Persistence of data-driven knowledge to predict breast cancer survival

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Machine learning (ML) models can enhance breast cancer survival prediction, particularly when tailored to cancer stage at diagnosis.
• The influence of clinical parameters and the predictive accuracy of ML models may evolve over time.

Purpose of the Study:

• To assess whether findings on clinical parameters and ML model performance in breast cancer survival prediction are temporary or permanent.
• To determine the validity period of newly generated knowledge in breast cancer survival prediction.

Main Methods:

• A systematic review identified 15 relevant conclusions on ML for breast cancer survival prediction.
• Data-driven ML models were constructed using SEER database breast cancer data (1988-2009) to predict five-year survival.
• Persistence analysis evaluated the temporal validity of model predictive quality and clinical parameter importance across three ML methods and stage-specific/joint models.

Main Results:

• Only 53% of previously published conclusions held true for SEER cases from 1988-2009, with only 75% remaining true over time.
• Previously accepted conclusions, such as the limited impact of additional data on frequent stages or the importance of tumor grade for distant metastasis, were found to be false under temporal analysis.
• Significant shifts in the predictive value of clinical parameters and model performance were observed over the study period.

Conclusions:

• Data-driven insights derived from machine learning in breast cancer research require rigorous temporal validation.
• The clinical and professional application of ML-generated knowledge is contingent upon demonstrating its durability over time.
• Continuous monitoring and re-evaluation of ML models are essential for reliable breast cancer survival prediction.