A deep learning framework (DLF) to identify clinical predictors of disease progression and mortality in metastatic castration-resistant prostate cancer (mCRPC) patients treated with docetaxel/prednisone (DP).

Abstract

e13550 Background: Treatment with DP improves survival in mCRCP but is associated with significant toxicity. The question remains as to whether the improved survival is worth the toxicity risk. In this study, we have investigated the ability of a DLF to identify those patients on which treatment with DP is likely to be beneficial. Methods: The dataset (n = 2028) included a compilation of records from 4 randomized phase 3 trials (NCT00273338, NCT00988208, NCT00617669, and NCT00519285) in which the comparator arm consisted of D (75 mg/m2 q3w) plus P (5mg PO bid) as first-line therapy for mCRPC patients. Data was obtained from www.projectdatasphere.org and contained more than 150 clinical variables at baseline. These were used to generate synthetic state representations (SSR) of every patient that were then input into the DLF to identify subgroup of patients based on their similarities. The resultant subgroups were correlated with progression-free survival (PFS) and overall survival (OS). Results: DLF identified three patient subgroups with specific clinical traits: LL (n = 438), HL (n = 386) and HH (n = 1204). These subpopulations varied in clinical outcome after DP treatment. LL patients (median PFS 19.3 months) had a lower risk of progression compared to HL (median PFS 8.2 months; HR 0.32, 95% CI 0.26-0.40, p < .0001) and HH (median PFS: 9.2 months; HR 0.44, 95% CI 0.36-0.52, p < .0001). No differences were observed for HL and HH. In reference to OS, patients in LL (median OS not reached) and HL (median OS 27 .2 months) did not show any difference; however, both subpopulations showed a lower risk of death compared to HH (median OS 17.7 months) (HR 0.45, 95% CI 0.38-0.54, p < 0.001; HR 0.55 95% CI 0.46-0.62, p < 0.001, respectively). Feature contribution analysis showed that LL signature was associated with ECOG0 and lower levels of PSA, LDH, ALP and AST. LL patients had received less cancer therapy since diagnosis and were more treated with biguanides. In contrast, HL signature had more patients with ECOG 1, and intermediate levels of PSA, LDH, ALP and AST. HL patients had received more hormonal therapy since diagnosis and were more treated with HMG-COA reductase inhibitors. Finally, HH signature was characterised with the highest levels of PSA, LDH, ALP and AST and were more treated with opioids. Other major differences were observed on anthropometrics, vital signs, testosterone, albumin and ALT levels, and non-opioid concomitant medications. Conclusions: We show a methodology that identifies distinct baseline clinical features correlated to the risk of progression and death after DP treatment. While evident for LL (PFS and OS) and HL (OS), HH patients would not receive any treatment benefit on survival. Further work is required to validate this approach as a novel predictive tool for DP treatment decision making on mCRPC patients.