Presentation (CTI)

OBJECTIVES: Traditional modeling methods such as decision trees, Markov cohort models often incorporate highly simplified care pathways that do not align with actual clinical complexity, due to their manual construction and limited flexibility. This study investigates a data-driven approach for constructing simulation model structures directly from real-world data (RWD). This supports more accurate reflection of real-world clinical complexity and patient heterogeneity as compared to guideline or expert-based model structures, enhancing the value of health economic modeling.
METHODS: We use process mining (PM) techniques to identify patient pathways from event logs, which are detailed, timestamped records of healthcare information systems. These pathways inform the structure of a Generalized Stochastic Petri Net (GSPN) with immediate transitions, serving as a simulation framework. This approach allows semi-automated development of simulation models. To demonstrate this, we generate synthetic event logs using the well-known Sick-Sicker micro-simulation model and use PM techniques to reconstruct its underlying logic.
RESULTS: We ran the Sick-Sicker simulation model 500 times for 10,000 individuals, producing a total of 500 logs to inform the GSPN model's structure and transition probabilities by PM. The root mean squared error (RMSE) for transitions probabilities ranged from 0.00000 and 0.00014, indicating minimal error. The resulting GSPN model was also run 500 times, with each simulated log achieving a fitness score of 1.00, showing that the GSPN model accurately captured the inherent patient trajectory dynamics in the data. Additionally, the RMSE and mean absolute error (MAE) for cost and utility measurements were low, further validating the model’s accuracy.
CONCLUSIONS: This study offers a formal mechanism for integrating RWD into health economic modeling and demonstrates that Petri nets, particularly GSPN, informed by PM can accurately reflect complex real-world care pathways. Furthermore, it shows that combining process mining and GSPN supports assessing ‘real-world cost-effectiveness’ while reducing modelling time and efforts.