Quantified treatment effect at the individual level is more indicative for personalized radical prostatectomy recommendation: implications for prostate cancer treatment using deep learning.

Background: There are potential uncertainties and overtreatment existing in radical prostatectomy (RP) for prostate cancer (PCa) patients, thus identifying optimal candidates is quite important.

Purpose: This study aims to establish a novel causal inference deep learning (DL) model to discern whether a patient can benefit more from RP and to identify heterogeneity in treatment responses among PCa patients.

Methods: We introduce the Self-Normalizing Balanced individual treatment effect for survival data (SNB). Six models were trained to make individualized treatment recommendations for PCa patients. Inverse probability treatment weighting (IPTW) was used to avoid treatment selection bias.

Results: 35,236 patients were included. Patients whose actual treatment was consistent with SNB recommendations had better survival outcomes than those who were inconsistent (multivariate hazard ratio (HR): 0.76, 95% confidence interval (CI), 0.64-0.92; IPTW-adjusted HR: 0.77, 95% CI, 0.61-0.95; risk difference (RD): 3.80, 95% CI, 2.48-5.11; IPTW-adjusted RD: 2.17, 95% CI, 0.92-3.35; the difference in restricted mean survival time (dRMST): 3.81, 95% CI, 2.66-4.85; IPTW-adjusted dRMST: 3.23, 95% CI, 2.06-4.45). Keeping other covariates unchanged, patients with 1 ng/mL increase in PSA levels received RP caused 1.77 months increase in the time to 90% mortality, and the similar results could be found in age, Gleason score, tumor size, TNM stages, and metastasis status.

Conclusions: Our highly interpretable and reliable DL model (SNB) may identify patients with PCa who could benefit from RP, outperforming other models and clinical guidelines. Additionally, the DL-based treatment guidelines obtained can provide priori evidence for subsequent studies.