More than 60% of patients hospitalized with COVID-19 who underwent transthoracic echocardiography (TTE) for suspected cardiac involvement had myocardial injury based on cardiac troponin (cTn) levels being greater than the upper reference limit for the assay in use at that institution (J Am Coll Cardiol 2020;76:2043-55).
The overall in-hospital mortality rate was 5.2% for all patients who were part of this multicenter cohort study at seven hospitals in New York City and Milan, Italy. Elevated cTn levels alone were associated with an increased mortality rate (18.6%), but patients who had both elevated cTn levels and abnormalities found by TTE fared the worst, with a 31.7% in-hospital mortality rate.
Univariate odds ratios of death, acute kidney injury, shock, or ventricular arrhythmia were 6.67, 6.13, 4.40, and 3.72, respectively for those with cTn levels indicative of myocardial injury in comparison to those without elevated levels.
Patients with myocardial injury also had more electrocardiographic abnormalities, higher inflammatory biomarker levels, and increased prevalence of major echocardiographic abnormalities. Median values of analytes in patients with myocardial injury that were notably elevated in comparison to levels in patients without myocardial injury were interleukin-6 (116 pg/mL versus 58 pg/mL), lactate dehydrogenase (762 U/L versus 445 U/L), ferritin (1,624 ng/mL versus 701 ng/mL), D-dimer (3.7 µg/mL versus 1.5 µg/mL), and procalcitonin (1.3 ng/mL versus 0.2 ng/mL).
Major types of abnormalities detected by TTE included right ventricular dysfunction (26.3%), left ventricular global dysfunction (18.4%), diastolic dysfunction grade II or III (13.2%), and left ventricular wall motion abnormalities (23.7%). In 7.2% of these cases, patients also had pericardial effusion abnormalities.
American College of Cardiology (ACC) guidance recommends measuring cTn levels in patients with SARS-CoV-2 infection who also are being evaluated for acute myocardial infarction. However, this guidance “seems somehow inadequate” in light of this study, according to an accompanying editorial (J Am Coll Cardiol 2020;76:2056-9). The editorialists added that the ACC criteria for assessing cTn would be expanded to include all patients with COVID-19, “not only those with a clinical suspicion of cardiac ischemia.”
Best Practice Alert Successfully Curbs Antibiotic Usage for Lower Respiratory Tract Infections
An electronic medical record (EMR) best practice alert (BPA) based on procalcitonin (PCT) and polymerase chain reaction (PCR) results reduced inpatient antibiotic days by 2.2 days in patients with lower respiratory tract infections (Clin Infect Dis 2020;71:1684-9). These results demonstrate that “well-constructed EMR provider alerts that integrate PCR, PCT, and antibiotic data can target patients in whom antibiotic therapy can be rapidly narrowed, without need for direct antimicrobial stewardship oversight,” according to the investigators. They went on to suggest that this minimally invasive stewardship practice could be emulated “easily” by other organizations.
The researchers sought to determine whether an automated antimicrobial stewardship provider alert would lower antibiotic use. The BPA alert would be activated for adult patients who met three criteria within 48 hours of each other, including PCT results ≤0.25 ng/mL, virus detected via PCR from respiratory specimens, and active use of systemic antibiotics. The study took place at five hospitals in the Saint Luke’s Health System in Kansas City, Missouri. The primary outcome was inpatient antibiotic days of therapy.
Whenever these criteria were met, the BPA stated “Antimicrobial Stewardship Alert: Your patient has a positive viral PCR + negative procalcitonin + one or more antibiotics ordered. These results suggest viral infection – please reassess necessity of antibiotics as indicated.” Providers in response to the BPA had three choices, including “acknowledge,” “does not meet criteria,” and “not making antimicrobial decisions.” The two former alerts suppressed the alert permanently, while the latter kept firing the alert each time a provider accessed the EMR until he or she chose one of the other options.
The study involved pre-post evaluation of antibiotic use in the proscribed patient population. The authors found that the BPA not only reduced inpatient antibiotic days by a mean of 2.2 days but also the percentage of patients prescribed antibiotics on discharge (20% versus 47%).
Machine Learning Models Accurately Predict Familial Hypercholesterolemia, but Clinical Utility Differs Substantially
An analysis of five common machine learning (ML) algorithms found that four show high accuracy in predicting familial hypercholesterolemia (FH) but that the clinical case finding workload to yield cases would vary substantially between models (NPJ Digit Med 2020;3:142).
This retrospective cohort study involved routine primary care records of more than 4 million individuals in the U.K. who had a recorded cholesterol measurement. Of these, 7,928 had diagnosed FH. The authors randomly split the study population into a training cohort comprising 75% of patients and a validation cohort of the remaining 25%. They used the training cohort to derive the FH algorithms and the validation cohort to apply and test the algorithms.
The four high-performing ML algorithms were deep learning model (DLM), gradient boosting model, random forest model, and ensemble learning model (ELM). All of these yielded areas under the receiver operating characteristic curve (AUC) >0.89. The logistic regression model was the poor performer with AUC >0.81.
The models incorporated 45 predictor variables for FH derived from known associations between these features and FH, as detailed in the scientific literature, recommended diagnostic criteria, previously developed algorithms, and expert clinical opinions.
Sensitivity of the models varied considerably, from 30.5% in ELM to 72.6% for DLM. Specificity was more uniform, varying from 90% for DLM to 99.3% for ELM. Positive predictive values ranged from 2.8% for DLM to 15.5% for ELM; negative predictive values were nearly uniform, ranging from 99.7% to 99.9%. Assuming FH prevalence of 1 per 250 population, DLM would identify about 10% of the population as probable FH, whereas ELM would identify just 0.73%.