Explainable ML models for a deeper insight on treatment decision for localized prostate cancer

Area of Science:

• Urology
• Oncology
• Health Informatics

Background:

• Existing prostate cancer (PCa) treatment decision aids lack consistency and certainty.
• Understanding the multifaceted factors influencing treatment choices is crucial for localized PCa patients.

Purpose of the Study:

• To develop and validate predictive models for localized prostate cancer treatment decisions.
• To identify key oncologic, demographic, socioeconomic, and geographic factors influencing treatment selection.
• To visualize complex feature interactions in the PCa treatment decision-making process.

Main Methods:

• Utilized a large real-world database (SEER-PWW) of 255,837 men diagnosed with localized PCa (2010-2015).
• Developed two prediction models: one for the entire cohort (AS/WW, RP, RT) and one for low-risk patients (AS/WW).
• Evaluated model discrimination using multiclass AUC and employed Shapley Additive Explanations (SHAP) for interpretability.

Main Results:

• The overall decision-predicting model achieved a multiclass AUC of 0.77; the low-risk model achieved an AUC of 0.74.
• Oncological variables significantly impacted radical prostatectomy (RP) decisions.
• Geographic factors were highly influential in radiation therapy (RT) decisions.

Conclusions:

• Developed accurate predictive models for localized prostate cancer treatment selection using real-world data.
• Demonstrated the significant influence of diverse factors, including geographic location, on treatment decisions.
• Visualized nonlinear feature interactions, offering insights into the complex PCa treatment decision process.