by an asymptomatic onset and slow progression, so most patients have late diagnoses and poor prognoses2. Screening is an important way to identify early cases of PCA.Currently, China’s PCA screening guidelines recommend the use of total prostate-specific antigen (TPSA) for primary screening3. However, a recent meta-analysis based on a Chinese population reported that benign diseases such as benign prostatic hyperplasia (BPH) were also related to elevated levels of TPSA. Therefore, this indicator has high sensitivity (pooled estimate: 91%) but low specificity (41%)4. A meta-analysis including 19 studies from different countries also reported that the TPSA had a high sensitivity (pooled estimate: 93%) but a low specificity (20%)5. This may lead to patient anxiety, increased costs, and potential harm related to unnecessary biopsies. A recent systematic review reported that most PCA models had good performance, with a pooled area under the receiver operating characteristic curve (AUC) of 0.78 (95% CI: 0.73, 0.82) (I2 = 96%)6. In routine care in China, patients who have enlarged prostate glands identified through digital rectal examination or ultrasound images with moderate or severe lower urinary tract symptoms are usually referred for prostate biopsy. However, few current studies have reported models specific for patients who often experience a more significant impact on their quality of life than for those with mild symptoms7. Precisely assessing PCA risk in this group is important for addressing immediate concerns and reducing unnecessary biopsies.Thymidine kinase 1 (TK1) is recognized as a cell proliferation biomarker that has potential value in cancer risk assessment8,9,10. A recent study showed that the serum TK1 protein (STK1p) concentration is associated with PCA11. A model involving STK1p may help to improve the accuracy of PCA risk assessment. Furthermore, the use of machine learning for the classification and prediction of disease outcomes has been shown to be highly efficient in the field of clinical research12. Therefore, this study aimed to develop a machine learning model using STK1p and other noninvasive predictors for classifying PCA and BPH to help precisely select high-risk patients for biopsy.MethodsDataset preparationThis study used a cross-sectional design. Male patients with suspected PCA who underwent prostate biopsies (Fig. 1a) at the First People’s Hospital of Longquanyi District, China, from June 1, 2022, to April 1, 2024, and at the Daping Hospital of the Third Military Medical University, China, from May 1, 2016, to September 1, 2017, were considered eligible. Patients with a history of cancer diagnosis prior to biopsy were excluded. Clinical methods were carried out in accordance with the Guidelines for the Diagnosis and Treatment of Prostate Cancer by the Chinese Society of Clinical Oncology. The requirement to obtain informed consent for inclusion in this analysis was waived by the Medical Ethics Committee of First People’s Hospital of Longquanyi District, Chengdu (No. AF-KY-2021002).Fig. 1(a) Flowchart of the study participants. Center 1 refers to the First People’s Hospital of Longquanyi District; Center 2 refers to the Daping Hospital of the Third Military Medical University. The Gleason scores of the prostatic carcinoma patients were between 6 and 9. (b) The procedure for developing machine learning models and a logistic model. Histology images of prostatic carcinoma (c) and benign prostatic hyperplasia (d) from the study participants.Full size imageNoninvasive parameters measured at inpatient departments within one day before biopsy were included as candidate predictors in this study. These parameters were chosen based on a thorough review of the existing literature, which suggested their potential relevance to the pathophysiology of PCA and could be measured using non-invasive, well-established, and validated methods. Parameters with a missing data rate of less than 40% were included. These included age, irregular margin (TRUS) parameters, and 13 serum biomarkers.The 13 serum biomarkers included STK1p, TPSA, free prostate-specific antigen (FPSA), free/total prostate-specific antigen ratio (FTPSA), alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), serum ferritin (SF), neuron-specific enolase (NSE), β-human chorionic gonadotropin (β-HCG), cancer antigen 125 (CA125), cancer antigen 15 − 3 (CA15-3), cancer antigen 72 − 4 (CA72-4), and cancer antigen 19 − 9 (CA19-9). STK1p was detected by a TK1 antibody-based enhanced chemiluminescence dot blot and analyzer, which was validated in previous research8. Other serum biomarkers were detected by a chemiluminescence dot blot and analyzer (Cobas 1601 auto analyzer of Roche, Switzerland).Two senior pathologists confirmed the biopsy results according to the Guidelines for the Diagnosis and Treatment of Prostate Cancer by the Chinese Society of Clinical Oncology. Patients with a positive biopsy result (PCA) were referred for prostate surgery and further pathological tests. Examples of histology images of PCA and BPH from the study participants are presented in Fig. 1c and d, respectively.Model development and validationThe Lasso logistic algorithm was also used to select significant predictors with 10-fold cross-validation for overfitting reduction (Fig. 1b). The predictors with the minimum cross-validation mean variance selected by the Lasso procedure were used to develop machine learning models. Five machine learning methods, namely, extreme gradient boosting (XGBOOST), decision tree learning, lasso, neural network (NNET), and support vector machine (SVM), were used to model PCA. The XGBoost model was optimized using a grid search approach with the following hyperparameters: learning rate (0.1), maximum depth of trees (5), and subsample ratio (0.8). The mean Shapley additive explanations (SHAP) approach was used to measure the importance of each candidate predictor.Forward stepwise selection was used to screen candidate predictors and fit the logistic model. The odds ratio (OR), 95% confidence interval (CI), and logistic regression coefficients were calculated for each selected predictor.The dataset of 310 patients was randomly split into training (70%) and test (30%) sets to evaluate the model performance (Table 1). To ensure the models’ stability and generalizability, 10-fold cross-validation was employed during the training phase. This process involved partitioning the training data into 10 equal parts, with each part serving as the validation set once while the remaining nine parts were used for training. The AUC of each model was calculated to assess its performance. The cutoff point for the logistic and lasso models was identified based on the highest specificity when the sensitivity was ≥ 91% for comparison with the TPSA4. The net benefits of machine learning models were compared with those of logistic models. The net benefit was analyzed using a decision curve to assist in the clinical trade-off between the potential benefits (e.g., PCA diagnosis) and harms (e.g., unnecessary biopsies) of a method.Table 1 Characteristics of the study participants.Full size tableStatistical analysisStatistical analyses were performed using R version 4.2.2 and Stata MP version 17.0. All the statistical tests were two-sided, and a p value less than 0.05 was considered to indicate statistical significance. The predictive mean matching method was used for the imputation of missing candidate predictors, while k-nearest neighbors imputation with k set to five was used. The Wilcoxon rank-sum test was used to compare the levels of continuous variables (as they had skewed distributions) between the PCA and non-PCA groups. The number needed to biopsy (NNB) was calculated as the inverse of the absolute risk difference between the PCA and non-PCA groups.ResultsParticipantsA total of 310 men met the inclusion and exclusion criteria and were included in this study. The age of the 310 study participants ranged from 32 to 93 years; 126 (40.65%) patients had positive biopsy results (PCA) (Gleason score: 6–9), and the remaining 184 non-PCA patients had negative biopsy results (moderate or severe BPH). Age (P = 0.001), irregular margin (P