Profiling of clinically relevant subgroups of asymptomatic or mild-symptomatic metastatic castration-resistant prostate cancer (mCRPC) patients using PSA and testosterone levels and deep learning.

Abstract

e17032 Background: The optimal time to begin therapy in mCRPC patients who have either no, or minimal symptoms is not defined yet. Estimating prognosis of these patients can avoid undertreatment or overtreatment and guide the follow-up intensity. This study has investigated the ability of a deep learning framework (DLF) to identify asymptomatic or mildly symptomatic mCRPC patients with high risk of disease progression and mortality using PSA and testosterone levels. Methods: Data from the control arm of the NCT00554229 trial was obtained from www.projectdatasphere.org . These mCRPC patients were treated with placebo alone. The dataset consisted of more than 150 clinical variables. We generated synthetic fingerprints (SF) for each patient through the integration of PSA values collected at baseline and during the first 90 days of follow-up and testosterone levels at baseline. These SF were then input into a DLF to identify subgroup of patients based on their similarities. The resultant subgroups were correlated with progression-free survival (PFS) and overall survival (OS). Feature contribution analysis identified the significant predictive clinical variables in each subgroup. Results: After discarding missing data, 189 patients were eligible for this exploratory study. Median PFS and OS for the total population were 8.7 months and 24.5 months, respectively. DLF identified two different subpopulations, SPA (n = 89) and SPB (n = 100), with significantly different survival outcome. Patients in SPA had a lower risk of progression (median PFS 12.2 months vs. 7.2 months; hazard ratio 0.54, 95% CI 0.41-0.71; p < .0001) and death (median OS not reached vs. 20.3 months; hazard ratio 0.33, 95% CI 0.21-0.52; p < .0001) compared to patients in SPB. SPA signature was significantly correlated to leuprolide treatment. These patients also had higher levels of haemoglobin, red blood cells, haematocrit, total bilirubin, Ca, phosphate, and K. Instead, SPB signature was significantly correlated with presence of > 3 bone metastasis, lack of energy or forced time to bed in the FACT-P and moderate problems in selfcare in the EQ-5D, and treatment with goserelin, tamsulosin, opioids and contact laxatives. In addition, these patients had higher levels of platelets, basophils, monocytes, ALP, AST, LDH, PSA, free PSA, Mg, Na, and specific gravity. Conclusions: DLF identifies mCRPC subtypes that correlate with higher risk of progression and death according to PSA and testosterone levels. Model predictions also suggest complex interactions among many clinical features that might be important modulators of key clinical traits and outcome. Further work is required to validate this approach to anticipate the clinical course of asymptomatic or mildly symptomatic mCRPC patients and help to better decide when to start appropriate therapy.