Prediction models have been widely adopted in clinical research and medical practice, yet there are often major concerns about whether such models can be generalized across different populations and clinical settings. Relatively little research has been done to develop general approaches that can improve model performance for heterogeneous populations within different clinical practices ("clusters"). The authors set out to fill this gap.
Using a simulation, they sought to measure how much we can potentially improve accuracy by using dynamic prediction models in the context of clustered data. These models update predictions as new data accrue and can account for site-specific differences.
They found that dynamic mixed-effects models substantially improved prediction model accuracy across a broad range of simulated conditions. These models can serve as a useful alternative to standard static ones, for improving the generalizability of clinical prediction in settings of clustered data. For example, when we use existing models in varied settings where patient populations get cardiac surgeries, our predictions of the risks involved may prove to be inaccurate. Dynamic prediction modeling can improve the accuracy by updating models to account for these settings. Thus they are well worth the logistical challenges that may accompany their implementation in practice.