Calibration and Validation of the Colorectal Cancer and Adenoma Incidence and Mortality (CRC-AIM) Microsimulation Model Using Deep Neural Networks
2023
This study explores the efficacy of machine learning (ML)-based emulators in calibrating complex microsimulation models, using the Colorectal Cancer (CRC)-Adenoma Incidence and Mortality (CRC-AIM) model as a case study. ML algorithms, including deep neural networks (DNN), were trained and compared using data generated from the CRC-AIM model to predict various outcomes. The DNN outperformed other algorithms and efficiently predicted outcomes, reducing computational burden significantly. The calibrated CRC-AIM model demonstrated cross-model validity against established CISNET models and external validity against a randomized controlled trial. Calibration targets greatly influenced model outcomes, particularly life-year gains with screening. The study highlights the potential of carefully selected and trained ML-based emulators to streamline the calibration process in microsimulation models, enhancing their computational efficiency and accuracy. The methods involved using a DNN model for calibration, generating and training the model with CRC-AIM data, and validating the calibrated model against established models and clinical trials.
Source:
Vahdat V, Alagoz O, Chen JV et al. Calibration and Validation of the Colorectal Cancer and Adenoma Incidence and Mortality (CRC-AIM) Microsimulation Model Using Deep Neural Networks. Medical Decision Making 2023; 43 (6): 719-736. https://doi.org/10.1177/0272989X231184175