FACETS: AN UNCERTAINTY-AWARE PATIENT-LEVEL SURVIVAL SIMULATION FRAMEWORK FOR PRECISION HEALTH TECHNOLOGY ASSESSMENT: VALIDATION ACROSS REDUCED AND ENRICHED METABRIC BREAST CANCER COHORTS
Author(s)
Salma Barkaoui, PhD1, Mohammed BENNANI, PhD2, Hadhami Mejbri, Master3, Ikrame Nouigua, Master3, Tommy Dang, Master4, Caroline Philippe, Master2, Jerome Vetillard, PhD5.
1Qualees, PARIS, France, 2QUALEES, PARIS, France, 3QUALEES, Paris, France, 4Qualees, Qualees, France, 5Qualees, Paris, France.
1Qualees, PARIS, France, 2QUALEES, PARIS, France, 3QUALEES, Paris, France, 4Qualees, Qualees, France, 5Qualees, Paris, France.
OBJECTIVES: Health technology assessment (HTA) increasingly relies on patient-level microsimulation to support precision medicine, subgroup evaluation, and decision making. This requires survival models combining individualized prediction, predictive uncertainty, and scenario simulation. We evaluated FACETS, a transformer-based survival simulation framework, using reduced and enriched versions of the METABRIC breast cancer cohort.
METHODS: The METABRIC cohort (n=1,904; long-term overall survival) was evaluated using (A) a reduced dataset containing nine clinical-biological variables and (B) an enriched dataset including 35 clinical, pathological, treatment, and molecular variables. FACETS represents each patient characteristic as an independent transformer token and jointly optimizes discrete-time survival likelihood and concordance ranking. Predictive uncertainty was estimated using deep ensembles and Monte Carlo Dropout. Performance was compared with Cox proportional hazards (Cox PH), Random Survival Forest (RSF), Gradient Boosting Survival (GBS), and DeepSurv using Harrell's C-index and Integrated Brier Score (IBS). Treatment scenarios modified treatment inputs and were interpreted as predictive rather than causal.
RESULTS: In the reduced dataset, FACETS achieved the highest predictive performance (C-index=0.654; IBS=0.165), outperforming Cox PH (0.619/0.173), RSF (0.628/0.169), GBS (0.615/0.174), and DeepSurv (0.610/0.181). In the enriched dataset, discrimination improved (C-index=0.700; IBS=0.178). Although RSF and GBS achieved similar discrimination (0.695/0.157 and 0.686/0.156), neither provided uncertainty quantification or integrated scenario simulation, while DeepSurv showed poorer calibration (IBS=0.236). FACETS generated individualized 90% predictive intervals with mean widths of 0.20 and 0.25 at five years in the reduced and enriched datasets, respectively. Among estrogen receptor-positive patients, hormone therapy scenarios predicted absolute five-year survival differences of 0.7 and 3.9 percentage points, highlighting the added value of richer molecular information.
CONCLUSIONS: FACETS demonstrated competitive predictive performance while integrating survival prediction, predictive uncertainty, and patient-level treatment scenario simulation. FACETS provides an upstream survival simulation engine supporting uncertainty-aware microsimulation, subgroup analyses, and precision-oriented economic evaluation. External validation and causal treatment modelling remain important next steps.
METHODS: The METABRIC cohort (n=1,904; long-term overall survival) was evaluated using (A) a reduced dataset containing nine clinical-biological variables and (B) an enriched dataset including 35 clinical, pathological, treatment, and molecular variables. FACETS represents each patient characteristic as an independent transformer token and jointly optimizes discrete-time survival likelihood and concordance ranking. Predictive uncertainty was estimated using deep ensembles and Monte Carlo Dropout. Performance was compared with Cox proportional hazards (Cox PH), Random Survival Forest (RSF), Gradient Boosting Survival (GBS), and DeepSurv using Harrell's C-index and Integrated Brier Score (IBS). Treatment scenarios modified treatment inputs and were interpreted as predictive rather than causal.
RESULTS: In the reduced dataset, FACETS achieved the highest predictive performance (C-index=0.654; IBS=0.165), outperforming Cox PH (0.619/0.173), RSF (0.628/0.169), GBS (0.615/0.174), and DeepSurv (0.610/0.181). In the enriched dataset, discrimination improved (C-index=0.700; IBS=0.178). Although RSF and GBS achieved similar discrimination (0.695/0.157 and 0.686/0.156), neither provided uncertainty quantification or integrated scenario simulation, while DeepSurv showed poorer calibration (IBS=0.236). FACETS generated individualized 90% predictive intervals with mean widths of 0.20 and 0.25 at five years in the reduced and enriched datasets, respectively. Among estrogen receptor-positive patients, hormone therapy scenarios predicted absolute five-year survival differences of 0.7 and 3.9 percentage points, highlighting the added value of richer molecular information.
CONCLUSIONS: FACETS demonstrated competitive predictive performance while integrating survival prediction, predictive uncertainty, and patient-level treatment scenario simulation. FACETS provides an upstream survival simulation engine supporting uncertainty-aware microsimulation, subgroup analyses, and precision-oriented economic evaluation. External validation and causal treatment modelling remain important next steps.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
HTA351
Topic
Clinical Outcomes, Health Technology Assessment, Methodological & Statistical Research
Topic Subcategory
Decision & Deliberative Processes
Disease
Oncology