A deep learning algorithm with good prediction efficacy for cancer-specific survival in osteosarcoma: A retrospective...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning in Medicine

Background:

• Accurate prognosis is essential for effective osteosarcoma (OSC) management and treatment.
• Individualized treatment strategies require reliable prediction of cancer-specific survival (CSS).
• Deep learning and traditional statistical models offer potential for improving prognostic accuracy.

Purpose of the Study:

• To predict cancer-specific survival (CSS) rates in osteosarcoma (OSC) patients.
• To compare the efficacy of deep learning (DeepSurv) and Cox proportional hazard models for OSC prognosis.
• To support individualized treatment decisions for OSC patients through accurate survival prediction.

Main Methods:

• Utilized data from 3218 osteosarcoma patients (2004-2017) from the Surveillance, Epidemiology, and End Results (SEER) database.
• Randomly assigned patients into training (70%) and validation (30%) cohorts.
• Developed prognostic models using the DeepSurv algorithm and the Cox proportional hazard model, evaluating with C-index, IBS, RMSE, and SME.

Main Results:

• Both DeepSurv and Cox models demonstrated strong predictive performance for CSS in OSC patients, with C-indices exceeding 0.74.
• The DeepSurv model did not show significant superiority over the Cox model in predicting survival based on validation metrics.
• A total of 3218 patients were included, split into training (n=2252) and validation (n=966) groups.

Conclusions:

• The CSS prediction model for osteosarcoma patients based on the DeepSurv algorithm achieved satisfactory prediction efficacy after validation.
• The developed DeepSurv model provides a practical and convenient webpage calculator for predicting cancer-specific survival in OSC.
• The study highlights the utility of advanced algorithms like DeepSurv in conjunction with traditional methods for robust prognostic modeling in oncology.