Informatics Approaches for Characterizing and Predicting Acute Cardiovascular and Respiratory Decompensation

Abstract

Decompensation is a rapid deterioration of the physiological state of critically ill patients whereby major organ systems cannot adjust to stressors to maintain organ function within normal ranges. Two common examples of decompensation are acute heart failure (AHF) and acute respiratory failure (ARF). Heart failure is a leading cause of mortality and morbidity worldwide. AHF, broadly defined as rapid onset of new or worsening signs and symptoms of heart failure, often requires hospitalization and admission to the intensive care unit (ICU). This acute condition is highly heterogeneous and less well-understood as compared to chronic heart failure. ARF is the most common syndrome in patients requiring admission to the ICU. Similarly, acute respiratory failure represents a significant source of strain on the United States’ healthcare system in terms of number of admissions and financial burden. ARF is generally treated with invasive mechanical ventilation or noninvasive respiratory support strategies. The efficacies of the various strategies, however, are not fully understood. ICUs, through detailed and continuously monitored patient data, provides an opportunity to retrospectively analyze decompensation, namely AHF and ARF, to evaluate physiological states, mechanisms of decompensation, and patient outcomes. There exist, however, significant challenges to large-scale observational analysis of such patient data. Challenges include data curation, data quality, consistency, and accuracy, and granularity of retrospective data required to extract meaningful information and identify usable representations within the data. In this work, we apply informatics methods to address these challenges in critical care medicine and ultimately, to accurately identify, characterize, and predict decompensation events to allow for prevention and mitigation during an ICU stay. The objectives of this work were to address knowledge gaps related to ventilation therapy strategies across diverse patient populations and to examine the prevalence of cardiovascular risk factors among those admitted to ICUs. There is a need for accurate therapy-based phenotyping for secondary analyses of electronic health record data to answer research questions regarding respiratory management of ARF patients and outcomes associated with different respiratory therapy strategies. Similarly, ICU data provides opportunity to evaluate combinations of clinical features that are predictive of decompensation events, such as onset of AHF and ARF requiring various methods of oxygen therapy using machine learning techniques. To accomplish these objectives, we developed an electronic phenotyping algorithm for identification of patients with ARF and cohort definitions for AHF patients with and without various heart failure risk factors. Our goal was to develop computable phenotypes and cohort definitions for both decompensation use cases for robustness and reproducibility. This approach permits analyses across broad patient populations of interest with the ability to sub-phenotype and address additional research questions. We then characterize clinical subpopulations of ARF and AHF patients for deeper analysis and apply machine learning methods (logistic regression and random forest) for feature selection of data inputs which might be indicative of decompensation. We use these same machine learning methods as baseline prediction models for comparison to more powerful deep learning models (recurrent neural networks) which incorporate granular time series data for prediction of decompensation events in critically ill patients. Lastly, we use remotely monitored ICU data to begin to evaluate remote, critical care telemedicine decision-making and discuss future opportunities for prediction model implementation. Our results show that meaningful information regarding physiological state of critically ill patients can be extracted from retrospective, observational data and used for detailed characterization of clinical sub-groups across diverse patient populations. In addition, data can be leveraged to predict which patients will experience decompensation during an ICU visit. These accurate predictions are also made early enough for clinicians to intervene or reevaluate treatment path trajectory if necessary. Remotely monitored critical care data also suggests that implementation of remote times, broadly, does not interrupt ICU operations. This suggests remote ICU may serve as an implementation platform for additional validation beyond the internal cross-validation and testing in this work.