Enhancing Prognostic Prediction of Gastrointestinal Stromal Tumors Using Semi-Supervised Regression Based on CT...

Area of Science:

• Oncology
• Radiology
• Machine Learning

Background:

• Prognostic prediction models are vital for personalized cancer treatment.
• Supervised survival analysis methods struggle with weakly supervised censored data.
• Censored data in patient follow-up limits the accuracy of current prognostic models.

Purpose of the Study:

• To introduce a semi-supervised prognostic prediction model, the Prognostic Co-Training Regression (PCTR) algorithm.
• To enhance prognostic accuracy by effectively integrating information from censored data.
• To develop a more robust machine learning approach for cancer prognosis.

Main Methods:

• Developed the Prognostic Co-Training Regression (PCTR) algorithm using co-training of two KNN regressors.
• Integrated prior information from censored data to extract latent prognostic insights.
• Extracted and selected radiomic features from CT imaging data of 523 gastrointestinal stromal tumor patients.

Main Results:

• The PCTR algorithm demonstrated superior performance compared to Cox Proportional Hazards and Random Survival Forest algorithms.
• PCTR achieved enhanced prognostic accuracy in an external test cohort.
• The model effectively leveraged censored data for improved predictive power.

Conclusions:

• The PCTR algorithm offers a more effective method for developing accurate prognostic models.
• This semi-supervised approach addresses limitations posed by censored data in survival analysis.
• PCTR provides a valuable tool for clinical researchers in oncology and personalized medicine.