Zhonghua Nan Ke Xue · Jul 2026
Research on screening, validation and risk prediction of key biomarkers for prostate cancer
By Han Y, Xu W, Wang C et al.
Objective A non-invasive risk prediction tool based on the localized DeepSeek large model will be proposed in this article, aiming to address the high invasiveness and data privacy issues of traditional prostate cancer screening methods. Methods Utilizing the 2 652 cases of prostate cancer dataset from the National Population Health Science Data Repository, key indicators were screened through LASSO regression and Logistic regression. And training and validation were conducted using the locally deployed DeepSeek large model. Results Nine indicators such as apolipoprotein B, fPSA, inorganic phosphorus, tPSA, and β -globulin were independent risk factors for prostate cancer. The prediction accuracy rates of the DeepSeek model in the prostate cancer group, the hyperplasia group, and the cancer with hyperplasia group were 79% (95%CI: 75.1%-82.4%), 84.1% (95%CI: 80.5%-87.2%), and 84.5% (95%CI: 81.2% -87.3%), respectively. The AUC of the ROC curve of the model was 0.869, demonstrating a high discriminatory ability. Conclusion DeepSeek can precisely distinguish prostate cancer from hyperplasia through biochemical indicators, reducing unnecessary puncture biopsies. The localized closed-loop application of DeepSeek provides efficient, privacy-protected and secure screening solutions for primary medical institutions. In the future, it is necessary to combine multimodal data to optimize the generalization ability of the model and promote the development of individualized diagnosis and treatment.