Kidney Outcomes and Trajectories of Tubular Injury and Function in Critically Ill Persons with and without Coronavirus-2019 (preprint)

Background Coronavirus disease-2019 (COVID-19) may injure the kidney tubules via activation of inflammatory host responses and/or direct viral infiltration. Most studies of kidney injury in COVID-19 lacked contemporaneous controls or measured kidney biomarkers at a single time point. To better understand mechanisms of AKI in COVID-19, we compared kidney outcomes and trajectories of tubular injury, viability, and function in prospectively enrolled critically ill adults with and without COVID-19.Methods The COVID-19 Host Response and Outcomes (CHROME) study prospectively enrolled patients admitted to intensive care units in Washington state with symptoms of lower respiratory tract infection, determining COVID-19 status by nucleic acid amplification on arrival. We evaluated major adverse kidney events (MAKE) defined as a doubling of serum creatinine, kidney replacement therapy, or death, in 330 patients after inverse probability weighting. In the 181 patients with available biosamples, we determined trajectories of urine kidney injury molecule-1 (KIM-1) and epithelial growth factor (EGF), and urine:plasma ratios of endogenous markers of tubular secretory clearance.Results At ICU admission, mean age was 55\(\pm\)16 years; 45% required mechanical ventilation; and mean serum creatinine concentration was 1.1 mg/dL. COVID-19 was associated with a 70% greater incidence of MAKE (95% CI 1.05, 2.74) and a 741% greater incidence of KRT (95% CI 1.69, 32.41). The biomarker cohort had a median of three follow-up measurements. Urine EGF, secretory clearance ratios, and eGFR increased over time in the COVID-19 negative group but remained unchanged in the COVID-19 positive group. In contrast, urine KIM-1 concentrations did not significantly change over the course of the study in either group.Conclusions Among critically ill adults, COVID-19 is associated with a more protracted course of proximal tubular dysfunction.

Mediators of monocyte chemotaxis and matrix remodeling are associated with the development of fibrosis in patients with COVID-19 (preprint)

Acute respiratory distress syndrome (ARDS) has a fibroproliferative phase that may be followed by pulmonary fibrosis. This has been described in patients with COVID-19 pneumonia, but the underlying mechanisms have not been completely defined. We hypothesized that protein mediators of tissue remodeling and monocyte chemotaxis are elevated in the plasma and endotracheal aspirates of critically ill patients with COVID-19 who subsequently develop radiographic fibrosis. We enrolled COVID-19 patients admitted to the ICU who had hypoxemic respiratory failure, were hospitalized and alive for at least 10 days, and had chest imaging done during hospitalization (n = 119). Plasma was collected within 24h of ICU admission and at 7d. In mechanically ventilated patients, endotracheal aspirates (ETA) were collected at 24h and 48-96h. Protein concentrations were measured by immunoassay. We tested for associations between protein concentrations and radiographic evidence of fibrosis using logistic regression adjusting for age, sex, and APACHE score. We identified 39 patients (33%) with features of fibrosis. Within 24h of ICU admission, plasma proteins related to tissue remodeling (MMP-9, Amphiregulin) and monocyte chemotaxis (CCL-2/MCP-1, CCL-13/MCP-4) were associated with the subsequent development of fibrosis whereas markers of inflammation (IL-6, TNF-) were not. After 1 week, plasma MMP-9 increased in patients without fibrosis. In ETAs, only CCL-2/MCP-1 was associated with fibrosis at the later timepoint. This cohort study identifies proteins of tissue remodeling and monocyte recruitment that may identify early fibrotic remodeling following COVID-19. Measuring changes in these proteins over time may allow for early detection of fibrosis in patients with COVID-19.

Identifying ARDS using the Hierarchical Attention Network with Sentence Objectives Framework (preprint)

Acute respiratory distress syndrome (ARDS) is a life-threatening condition that is often undiagnosed or diagnosed late. ARDS is especially prominent in those infected with COVID-19. We explore the automatic identification of ARDS indicators and confounding factors in free-text chest radiograph reports. We present a new annotated corpus of chest radiograph reports and introduce the Hierarchical Attention Network with Sentence Objectives (HANSO) text classification framework. HANSO utilizes fine-grained annotations to improve document classification performance. HANSO can extract ARDS-related information with high performance by leveraging relation annotations, even if the annotated spans are noisy. Using annotated chest radiograph images as a gold standard, HANSO identifies bilateral infiltrates, an indicator of ARDS, in chest radiograph reports with performance (0.87 F1) comparable to human annotations (0.84 F1). This algorithm could facilitate more efficient and expeditious identification of ARDS by clinicians and researchers and contribute to the development of new therapies to improve patient care.

Deep learning models for COVID-19 chest x-ray classification: Preventing shortcut learning using feature disentanglement (preprint)

In response to the COVID-19 global pandemic, recent research has proposed creating deep learning based models that use chest radiographs (CXRs) in a variety of clinical tasks to help manage the crisis. However, the size of existing datasets of CXRs from COVID-19+ patients are relatively small, and researchers often pool CXR data from multiple sources, for example, using different x-ray machines in various patient populations under different clinical scenarios. Deep learning models trained on such datasets have been shown to overfit to erroneous features instead of learning pulmonary characteristics - a phenomenon known as shortcut learning. We propose adding feature disentanglement to the training process, forcing the models to identify pulmonary features from the images while penalizing them for learning features that can discriminate between the original datasets that the images come from. We find that models trained in this way indeed have better generalization performance on unseen data; in the best case we found that it improved AUC by 0.13 on held out data. We further find that this outperforms masking out non-lung parts of the CXRs and performing histogram equalization, both of which are recently proposed methods for removing biases in CXR datasets.