Abstract

Abstract for: Calibrating a Hypertension Model Across 20 Cities Using Open Data: Workflow and Parameter Identifiability

Introduction Deploying System Dynamics models across many cities requires scalable calibration. Each city demands its own parameter set, but manual tuning does not scale and full Bayesian methods require substantial infrastructure. A standardized, reproducible calibration workflow using publicly available data is needed.
Approach We calibrate a sex-stratified hypertension cascade-of-care and CVD model across 20 cities using 500 Latin Hypercube starting points per city, staged Powell optimization, and three open data sources (UN WPP, NCD-RisC, GBD). All optimized solutions are retained for identifiability analysis.
Results The calibration achieves good fit across all 20 cities. The multi-start design reveals structural parameter non-identifiability: the same compensating parameter pairs recur across all cities, indicating model-inherent property rather than data-specific ambiguity.
Discussion The non-identifiability is rooted in the cascade architecture and cannot be resolved by adding more cities. We propose a four-step workflow where identifiability diagnostics from calibration directly inform which parameters require expert review, enabling more efficient stakeholder engagement.

Introduction	Deploying System Dynamics models across many cities requires scalable calibration. Each city demands its own parameter set, but manual tuning does not scale and full Bayesian methods require substantial infrastructure. A standardized, reproducible calibration workflow using publicly available data is needed.
Approach	We calibrate a sex-stratified hypertension cascade-of-care and CVD model across 20 cities using 500 Latin Hypercube starting points per city, staged Powell optimization, and three open data sources (UN WPP, NCD-RisC, GBD). All optimized solutions are retained for identifiability analysis.
Results	The calibration achieves good fit across all 20 cities. The multi-start design reveals structural parameter non-identifiability: the same compensating parameter pairs recur across all cities, indicating model-inherent property rather than data-specific ambiguity.
Discussion	The non-identifiability is rooted in the cascade architecture and cannot be resolved by adding more cities. We propose a four-step workflow where identifiability diagnostics from calibration directly inform which parameters require expert review, enabling more efficient stakeholder engagement.