Azra Bihorac

The past decade has seen an explosion in the amount of digital information stored in electronic health records (EHRs). While primarily designed for archiving patient information and performing administrative healthcare tasks like billing, many researchers have found secondary use of these records for various clinical informatics applications. Over the same period, the machine learning community has seen widespread advances in the field of deep learning. In this review, we survey the current research on applying deep learning to clinical tasks based on EHR data, where we find a variety of deep learning techniques and frameworks being applied to several types of clinical applications including information extraction, representation learning, outcome prediction, phenotyping, and deidentification. We identify several limitations of current research involving topics such as model interpretability, data heterogeneity, and lack of universal benchmarks. We conclude by summarizing the state of the field and identifying avenues of future deep EHR research.

Consensus definitions have been reached for both acute kidney injury (AKI) and chronic kidney disease (CKD) and these definitions are now routinely used in research and clinical practice. The KDIGO guideline defines AKI as an abrupt decrease in kidney function occurring over 7 days or less, whereas CKD is defined by the persistence of kidney disease for a period of >90 days. AKI and CKD are increasingly recognized as related entities and in some instances probably represent a continuum of the disease process. For patients in whom pathophysiologic processes are ongoing, the term acute kidney disease (AKD) has been proposed to define the course of disease after AKI; however, definitions of AKD and strategies for the management of patients with AKD are not currently available. In this consensus statement, the Acute Disease Quality Initiative (ADQI) proposes definitions, staging criteria for AKD, and strategies for the management of affected patients. We also make recommendations for areas of future research, which aim to improve understanding of the underlying processes and improve outcomes for patients with AKD.

INTRODUCTION: Acute kidney injury (AKI) can evolve quickly and clinical measures of function often fail to detect AKI at a time when interventions are likely to provide benefit. Identifying early markers of kidney damage has been difficult due to the complex nature of human AKI, in which multiple etiologies exist. The objective of this study was to identify and validate novel biomarkers of AKI. METHODS: We performed two multicenter observational studies in critically ill patients at risk for AKI - discovery and validation. The top two markers from discovery were validated in a second study (Sapphire) and compared to a number of previously described biomarkers. In the discovery phase, we enrolled 522 adults in three distinct cohorts including patients with sepsis, shock, major surgery, and trauma and examined over 300 markers. In the Sapphire validation study, we enrolled 744 adult subjects with critical illness and without evidence of AKI at enrollment; the final analysis cohort was a heterogeneous sample of 728 critically ill patients. The primary endpoint was moderate to severe AKI (KDIGO stage 2 to 3) within 12 hours of sample collection. RESULTS: Moderate to severe AKI occurred in 14% of Sapphire subjects. The two top biomarkers from discovery were validated. Urine insulin-like growth factor-binding protein 7 (IGFBP7) and tissue inhibitor of metalloproteinases-2 (TIMP-2), both inducers of G1 cell cycle arrest, a key mechanism implicated in AKI, together demonstrated an AUC of 0.80 (0.76 and 0.79 alone). Urine [TIMP-2]·[IGFBP7] was significantly superior to all previously described markers of AKI (P <0.002), none of which achieved an AUC >0.72. Furthermore, [TIMP-2]·[IGFBP7] significantly improved risk stratification when added to a nine-variable clinical model when analyzed using Cox proportional hazards model, generalized estimating equation, integrated discrimination improvement or net reclassification improvement. Finally, in sensitivity analyses [TIMP-2]·[IGFBP7] remained significant and superior to all other markers regardless of changes in reference creatinine method. CONCLUSIONS: Two novel markers for AKI have been identified and validated in independent multicenter cohorts. Both markers are superior to existing markers, provide additional information over clinical variables and add mechanistic insight into AKI. TRIAL REGISTRATION: ClinicalTrials.gov number NCT01209169.

BACKGROUND: Long-term survival after acute kidney injury (AKI) is poorly studied. We report the relationship between long-term mortality and AKI with small changes in serum creatinine during hospitalization after various cardiothoracic surgery procedures. METHODS AND RESULTS: This was a retrospective study of 2973 patients with no history of chronic kidney disease who were discharged from the hospital after cardiothoracic surgery between 1992 and 2002. AKI was defined by the RIFLE classification (Risk, Injury, Failure, Loss, and End stage), which requires at least a 50% increase in serum creatinine and stratifies patients into 3 grades of AKI: Risk, injury, and failure. Patient survival was determined through the National Social Security Death Index. Long-term survival was analyzed with a risk-adjusted Cox proportional hazards regression model. Survival was worse among patients with AKI and was proportional to its severity, with an adjusted hazard ratio of 1.23 (95% CI 1.06 to 1.42) for the least severe RIFLE risk class and 2.14 (95% CI 1.73 to 2.66) for the RIFLE failure class compared with patients without AKI. Survival was worse among all subgroups of cardiothoracic surgery with AKI except for valve surgery. Patients with complete renal recovery after AKI still had an increased adjusted hazard ratio for death of 1.28 (95% CI 1.11 to 1.48) compared with patients without AKI. CONCLUSIONS: The risk of death associated with AKI after cardiothoracic surgery remains high for 10 years regardless of other risk factors, even for those patients with complete renal recovery. Improved renal protection and closer postdischarge follow-up of renal function may be warranted.

Surgical intensive care unit (ICU) stay of longer than 10 days is often described by the experienced intensivist as a "complicated clinical course" and is frequently attributed to persistent immune dysfunction. "Systemic inflammatory response syndrome" (SIRS) followed by "compensatory anti-inflammatory response syndrome" (CARS) is a conceptual framework to explain the immunologic trajectory that ICU patients with severe sepsis, trauma, or emergency surgery for abdominal infection often traverse, but the causes, mechanisms, and reasons for persistent immune dysfunction remain unexplained. Often involving multiple-organ failure (MOF) and death, improvements in surgical intensive care have altered its incidence, phenotype, and frequency and have increased the number of patients who survive initial sepsis or surgical events and progress to a persistent inflammation, immunosuppression, and catabolism syndrome (PICS). Often observed, but rarely reversible, these patients may survive to transfer to a long-term care facility only to return to the ICU, but rarely to self-sufficiency. We propose that PICS is the dominant pathophysiology and phenotype that has replaced late MOF and prolongs surgical ICU stay, usually with poor outcome. This review details the evolving epidemiology of MOF, the clinical presentation of PICS, and our understanding of how persistent inflammation and immunosuppression define the pathobiology of prolonged intensive care. Therapy for PICS will involve innovative interventions for immune system rebalance and nutritional support to regain physical function and well-being.