Performance of existing and novel surveillance case definitions for COVID-19 in household contacts of PCR-confirmed COVID-19.

BACKGROUND: Optimized symptom-based COVID-19 case definitions that guide public health surveillance and individual patient management in the community may assist pandemic control. METHODS: We assessed diagnostic performance of existing cases definitions (e.g. influenza-like illness, COVID-like illness) using symptoms reported from 185 household contacts to a PCR-confirmed case of COVID-19 in Wisconsin and Utah, United States. We stratified analyses between adults and children. We also constructed novel case definitions for comparison. RESULTS: Existing COVID-19 case definitions generally showed high sensitivity (86-96%) but low positive predictive value (PPV) (36-49%; F-1 score 52-63) in this community cohort. Top performing novel symptom combinations included taste or smell dysfunction and improved the balance of sensitivity and PPV (F-1 score 78-80). Performance indicators were generally lower for children (< 18 years of age). CONCLUSIONS: Existing COVID-19 case definitions appropriately screened in household contacts with COVID-19. Novel symptom combinations incorporating taste or smell dysfunction as a primary component improved accuracy. Case definitions tailored for children versus adults should be further explored.

Opioid Prescribing Rates in Nonmetropolitan and Metropolitan Counties Among Primary Care Providers Using an Electronic Health Record System - United States, 2014-2017.

Drug overdose is the leading cause of unintentional injury-associated death in the United States. Among 70,237 fatal drug overdoses in 2017, prescription opioids were involved in 17,029 (24.2%) (1). Higher rates of opioid-related deaths have been recorded in nonmetropolitan (rural) areas (2). In 2017, 14 rural counties were among the 15 counties with the highest opioid prescribing rates.* Higher opioid prescribing rates put patients at risk for addiction and overdose (3). Using deidentified data from the Athenahealth electronic health record (EHR) system, opioid prescribing rates among 31,422 primary care providers† in the United States were analyzed to evaluate trends from January 2014 to March 2017. This analysis assessed how prescribing practices varied among six urban-rural classification categories of counties, before and after the March 2016 release of CDC's Guideline for Prescribing Opioids for Chronic Pain (Guideline) (4). Patients in noncore (the most rural) counties had an 87% higher chance of receiving an opioid prescription compared with persons in large central metropolitan counties during the study period. Across all six county groups, the odds of receiving an opioid prescription decreased significantly after March 2016. This decrease followed a flat trend during the preceding period in micropolitan and large central metropolitan county groups; in contrast, the decrease continued previous downward trends in the other four county groups. Data from EHRs can effectively supplement traditional surveillance methods for monitoring trends in opioid prescribing and other areas of public health importance, with minimal lag time under ideal conditions. As less densely populated areas appear to indicate both substantial progress in decreasing opioid prescribing and ongoing need for reduction, community health care practices and intervention programs must continue to be tailored to community characteristics.

Chief complaint classification with recurrent neural networks.

Syndromic surveillance detects and monitors individual and population health indicators through sources such as emergency department records. Automated classification of these records can improve outbreak detection speed and diagnosis accuracy. Current syndromic systems rely on hand-coded keyword-based methods to parse written fields and may benefit from the use of modern supervised-learning classifier models. In this paper, we implement two recurrent neural network models based on long short-term memory (LSTM) and gated recurrent unit (GRU) cells and compare them to two traditional bag-of-words classifiers: multinomial naïve Bayes (MNB) and a support vector machine (SVM). The MNB classifier is one of only two machine learning algorithms currently being used for syndromic surveillance. All four models are trained to predict diagnostic code groups as defined by Clinical Classification Software, first to predict from discharge diagnosis, and then from chief complaint fields. The classifiers are trained on 3.6 million de-identified emergency department records from a single United States jurisdiction. We compare performance of these models primarily using the F1 score, and we measure absolute model performance to determine which conditions are the most amenable to surveillance based on chief complaint alone. Using discharge diagnoses, the LSTM classifier performs best, though all models exhibit an F1 score above 96.00. Using chief complaints, the GRU performs best (F1 = 47.38), and MNB with bigrams performs worst (F1 = 39.40). We also note that certain syndrome types are easier to detect than others. For example, chief complaints using the GRU model predicts alcohol-related disorders well (F1 = 78.91) but predicts influenza poorly (F1 = 14.80). In all instances, the RNN models outperformed the bag-of-words classifiers suggesting deep learning models could substantially improve the automatic classification of unstructured text for syndromic surveillance.

Declines in Opioid Prescribing After a Private Insurer Policy Change - Massachusetts, 2011-2015.

Overdose deaths involving opioid pain medications are epidemic in the United States, in part because of high opioid prescribing rates and associated abuse of these drugs (1). In 2014, nearly 2 million U.S. residents either abused or were dependent on prescription opioids (2). In Massachusetts, unintentional opioid-related overdose deaths, including deaths involving heroin, increased 45% from 2012 to 2013.* In 2014, the rate of these deaths reached 20.0 per 100,000, nearly 2.5 times higher than the U.S. rate overall (3,4). On July 1, 2012, Blue Cross Blue Shield of Massachusetts (BCBSMA), the largest insurer in the state with approximately 2.8 million members,† implemented a comprehensive opioid utilization program after learning that many of its members were receiving new prescriptions with a >30-day supply of opioids. The 2016 CDC Guideline for Prescribing Opioids for Chronic Pain recommends avoiding opioids as a first-line therapy for chronic pain and limiting quantities when initiating opioids for acute pain (5). CDC analyzed BCBSMA prescription claims data for the period 2011-2015 to assess the effect of the new utilization program on opioid prescribing rates. During the first 3 years after policy implementation, the average monthly prescribing rate for opioids decreased almost 15%, from 34 per 1,000 members to 29. The percentage of BCBSMA members per month with current opioid prescriptions also declined. The temporal association between implementation of the program and statistically significant declines in both prescribing rates and proportion of members using opioids suggests that the BCBSMA initiative played a role in reducing the use of prescription opioids among its members. Public and private insurers in the United States could benefit from developing their own best practices for prescription opioid utilization that ensure accessible pain care, while reducing the risk for dependence and abuse associated with these drugs.

Lower yield of endoscopic ultrasound-guided fine-needle aspiration in patients with pancreatic head mass with a biliary stent.

BACKGROUND: In jaundiced patients with suspected pancreatic cancer, endoscopic retrograde cholangiopancreatography (ERCP) with biliary stent is frequently performed prior to histologic diagnosis by endoscopic ultrasound-guided fine-needle aspiration (EUS-FNA). AIMS: To compare the accuracy of jaundiced patients undergoing EUS-FNA for a pancreatic mass with or without a biliary stent. METHODS: Consecutive patients with a pancreatic mass and jaundice who underwent EUS-FNA between 4/2005 and 4/2013 were identified. Final diagnosis was based on histologic or radiologic evaluation with >6-month follow-up if the index EUS-FNA was negative for malignancy. Primary endpoint was histologic accuracy of EUS-FNA. RESULTS: Mean age of the 180 patients was 65 ± 12 years; 83 (46 %) had ERCP prior to EUS-FNA and 75 (42 %) received a biliary stent. Onsite cytopathologist was present for 81 (45 %) procedures. Final diagnosis revealed malignancy in 172 (96 %) patients, with adenocarcinoma in 159 (88 %). Patients with biliary stents had lower accuracy of EUS-FNA for malignancy than those without a biliary stent: 77 % (95 % CI 67-85 %) versus 89 % (95 % CI 81-93 %). On multivariate analysis, having a biliary stent (OR = 0.37, 95 % CI 0.15-0.90), onsite cytopathologist (OR = 9.24, 95 % CI 2.64-32.37), and receiving a core biopsy (OR = 2.60, 95 % CI 1.07-6.29) were associated with accuracy of EUS-FNA. CONCLUSIONS: Presence of a biliary stent was associated with a significant decrease in the accuracy of EUS-FNA for histologic diagnosis of pancreatic cancer, while accuracy was increased when a cytopathologist was onsite. EUS-FNA should be performed prior to ERCP in jaundiced patients with suspected pancreatic cancer.

Development and validation of a deep learning model for detecting signs of tuberculosis on chest radiographs among US-bound immigrants and refugees.

Immigrants and refugees seeking admission to the United States must first undergo an overseas medical exam, overseen by the US Centers for Disease Control and Prevention (CDC), during which all persons ≥15 years old receive a chest x-ray to look for signs of tuberculosis. Although individual screening sites often implement quality control (QC) programs to ensure radiographs are interpreted correctly, the CDC does not currently have a method for conducting similar QC reviews at scale. We obtained digitized chest radiographs collected as part of the overseas immigration medical exam. Using radiographs from applicants 15 years old and older, we trained deep learning models to perform three tasks: identifying abnormal radiographs; identifying abnormal radiographs suggestive of tuberculosis; and identifying the specific findings (e.g., cavities or infiltrates) in abnormal radiographs. We then evaluated the models on both internal and external testing datasets, focusing on two classes of performance metrics: individual-level metrics, like sensitivity and specificity, and sample-level metrics, like accuracy in predicting the prevalence of abnormal radiographs. A total of 152,012 images (one image per applicant; mean applicant age 39 years) were used for model training. On our internal test dataset, our models performed well both in identifying abnormalities suggestive of TB (area under the curve [AUC] of 0.97; 95% confidence interval [CI]: 0.95, 0.98) and in estimating sample-level counts of the same (-2% absolute percentage error; 95% CIC: -8%, 6%). On the external test datasets, our models performed similarly well in identifying both generic abnormalities (AUCs ranging from 0.89 to 0.92) and those suggestive of TB (AUCs from 0.94 to 0.99). This performance was consistent across metrics, including those based on thresholded class predictions, like sensitivity, specificity, and F1 score. Strong performance relative to high-quality radiological reference standards across a variety of datasets suggests our models may make reliable tools for supporting chest radiography QC activities at CDC.

Sensitive and Feasible Specimen Collection and Testing Strategies for Diagnosing Tuberculosis in Young Children.

Importance: Criterion-standard specimens for tuberculosis diagnosis in young children, gastric aspirate (GA) and induced sputum, are invasive and rarely collected in resource-limited settings. A far less invasive approach to tuberculosis diagnostic testing in children younger than 5 years as sensitive as current reference standards is important to identify. Objective: To characterize the sensitivity of preferably minimally invasive specimen and assay combinations relative to maximum observed yield from all specimens and assays combined. Design, Setting, and Participants: In this prospective cross-sectional diagnostic study, the reference standard was a panel of up to 2 samples of each of 6 specimen types tested for Mycobacterium tuberculosis complex by Xpert MTB/RIF assay and mycobacteria growth indicator tube culture. Multiple different combinations of specimens and tests were evaluated as index tests. A consecutive series of children was recruited from inpatient and outpatient settings in Kisumu County, Kenya, between October 2013 and August 2015. Participants were children younger than 5 years who had symptoms of tuberculosis (unexplained cough, fever, malnutrition) and parenchymal abnormality on chest radiography or who had cervical lymphadenopathy. Children with 1 or more evaluable specimen for 4 or more primary study specimen types were included in the analysis. Data were analyzed from February 2015 to October 2020. Main Outcomes and Measures: Cumulative and incremental diagnostic yield of combinations of specimen types and tests relative to the maximum observed yield. Results: Of the 300 enrolled children, the median (interquartile range) age was 2.0 (1.0-3.6) years, and 151 (50.3%) were female. A total of 294 met criteria for analysis. Of 31 participants with confirmed tuberculosis (maximum observed yield), 24 (sensitivity, 77%; interdecile range, 68%-87%) had positive results on up to 2 GA samples and 20 (sensitivity, 64%; interdecile range, 53%-76%) had positive test results on up to 2 induced sputum samples. The yields of 2 nasopharyngeal aspirate (NPA) samples (23 of 31 [sensitivity, 74%; interdecile range, 64%-84%]), of 1 NPA sample and 1 stool sample (22 of 31 [sensitivity, 71%; interdecile range, 60%-81%]), or of 1 NPA sample and 1 urine sample (21.5 of 31 [sensitivity, 69%; interdecile range, 58%-80%]) were similar to reference-standard specimens. Combining up to 2 each of GA and NPA samples had an average yield of 90% (28 of 31). Conclusions and Relevance: NPA, in duplicate or in combination with stool or urine specimens, was readily obtainable and had diagnostic yield comparable with reference-standard specimens. This combination could improve tuberculosis diagnosis among children in resource-limited settings. Combining GA and NPA had greater yield than that of the current reference standards and may be useful in certain clinical and research settings.

A comparison of machine learning algorithms for the surveillance of autism spectrum disorder.

OBJECTIVE: The Centers for Disease Control and Prevention (CDC) coordinates a labor-intensive process to measure the prevalence of autism spectrum disorder (ASD) among children in the United States. Random forests methods have shown promise in speeding up this process, but they lag behind human classification accuracy by about 5%. We explore whether more recently available document classification algorithms can close this gap. MATERIALS AND METHODS: Using data gathered from a single surveillance site, we applied 8 supervised learning algorithms to predict whether children meet the case definition for ASD based solely on the words in their evaluations. We compared the algorithms' performance across 10 random train-test splits of the data, using classification accuracy, F1 score, and number of positive calls to evaluate their potential use for surveillance. RESULTS: Across the 10 train-test cycles, the random forest and support vector machine with Naive Bayes features (NB-SVM) each achieved slightly more than 87% mean accuracy. The NB-SVM produced significantly more false negatives than false positives (P = 0.027), but the random forest did not, making its prevalence estimates very close to the true prevalence in the data. The best-performing neural network performed similarly to the random forest on both measures. DISCUSSION: The random forest performed as well as more recently available models like the NB-SVM and the neural network, and it also produced good prevalence estimates. NB-SVM may not be a good candidate for use in a fully-automated surveillance workflow due to increased false negatives. More sophisticated algorithms, like hierarchical convolutional neural networks, may not be feasible to train due to characteristics of the data. Current algorithms might perform better if the data are abstracted and processed differently and if they take into account information about the children in addition to their evaluations. CONCLUSION: Deep learning models performed similarly to traditional machine learning methods at predicting the clinician-assigned case status for CDC's autism surveillance system. While deep learning methods had limited benefit in this task, they may have applications in other surveillance systems.

Natural language generation for electronic health records.

One broad goal of biomedical informatics is to generate fully-synthetic, faithfully representative electronic health records (EHRs) to facilitate data sharing between healthcare providers and researchers and promote methodological research. A variety of methods existing for generating synthetic EHRs, but they are not capable of generating unstructured text, like emergency department (ED) chief complaints, history of present illness, or progress notes. Here, we use the encoder-decoder model, a deep learning algorithm that features in many contemporary machine translation systems, to generate synthetic chief complaints from discrete variables in EHRs, like age group, gender, and discharge diagnosis. After being trained end-to-end on authentic records, the model can generate realistic chief complaint text that appears to preserve the epidemiological information encoded in the original record-sentence pairs. As a side effect of the model's optimization goal, these synthetic chief complaints are also free of relatively uncommon abbreviation and misspellings, and they include none of the personally identifiable information (PII) that was in the training data, suggesting that this model may be used to support the de-identification of text in EHRs. When combined with algorithms like generative adversarial networks (GANs), our model could be used to generate fully-synthetic EHRs, allowing healthcare providers to share faithful representations of multimodal medical data without compromising patient privacy. This is an important advance that we hope will facilitate the development of machine-learning methods for clinical decision support, disease surveillance, and other data-hungry applications in biomedical informatics.

Oral Mucosa Harbors a High Frequency of Endothelial Cells: A Novel Postnatal Cell Source for Angiogenic Regeneration.

Endothelial progenitor cells/endothelial cells (EPCs/ECs) have great potential to treat pathological conditions such as cardiac infarction, muscle ischemia, and bone fractures, but isolation of EPC/ECs from existing cell sources is challenging due to their low EC frequency. We have isolated endothelial progenitor (EP)-like cells from rat oral mucosa and characterized their yield, immunophenotype, growth, and in vivo angiogenic potential. The frequency of EP-like cells derived from oral mucosa is thousands of folds higher than EPCs derived from donor-match bone marrow samples. EP-like cells from oral mucosa were positive for EC markers CD31, VE-Cadherin, and VEGFR2. Oral mucosa-derived EP-like cells displayed robust uptake of acetylated low-density lipoprotein and formed stable capillary networks in Matrigel. Subcutaneously implanted oral mucosa-derived EP-like cells anastomosed with host blood vessels, implicating their ability to elicit angiogenesis. Similar to endothelial colony-forming cells, EP-like cells from oral mucosa have a significantly higher proliferative rate than human umbilical vein endothelial cells. These findings identify a putative EPC source that is easily accessible in the oral cavity, potentially from discarded tissue specimens, and yet with robust yield and potency for angiogenesis in tissue and organ regeneration.

IFN-γ Release Assay Result Is Associated with Disease Site and Death in Active Tuberculosis.

RATIONALE: The IFN-γ release assays and tuberculin skin tests are used to support the diagnosis of both latent and active tuberculosis. However, we previously demonstrated that a negative tuberculin test in active tuberculosis is associated with disseminated disease and death. It is unknown whether the same associations exist for IFN-γ release assays. OBJECTIVES: To determine the association between these tests and site of tuberculosis and death among persons with active tuberculosis. METHODS: We analyzed IFN-γ release assays and tuberculin test results for all persons with culture-confirmed tuberculosis reported to the U.S. National Tuberculosis Surveillance System from 2010 to 2014. We used logistic regression to calculate the association between these tests and site of disease and death. MEASUREMENTS AND MAIN RESULTS: A total of 24,803 persons with culture-confirmed tuberculosis had either of these test results available for analysis. Persons with a positive tuberculin test had lower odds of disseminated disease (i.e., miliary or combined pulmonary and extrapulmonary disease), but there was no difference in the odds of disseminated disease with a positive IFN-γ release assay. However, persons who were positive to either of these tests had lower odds of death. An indeterminate IFN-γ release assay result was associated with greater odds of both disseminated disease and death. CONCLUSIONS: Despite perceived equivalence in clinical practice, IFN-γ release assays and tuberculin test results have different associations with tuberculosis site, yet similar associations with the risk of death. Furthermore, an indeterminate IFN-γ release assay result in a person with active tuberculosis is not unimportant, and rather carries greater odds of disseminated disease and death. Prospective study may improve our understanding of the underlying mechanisms by which these tests are associated with disease localization and death.