Prediction of Patient’s Response to an Acute Inflammation Treatment by Mixture of Experts

Acute inflammation is a progressive medical condition characterized as systemic inflammatory response to an infection. If inadequately diagnosed and treated, acute inflammation culminates in sepsis, which is the leading cause of death in non-coronary intensive care units in the United States. In practice, clinicians make diagnostic and therapeutic decisions based on their experience and understanding of the relationships between treatments and expected outcomes of such treatments, which is often suboptimal due to limited knowledge about patient response to therapy. To develop adequate sepsis treatment strategies, accurate prediction of patient response to therapy is essential. However, complex multi-stage nature of acute inflammation makes patient response to treatment difficult to predict. The complexity and heterogeneity of inflammatory response that consists of different stages with significantly different properties is one of the main reasons of unsuccessful treatments. Therefore, there is a critical need for tools for predicting patient’s response to therapy, which can help clinicians to design more effective therapies for this potentially fatal condition. Towards this objective we developed a data-driven prediction model based on mixture of experts. The proposed model identifies stages (regimes) of patient state and builds specialized predictors for each stage. In experiments conducted on 500 virtual patients generated by a mathematical model for acute inflammation, our method not only outperformed all baselines accuracy wise but also identified three stagespecific experts that can play an important role in improving perception of complex systems such as response to acute inflammation treatment.