(P-521) Towards OS approximation in early clinical trials using Prognostic Scores

Background: Drug development (in Oncology) is an extensive process composed of multiple phases, with early phases focusing on dosing/safety and later phases focusing on efficacy. In order to avoid: (1) abandoning potentially effective drugs too early, and (2) continuing to develop less or not effective drugs, it is critical to obtain a robust estimate of the drugs efficacy as early as possible. In Oncology trials, the gold-standard efficacy endpoint is Overall Survival (OS). However, in early phases, surrogate endpoints e.g., response rate, are often used due to the limited duration ( <1.5 years) and small number of patients ( <150). In many cases though, these surrogate endpoints do not correlate well with OS.

Objectives: To determine whether it is possible to estimate efficacy in initial clinical trials earlier and/or more accurately than other surrogate endpoints by using prognostic scores.

Methods: In this study we used the nationwide Flatiron Health (FH) electronic health record (EHR)-derived de-identified database. From FH, we selected a cohort of 89,201 patients with 16 different primary cancers, all of which had date-of-death available. For all patients, a total of 27 covariates were measured monthly between the date of death and 24 months before death. We developed two different prognostic scores (fit on start of treatment), one based on the Cox model (ROPRO) and another based on DeepSurv (DS). With these pre-trained models, we calculate Δrisk as the monthly variation (between month m and m+1) of the risk score for each patient and time point. We hypothesize that for Δrisk to be a good OS surrogate it should be: (1) lower at earlier time-points, (2) higher closer to death, and (3) continuously increasing over time.

Results: The median survival time of the selected cohort was 327 (95% CI 324-330) days. Both Δrisk scores are low at the beginning, high at the time of death and increasing in between. Additionally we observed that the two models have different patterns over time. Further away from the date of death (months -24 to -6) ROPRO has a higher slope (2.268e-4 ± 3.016e-5) than DS (1.261e-5 ± 2.299e-5). By contrast, closer to the date of death (from -6 months to death) ROPRO has a lower slope (4.154e-4 ± 1.265e-2) than DS (1.302e-2 ± 3.901e-4).

Conclusions: Both Δrisk scores correctly increase closer to the date of death. Therefore, we consider that both scores could be used as early OS proxies and thus could complement currently used surrogate endpoints. The ROPRO Δrisk increased gradually over time which might help signal a worsening of the patients’ condition earlier on. Conversely, the sharp increase in DS Δrisk when close to death might serve as a precise short-term indicator of the decrease of the patients’ health.