Cancer prognosis using support vector regression in imaging modality.

Area of Science:

• Biomedical data analysis
• Machine learning in oncology
• Radiomics and cancer imaging

Background:

• Cancer prognosis relies on accurate prediction from patient data.
• Support vector regression (SVR) offers a potential method for analyzing censored data in cancer prognosis.
• Imaging features present a rich source of information for cancer outcome prediction.

Purpose of the Study:

• To investigate the efficacy of SVR in cancer prognosis using imaging features.
• To adapt SVR methods for handling censored data in cancer prediction.
• To evaluate the impact of sampling and feature selection techniques on SVR performance.

Main Methods:

• Extracted cancer image features from patient data, recorded as censored data.
• Employed stratified sampling based on tumor size and lymph node status.
• Utilized three feature selection methods: none, individual, and forward selection.
• Compared SVR performance against Cox regression using concordance index (CI) and Brier score (BS).

Main Results:

• SVR demonstrated comparable or superior performance to Cox regression across different feature selection strategies.
• Stratified sampling of tumor size yielded optimal regression results.
• The best performance (CI: 0.6845, BS: 0.2065) was achieved with SVR, individual feature selection, and stratified sampling of lymph node status.
• SVR exhibited more consistent performance than Cox regression.

Conclusions:

• Combinational methods of SVR, feature selection, and sampling enhance cancer prognosis accuracy.
• While effective, further improvements may be achieved with more significant imaging features.
• SVR provides a robust alternative for cancer prognosis modeling with censored imaging data.