Multi-task oriented diffusion model for mortality prediction in shock patients with incomplete data

Abstract

Mortality prediction based on electronic medical records is crucial for treatment decisions of shock patients in the ICU. Although clinical data on such patients often contain many missing values, the multi-view property of medical data could compensate for such missing information. Traditionally, mortality prediction models are built as two-stage approaches with additional data imputation steps, leading to issues in which the local optimal model at each step may not necessarily obtain a globally optimal solution. To overcome this problem, we conducted a multi-centre study using real-world data and aimed to develop an end-to-end mortality prediction model for shock patients. A Multi-task Oriented Diffusion Model (MODM) is proposed to fill in missing values and predict mortality simultaneously. Specifically, the model incorporates label information from different tasks to guide the optimal direction and effectively reduce uncertainty in the diffusion process. In addition, we propose a self-adjusting training strategy that balances the convergence rates among different tasks. The two largest well-known ICU datasets were used in this study, where 14,278 shock patients from eICU-CRD (2018) were included in the internal experiment, and 5,310 shock patients from MIMIC-IV (2012) were used as an external test. Compared with 14 state-of-the-art methods, the proposed model achieved the best performance with an AUC of 0.7998 in mortality prediction and notably good performance in terms of RMSE (0.0058, 0.0034, 0.0030, 0.0027) and MAE (0.3959, 0.4358, 0.4975, 0.5435) at random missing rates (10%, 30%, 50%, 70%, respectively) in the data imputation stage. The experimental results indicate the superiority of the proposed end-to-end MODM algorithm in handling real-world data. We released our code at .