AI-DRIVEN SYNTHETIC PATIENT GENERATION FRAMEWORK TO MODEL OBESITY USING THE CONSTANCES COHORT
Author(s)
Diane Vincent, MSc1, Louise Dry, MSc1, Nina Temam, PharmD1, Antoine Movschin, MSc1, Sofiane Kab, PharmD2, Antoine Duclos, MD, PhD3, Pauline Guilmin, MSc1.
1Quinten Health, Paris, France, 2Paris Cité University, Paris-Saclay University, UVSQ, Inserm, Epidemiological Population Cohorts Unit (UMS11), Villejuif, France, 3Paris Cité University, Paris-Saclay University, UVSQ, Inserm, Epidemiological Population Cohorts Unit (UMS11), Villejuif, France and Paris Cité University and Sorbonne Paris Nord University, Inserm, INRAE, Centre for Research in Epidemiology and Statistic, Paris, France.
1Quinten Health, Paris, France, 2Paris Cité University, Paris-Saclay University, UVSQ, Inserm, Epidemiological Population Cohorts Unit (UMS11), Villejuif, France, 3Paris Cité University, Paris-Saclay University, UVSQ, Inserm, Epidemiological Population Cohorts Unit (UMS11), Villejuif, France and Paris Cité University and Sorbonne Paris Nord University, Inserm, INRAE, Centre for Research in Epidemiology and Statistic, Paris, France.
OBJECTIVES: Obesity is a chronic, heterogeneous disease associated with substantial cardiovascular burden and diverse patient profiles, creating challenges for evidence generation and healthcare decision-making. This study aimed to develop an artificial intelligence (AI)-driven framework capable of generating and evaluating realistic synthetic real-world Obesity patient cohorts that preserve both patient heterogeneity and privacy.
METHODS: Data were obtained from the French CONSTANCES cohort linked to the national health claims database (SNDS). Adults with obesity, defined as body mass index (BMI) ≥30 kg/m² or overweight defined as BMI ≥27 kg/m² with weight-related comorbidities, were identified using BMI measurements recorded in CONSTANCES. Baseline characteristics included clinical variables, medical history, and lifestyle. A Cox proportional hazards (CPH) model was developed to predict time to first 5-point major adverse cardiovascular event (5P-MACE) and evaluated on an independent test set. Synthetic baseline patient characteristics were generated using a Gaussian copula approach, and longitudinal outcomes were simulated using the trained CPH model combined with inverse transform sampling. Synthetic data performance was assessed across three domains: fidelity through comparisons of generated and observed variable distributions and correlation structures; utility, through agreement of generated and observed Kaplan-Meier outcome trajectories; and privacy through disclosure-risk assessment. Fidelity and utility were assessed in the overall population and across clinically relevant subgroups defined by cardiovascular comorbidities, diabetes and sex.
RESULTS: The framework was developed using data from 29,046 patients with Obesity. High fidelity and utility were observed across the overall population and clinically relevant subgroups, with close agreement in baseline characteristics and 5P-MACE trajectories, while maintaining promising privacy-preserving properties.
CONCLUSIONS: This AI-driven framework generated realistic synthetic Obesity cohorts that may enable augmentation of under-represented subgroups or simulation of alternative therapeutic trajectories, thereby supporting research on progression, cardiovascular outcomes, and long-term prevention strategies to inform clinical and healthcare decision-making in Obesity.
METHODS: Data were obtained from the French CONSTANCES cohort linked to the national health claims database (SNDS). Adults with obesity, defined as body mass index (BMI) ≥30 kg/m² or overweight defined as BMI ≥27 kg/m² with weight-related comorbidities, were identified using BMI measurements recorded in CONSTANCES. Baseline characteristics included clinical variables, medical history, and lifestyle. A Cox proportional hazards (CPH) model was developed to predict time to first 5-point major adverse cardiovascular event (5P-MACE) and evaluated on an independent test set. Synthetic baseline patient characteristics were generated using a Gaussian copula approach, and longitudinal outcomes were simulated using the trained CPH model combined with inverse transform sampling. Synthetic data performance was assessed across three domains: fidelity through comparisons of generated and observed variable distributions and correlation structures; utility, through agreement of generated and observed Kaplan-Meier outcome trajectories; and privacy through disclosure-risk assessment. Fidelity and utility were assessed in the overall population and across clinically relevant subgroups defined by cardiovascular comorbidities, diabetes and sex.
RESULTS: The framework was developed using data from 29,046 patients with Obesity. High fidelity and utility were observed across the overall population and clinically relevant subgroups, with close agreement in baseline characteristics and 5P-MACE trajectories, while maintaining promising privacy-preserving properties.
CONCLUSIONS: This AI-driven framework generated realistic synthetic Obesity cohorts that may enable augmentation of under-represented subgroups or simulation of alternative therapeutic trajectories, thereby supporting research on progression, cardiovascular outcomes, and long-term prevention strategies to inform clinical and healthcare decision-making in Obesity.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR130
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
Diabetes/Endocrine/Metabolic Disorders (including obesity)