Medical applications of machine learning (ML) have experienced a surge in popularity in recent years. Given the abundance of available data from electronic health records, the intensive care unit (ICU) is a natural habitat for ML. Models have been proposed to address numerous ICU prediction tasks like the early detection of complications. While authors frequently report state-of-the-art performance, it is challenging to verify claims of superiority. Datasets and code are not always published, and cohort definitions, preprocessing pipelines, and training setups are difficult to reproduce. This work introduces Yet Another ICU Benchmark (YAIB), a modular framework that allows researchers to define reproducible and comparable clinical ML experiments; we offer an end-to-end solution from cohort definition to model evaluation. The framework natively supports most open-access ICU datasets (MIMIC III/IV, eICU, HiRID, AUMCdb) and is easily adaptable to future ICU datasets. Combined with a transparent preprocessing pipeline and extensible training code for multiple ML and deep learning models, YAIB enables unified model development, transfer, and evaluation. Our benchmark comes with five predefined established prediction tasks (mortality, acute kidney injury, sepsis, kidney function, and length of stay) developed in collaboration with clinicians. Adding further tasks is straightforward by design. Using YAIB, we demonstrate that the choice of dataset, cohort definition, and preprocessing have a major impact on the prediction performance — often more so than model class — indicating an urgent need for YAIB as a holistic benchmarking tool. We provide our work to the clinical ML community to accelerate method development and enable real-world clinical implementations.

INTRODUCTION: In recent years, reinforcement learning (RL) has gained traction in the healthcare domain. In particular, RL methods have been explored for haemodynamic optimization of septic patients in the Intensive Care Unit. Most hospitals however, lack the data and expertise for model development, necessitating transfer of models developed using external datasets. This approach assumes model generalizability across different patient populations, the validity of which has not previously been tested. In addition, there is limited knowledge on safety and reliability. These challenges need to be addressed to further facilitate implementation of RL models in clinical practice. METHOD: We developed and validated a new reinforcement learning model for hemodynamic optimization in sepsis on the MIMIC intensive care database from the USA using a dueling double deep Q network. We then transferred this model to the European AmsterdamUMCdb intensive care database. T-Distributed Stochastic Neighbor Embedding and Sequential Organ Failure Assessment scores were used to explore the differences between the patient populations. We apply off-policy policy evaluation methods to quantify model performance. In addition, we introduce and apply a novel deep policy inspection to analyse how the optimal policy relates to the different phases of sepsis and sepsis treatment to provide interpretable insight in order to assess model safety and reliability. RESULTS: The off-policy evaluation revealed that the optimal policy outperformed the physician policy on both datasets despite marked differences between the two patient populations and physician's policies. Our novel deep policy inspection method showed insightful results and unveiled that the model could initiate therapy adequately and adjust therapy intensity to illness severity and disease progression which indicated safe and reliable model behaviour. Compared to current physician behavior, the developed policy prefers a more liberal use of vasopressors with a more restrained use of fluid therapy in line with previous work. CONCLUSION: We created a reinforcement learning model for optimal bedside hemodynamic management and demonstrated model transferability between populations from the USA and Europe for the first time. We proposed new methods for deep policy inspection integrating expert domain knowledge. This is expected to facilitate progression to bedside clinical decision support for the treatment of critically ill patients.