Medically complex children refer to patients who need intense medical care due to multisystem dysfunction, technology dependence, or complex medication needs. Medically complex patients often consume largely disproportionate amount of different care resources in hospitals. Understanding the nature of complexity and predicting the dynamics of complexity in patients can help design better personalized care plans in order to improve quality and reduce cost. Due to the heterogeneity in the complexity, we still do not have precise definitions of medically complex patients (MCP), nor the understanding of different MCP phenotypes. Finally, in order to directly impact care process, we also need predictive scores for each MCP phenotypes so that we can identify the high-risk patients earlier. Given the massive amount of Electronic Health Records (EHR) collected at Children’s healthcare of Atlanta (CHOA), now it is possible to develop machine learning techniques to gain insights into underlying patient phenotypes.
In this project, we pursue three aims:
Aim 1. Defining medically complex patients: The first step in analysis and management of medically complex patients is to understand which patients should be considered complex. While there are several approaches for defining MCP based on various medical conditions, the complete identification of all MCP is yet to be fully understood. We argue that a more general approach is to refine the existing definitions by incorporating the natural patterns found in EHR data. That is, to use co-occurrence patterns in EHR data to refine the definition of medically complex patients. The data-driven approach will identify additional patients that are similar to complex patients, but normally identified as non-complex by clinical rules.
Aim 2. Phenotyping medically complex patients: After identifying MCPs using the refined definition in aim 1, it is crucial to group patients into different homogeneous phenotypes so that we can customizing care plans and allocating the required hospital resources caring each phenotype group. We not only need to be able to find the number of phenotypes, but also we need to ensure clinical interpretation of each phenotype.
Aim 3. Predicting medically complex patients : Finally, after knowing all MCP phenotypes, it is important to identify high risk patients who are rapidly becoming complex, so that early intervention can be applied. To calculate the MCP risk for patients, we plan to design algorithms that capture the temporal dependency such as recurrent neural networks so that forecast the dynamics of complexity in each patient.
Collaboration with Children's Healthcare of Atlanta