Deep learning model outperforms traditional models in clinical data-based prognostic prediction for adult-type...

Area of Science:

• Neuro-oncology
• Computational biology
• Biostatistics

Background:

• Adult-type diffuse glioma (ADG) presents significant heterogeneity and data challenges.
• Accurate survival prediction is crucial for guiding treatment decisions in ADG patients.
• Traditional statistical models may not fully capture complex prognostic factors in ADG.

Purpose of the Study:

• To systematically compare traditional statistical methods with machine learning approaches for survival prediction in ADG.
• To evaluate the performance, interpretability, and clinical applicability of different predictive models.
• To identify the most effective model for predicting survival in heterogeneous ADG patient cohorts.

Main Methods:

• Utilized two public and one private retrospective ADG datasets.
• Developed and validated four survival prediction models: Cox Proportional Hazards, Random Survival Forest, Neural Multi Task Logistic Regression, and DeepSurv.
• Conducted sensitivity analyses for missing value imputation strategies.

Main Results:

• Deep learning models, particularly DeepSurv, outperformed traditional Cox models and Random Survival Forest in both internal and external validation cohorts.
• Identified key prognostic factors including age, molecular pathology, chemotherapy, extent of resection, extent of disease, and radiotherapy.
• Confirmed model stability through sensitivity analysis; the DeepSurv model is publicly available.

Conclusions:

• The DeepSurv model demonstrates superior survival prediction performance and stability for heterogeneous, partially missing ADG data compared to conventional methods.
• Deep learning offers a promising avenue for improving prognostic accuracy in neuro-oncology.
• The developed DeepSurv model provides a robust tool for clinical application in ADG patient management.