The construction of prognostic prediction models based on follow-up data is crucial for devising individualized treatment plans for patients. However, the performance of current supervised survival analysis methods is constrained due to the prevalence of weakly supervised censored samples during follow-up. To address this limitation, this study introduces the Prognostic Co-Training Regression (PCTR) algorithm, a semi-supervised prognostic prediction model developed through the co-training of two KNN regressors. By integrating the prior information of censored data, PCTR harnesses the prior information embedded in censored data, effectively extracting latent prognostic insights, thereby constructing machine learning models with enhanced prognostic accuracy. Validating this approach, we extracted and selected radiomic features from CT imaging data of 523 patients with gastrointestinal stromal tumors. The PCTR algorithm demonstrated superior performance over commonly used Cox Proportional Hazards and Random Survival Forest algorithms in external test cohort, offering clinical researchers a more effective method for prognostic model development.
