Search

Background Host gene expression has emerged as a complementary strategy to pathogen detection tests for the discrimination of bacterial and viral infection. The impact of immunocompromise on host-response tests remains unknown. We evaluated a host-response test discriminating bacterial, viral, and noninfectious conditions in immunocompromised subjects. Methods An 81-gene signature was measured using real-time–polymerase chain reaction in subjects with immunocompromise (chemotherapy, solid-organ transplant, immunomodulatory agents, AIDS) with bacterial infection, viral infection, or noninfectious illness. A regularized logistic regression model trained in immunocompetent subjects was used to estimate the likelihood of each class in immunocompromised subjects. Results Accuracy in the 136-subject immunocompetent training cohort was 84.6% for bacterial versus nonbacterial discrimination and 80.8% for viral versus nonviral discrimination. Model validation in 134 immunocompromised subjects showed overall accuracy of 73.9% for bacterial infection (P = .04 relative to immunocompetent subjects) and 75.4% for viral infection (P = .30). A scheme reporting results by quartile improved test utility. The highest probability quartile ruled-in bacterial and viral infection with 91.4% and 84.0% specificity, respectively. The lowest probability quartile ruled-out infection with 90.1% and 96.4% sensitivity for bacterial and viral infection, respectively. Performance was independent of the type or number of immunocompromising conditions. Conclusions A host gene expression test discriminated bacterial, viral, and noninfectious etiologies at a lower overall accuracy in immunocompromised patients compared with immunocompetent patients, although this difference was only significant for bacterial infection classification. With modified interpretive criteria, a host-response strategy may offer clinically useful diagnostic information for patients with immunocompromise.

Children with autism spectrum disorder (ASD) or attention deficit hyperactivity disorder (ADHD) have 2–3 times increased healthcare utilization and annual costs once diagnosed, but little is known about their utilization patterns early in life. Quantifying their early health system utilization could uncover condition-specific health trajectories to facilitate earlier detection and intervention. Patients born 10/1/2006–10/1/2016 with ≥ 2 well-child visits within the Duke University Health System before age 1 were grouped as ASD, ADHD, ASD + ADHD, or No Diagnosis using retrospective billing codes. An additional comparison group was defined by later upper respiratory infection diagnosis. Adjusted odds ratios (AOR) for hospital admissions, procedures, emergency department (ED) visits, and outpatient clinic encounters before age 1 were compared between groups via logistic regression models. Length of hospital encounters were compared between groups via Mann–Whitney U test. In total, 29,929 patients met study criteria (ASD N = 343; ADHD N = 1175; ASD + ADHD N = 140). ASD was associated with increased procedures (AOR = 1.5, p p p p p p p p p p p p = 0.003). Each condition was associated with increased health system utilization and distinctive patterns of utilization before age 1. Recognizing these patterns may contribute to earlier detection and intervention.

Background Monogenic diseases are individually rare but collectively common, and are likely underdiagnosed. Objectives The purpose of this study was to estimate the prevalence of monogenic cardiovascular diseases (MCVDs) and potentially missed diagnoses in a cardiovascular cohort. Methods Exomes from 8,574 individuals referred for cardiac catheterization were analyzed. Pathogenic/likely pathogenic (P/LP) variants associated with MCVD (cardiomyopathies, arrhythmias, connective tissue disorders, and familial hypercholesterolemia were identified. Electronic health records (EHRs) were reviewed for individuals harboring P/LP variants who were predicted to develop disease (G+). G+ individuals who did not have a documented relevant diagnosis were classified into groups of whether they may represent missed diagnoses (unknown, unlikely, possible, probable, or definite) based on relevant diagnostic criteria/features for that disease. Results In total, 159 P/LP variants were identified; 2,361 individuals harbored at least 1 P/LP variant, of whom 389 G+ individuals (4.5% of total cohort) were predicted to have at least 1 MCVD. EHR review of 342 G+ individuals predicted to have 1 MCVD with sufficient EHR data revealed that 52 had been given the relevant clinical diagnosis. The remaining 290 individuals were classified as potentially having an MCVD as follows: 193 unlikely (66.6%), 50 possible (17.2%), 30 probable (10.3%), and 17 definite (5.9%). Grouping possible, probable, definite, and known diagnoses, 149 were considered to have an MCVD. Novel MCVD pathogenic variants were identified in 16 individuals. Conclusions Overall, 149 individuals (1.7% of cohort) had MCVDs, but only 35% were diagnosed. These patients represents a “missed opportunity,” which could be addressed by greater use of genetic testing of patients seen by cardiologists.

Reinforcement learning (RL) has been widely used to aid training in language generation. This is achieved by enhancing standard maximum likelihood objectives with user-specified reward functions that encourage global semantic consistency. We propose a principled approach to address the difficulties associated with RL-based solutions, namely, high-variance gradients, uninformative rewards and brittle training. By leveraging the optimal transport distance, we introduce a regularizer that significantly alleviates the above issues. Our formulation emphasizes the preservation of semantic features, enabling end-to-end training instead of ad-hoc fine-tuning, and when combined with RL, it controls the exploration space for more efficient model updates. To validate the effectiveness of the proposed solution, we perform a comprehensive evaluation covering a wide variety of NLP tasks: machine translation, abstractive text summarization and image caption, with consistent improvements over competing solutions.

Background The Bethesda System for Reporting Thyroid Cytopathology (TBSRTC) comprises 6 categories used for the diagnosis of thyroid fine-needle aspiration biopsy (FNAB). Each category has an associated risk of malignancy, which is important in the management of a thyroid nodule. More accurate predictions of malignancy may help to reduce unnecessary surgery. A machine learning algorithm (MLA) was developed to evaluate thyroid FNAB via whole slide images (WSIs) to predict malignancy. Methods Files were searched for all thyroidectomy specimens with preceding FNAB over 8 years. All cytologic and surgical pathology diagnoses were recorded and correlated for each nodule. One representative slide from each case was scanned to create a WSI. An MLA was designed to identify follicular cells and predict the malignancy of the final pathology. The test set comprised cases blindly reviewed by a cytopathologist who assigned a TBSRTC category. The area under the receiver operating characteristic curve was used to assess the MLA performance. Results Nine hundred eight FNABs met the criteria. The MLA predicted malignancy with a sensitivity and specificity of 92.0% and 90.5%, respectively. The areas under the curve for the prediction of malignancy by the cytopathologist and the MLA were 0.931 and 0.932, respectively. Conclusions The performance of the MLA in predicting thyroid malignancy from FNAB WSIs is comparable to the performance of an expert cytopathologist. When the MLA and electronic medical record diagnoses are combined, the performance is superior to the performance of either alone. An MLA may be used as an adjunct to FNAB to assist in refining the indeterminate categories.

OBJECTIVES Host gene expression signatures discriminate bacterial and viral infection but have not been translated to a clinical test platform. This study enrolled an independent cohort of patients to describe and validate a first-in-class host response bacterial/viral test. DESIGN Subjects were recruited from 2006 to 2016. Enrollment blood samples were collected in an RNA preservative and banked for later testing. The reference standard was an expert panel clinical adjudication, which was blinded to gene expression and procalcitonin results. SETTING Four U.S. emergency departments. PATIENTS Six-hundred twenty-three subjects with acute respiratory illness or suspected sepsis. INTERVENTIONS Forty-five-transcript signature measured on the BioFire FilmArray System (BioFire Diagnostics, Salt Lake City, UT) in ~45 minutes. MEASUREMENTS AND MAIN RESULTS Host response bacterial/viral test performance characteristics were evaluated in 623 participants (mean age 46 yr; 45% male) with bacterial infection, viral infection, coinfection, or noninfectious illness. Performance of the host response bacterial/viral test was compared with procalcitonin. The test provided independent probabilities of bacterial and viral infection in ~45 minutes. In the 213-subject training cohort, the host response bacterial/viral test had an area under the curve for bacterial infection of 0.90 (95% CI, 0.84-0.94) and 0.92 (95% CI, 0.87-0.95) for viral infection. Independent validation in 209 subjects revealed similar performance with an area under the curve of 0.85 (95% CI, 0.78-0.90) for bacterial infection and 0.91 (95% CI, 0.85-0.94) for viral infection. The test had 80.1% (95% CI, 73.7-85.4%) average weighted accuracy for bacterial infection and 86.8% (95% CI, 81.8-90.8%) for viral infection in this validation cohort. This was significantly better than 68.7% (95% CI, 62.4-75.4%) observed for procalcitonin (p < 0.001). An additional cohort of 201 subjects with indeterminate phenotypes (coinfection or microbiology-negative infections) revealed similar performance. CONCLUSIONS The host response bacterial/viral measured using the BioFire System rapidly and accurately discriminated bacterial and viral infection better than procalcitonin, which can help support more appropriate antibiotic use.

Background Screening malaria-specific antibody responses on protein microarrays can help identify immune factors that mediate protection against malaria infection, disease, and transmission, as well as markers of past exposure to both malaria parasites and mosquito vectors. Most malaria protein microarray work has used serum as the sample matrix, requiring prompt laboratory processing and a continuous cold chain, thus limiting applications in remote locations. Dried blood spots (DBS) pose minimal biohazard, do not require immediate laboratory processing, and are stable at room temperature for transport, making them potentially superior alternatives to serum. The goals of this study were to assess the viability of DBS as a source for antibody profiling and to use DBS to identify serological signatures of low-density Plasmodium falciparum infections in malaria-endemic regions of Myanmar. Methods Matched DBS and serum samples from a cross-sectional study in Ingapu Township, Myanmar were probed on protein microarrays populated with P. falciparum antigen fragments. Signal and trends in both sample matrices were compared. A case-control study was then performed using banked DBS samples from malaria-endemic regions of Myanmar, and a regularized logistic regression model was used to identify antibody signatures of ultrasensitive PCR-positive P. falciparum infections. Results Approximately 30% of serum IgG activity was recovered from DBS. Despite this loss of antibody activity, antigen and population trends were well-matched between the two sample matrices. Responses to 18 protein fragments were associated with the odds of asymptomatic P. falciparum infection, albeit with modest diagnostic characteristics (sensitivity 58%, specificity 85%, negative predictive value 88%, and positive predictive value 52%). Conclusions Malaria-specific antibody responses can be reliably detected, quantified, and analysed from DBS, opening the door to serological studies in populations where serum collection, transport, and storage would otherwise be impossible. While test characteristics of antibody signatures were insufficient for individual diagnosis, serological testing may be useful for identifying exposure to asymptomatic, low-density malaria infections, particularly if sero-surveillance strategies target individuals with low previous exposure as sentinels for population exposure.

The primary objective of this study is to detect biomarkers and develop models that enable the identification of clinically significant prostate cancer and to understand the biologic implications of the genes involved. Peripheral blood samples (1018 patients) were split chronologically into independent training (n = 713) and validation (n = 305) sets. Whole transcriptome RNA sequencing was performed on isolated phagocytic CD14+ and non-phagocytic CD2+ cells and their gene expression levels were used to develop predictive models that correlate to adverse pathologic features. The immune-transcriptomic model with the highest performance for predicting adverse pathology, based on a subtraction of the log-transformed expression signals of the two cell types, displayed an area under the curve (AUC) of the receiver operating characteristic of 0.70. The addition of biomarkers in combination with traditional clinical risk factors (age, serum prostate-specific antigen (PSA), PSA density, race, digital rectal examination (DRE), and family history) enhanced the AUC to 0.91 and 0.83 for the training and validation sets, respectively. The markers identified by this approach uncovered specific pathway associations relevant to (prostate) cancer biology. Increased phagocytic activity in conjunction with cancer-associated (mis-)regulation is also represented by these markers. Differential gene expression of circulating immune cells gives insight into the cellular immune response to early tumor development and immune surveillance.

Context.— Prostate cancer is a common malignancy, and accurate diagnosis typically requires histologic review of multiple prostate core biopsies per patient. As pathology volumes and complexity increase, new tools to improve the efficiency of everyday practice are keenly needed. Deep learning has shown promise in pathology diagnostics, but most studies silo the efforts of pathologists from the application of deep learning algorithms. Very few hybrid pathologist–deep learning approaches have been explored, and these typically require complete review of histologic slides by both the pathologist and the deep learning system. Objective.— To develop a novel and efficient hybrid human–machine learning approach to screen prostate biopsies. Design.— We developed an algorithm to determine the 20 regions of interest with the highest probability of malignancy for each prostate biopsy; presenting these regions to a pathologist for manual screening limited the initial review by a pathologist to approximately 2% of the tissue area of each sample. We evaluated this approach by using 100 biopsies (29 malignant, 60 benign, 11 other) that were reviewed by 4 pathologists (3 urologic pathologists, 1 general pathologist) using a custom-designed graphical user interface. Results.— Malignant biopsies were correctly identified as needing comprehensive review with high sensitivity (mean, 99.2% among all pathologists); conversely, most benign prostate biopsies (mean, 72.1%) were correctly identified as needing no further review. Conclusions.— This novel hybrid system has the potential to efficiently triage out most benign prostate core biopsies, conserving time for the pathologist to dedicate to detailed evaluation of malignant biopsies.

Background: Distinguishing bacterial and viral respiratory infections is challenging. Novel diagnostics based on differential host gene expression patterns are promising but have not been translated to a clinical platform nor extensively tested. Here, we validate a microarray-derived host response signature and explore performance in microbiology-negative and coinfection cases. Methods: Subjects with acute respiratory illness were enrolled in participating emergency departments. Reference standard was an adjudicated diagnosis of bacterial infection, viral infection, both, or neither. An 87-transcript signature for distinguishing bacterial, viral, and noninfectious illness was measured from peripheral blood using RT-PCR. Performance characteristics were evaluated in subjects with confirmed bacterial, viral, or noninfectious illness. Subjects with bacterial-viral coinfection and microbiologically-negative suspected bacterial infection were also evaluated. Performance was compared to procalcitonin. Findings: 151 subjects with microbiologically confirmed, single-etiology illness were tested, yielding AUROCs 0•85–0•89 for bacterial, viral, and noninfectious illness. Accuracy was similar to procalcitonin (88% vs 83%, p = 0•23) for bacterial vs. non-bacterial infection. Whereas procalcitonin cannot distinguish viral from non-infectious illness, the RT-PCR test had 81% accuracy in making this determination. Bacterial-viral coinfection was subdivided. Among 19 subjects with bacterial superinfection, the RT-PCR test identified 95% as bacterial, compared to 68% with procalcitonin (p = 0•13). Among 12 subjects with bacterial infection superimposed on chronic viral infection, the RT-PCR test identified 83% as bacterial, identical to procalcitonin. 39 subjects had suspected bacterial infection; the RT-PCR test identified bacterial infection more frequently than procalcitonin (82% vs 64%, p = 0•02). Interpretation: The RT-PCR test offered similar diagnostic performance to procalcitonin in some subgroups but offered better discrimination in others such as viral vs. non-infectious illness and bacterial/viral coinfection. Gene expression-based tests could impact decision-making for acute respiratory illness as well as a growing number of other infectious and non-infectious diseases. Keywords: Biomarkers, Gene expression, Respiratory tract infections, Coinfection, Diagnosis, Precision medicine

Small bowel obstruction (SBO) results in >350,000 operations and >$2 billion annual health care expenditures in the US. Prompt, effective identification of patients at high/low surgery risk could improve survival, lower complication rates, and shorten hospitalization lengths. SBO surgery prediction models were developed based on SBO-related encounters in the Duke University Health System between 2013 and 2017. A total of 3,910 encounters among 3,374 unique patients were identified. Performance was assessed in each hour after admission when predicting whether patients will (a) receive surgery within 24 hours, and (b) receive surgery during the current encounter. Potential benefits of model-based discharge were assessed using the incorrect discharge rate and average reduction in hospital stay. Model-based discharge of low-risk patients was projected to reduce the average length of stay among patients not receiving surgery by >60 hours while maintaining an incorrect discharge rate lower than the observed readmission rate (9.3%). AUROC for the 24-hour prediction task increased from 0.644 to 0.779 at 12 and 72 hours post-admission, respectively. Future work will prospectively explore the benefits of model deployment in an inpatient setting.

Set classification is the task of predicting a single label from a set comprising multiple instances. The examples we consider are pathology slides represented by sets of patches and medical text data represented by sets of word embeddings. State-of-the-art methods, such as the transformer network, typically use attention mechanisms to learn representations of set data, by modeling interactions between instances of the set. These methods, however, have complex heuristic architectures comprising multiple heads and layers. The complexity of attention architectures hampers their training when only a small number of labeled sets is available, as is often the case in medical applications. To address this problem, we present a kernel-based representation learning framework that links learning affinity kernels to learning representations from attention architectures. We show that learning a combination of the sum and the product of kernels is equivalent to learning representations from multi-head multi-layer attention architectures. From our framework, we devise a simplified attention architecture which we term affinitention (affinity-attention) nets. We demonstrate the application of affinitention nets to the classification of the Set-Cifar10 dataset, thyroid malignancy prediction from pathology slides, as well as patient text-message triage. We show that affinitention nets provide competitive results compared to heuristic attention architectures and outperform other competing methods.

Balanced representation learning methods have been applied successfully to counterfactual inference from observational data. However, approaches that account for survival outcomes are relatively limited. Survival data are frequently encountered across diverse medical applications, i.e., drug development, risk profiling, and clinical trials, and such data are also relevant in fields like manufacturing (e.g., for equipment monitoring). When the outcome of interest is a time-to-event, special precautions for handling censored events need to be taken, as ignoring censored outcomes may lead to biased estimates. We propose a theoretically grounded unified framework for counterfactual inference applicable to survival outcomes. Further, we formulate a nonparametric hazard ratio metric for evaluating average and individualized treatment effects. Experimental results on real-world and semi-synthetic datasets, the latter of which we introduce, demonstrate that the proposed approach significantly outperforms competitive alternatives in both survival-outcome prediction and treatment-effect estimation., Accepted at ACM Conference on Health, Inference, and Learning (ACM CHIL 2021). Code at https://github.com/paidamoyo/counterfactual_survival_analysis

BACKGROUND AND AIMS Whether glycemic control, as opposed to diabetes status, is associated with the severity of NAFLD is open for study. We aimed to evaluate whether degree of glycemic control in the years preceding liver biopsy predicts the histological severity of NASH. APPROACH AND RESULTS Using the Duke NAFLD Clinical Database, we examined patients with biopsy-proven NAFLD/NASH (n = 713) and the association of liver injury with glycemic control as measured by hemoglobin A1c (HbA1c). The study cohort was predominantly female (59%) and White (84%) with median (interquartile range) age of 50 (42, 58) years; 49% had diabetes (n = 348). Generalized linear regression models adjusted for age, sex, race, diabetes, body mass index, and hyperlipidemia were used to assess the association between mean HbA1c over the year preceding liver biopsy and severity of histological features of NAFLD/NASH. Histological features were graded and staged according to the NASH Clinical Research Network system. Group-based trajectory analysis was used to examine patients with at least three HbA1c (n = 298) measures over 5 years preceding clinically indicated liver biopsy. Higher mean HbA1c was associated with higher grade of steatosis and ballooned hepatocytes, but not lobular inflammation. Every 1% increase in mean HbA1c was associated with 15% higher odds of increased fibrosis stage (OR, 1.15; 95% CI, 1.01, 1.31). As compared with good glycemic control, moderate control was significantly associated with increased severity of ballooned hepatocytes (OR, 1.74; 95% CI, 1.01, 3.01; P = 0.048) and hepatic fibrosis (HF; OR, 4.59; 95% CI, 2.33, 9.06; P

In many natural language processing applications, identifying predictive text can be as important as the predictions themselves. When predicting medical diagnoses, for example, identifying predictive content in clinical notes not only enhances interpretability, but also allows unknown, descriptive (i.e., text-based) risk factors to be identified. We here formalize this problem as predictive extraction and address it using a simple mechanism based on linear attention. Our method preserves differentiability, allowing scalable inference via stochastic gradient descent. Further, the model decomposes predictions into a sum of contributions of distinct text spans. Importantly, we require only document labels, not ground-truth spans. Results show that our model identifies semantically-cohesive spans and assigns them scores that agree with human ratings, while preserving classification performance.

Using structured elements from Electronic Health Records (EHR), we seek to:i) build predictive models to stratify patients tested for COVID-19 by their likelihood for hospitalization, ICU admission, mechanical ventilation and inpatient mortality, andii) identify the most important EHR-based features driving the predictions. We leveraged EHR data from the Duke University Health System tested for COVID-19 or hospitalized between March 11, 2020 and August 24, 2020, to build models to predict hospital admissions within 4 weeks. Models were also created for ICU admissions, need for mechanical ventilation and mortality following admission. Models were developed on a cohort of 86,355 patients with 112,392 outpatient COVID-19 tests or any-cause hospital admissions between March 11, 2020 and June 4, 2020. The four models considered resulted in AUROC=0.838 (CI: 0.832-0.844) and AP=0.272 (CI: 0.260-0.287) for hospital admissions, AUROC=0.847 (CI: 0.839-855) and AP=0.585 (CI: 0.565-0.603) for ICU admissions, AUROC=0.858 (CI: 0.846-0.871) and AP=0.434 (CI: 0.403-0.467) for mechanical ventilation, and AUROC=0.0.856 (CI: 0.842-0.872) and AP=0.243 (CI: 0.205-0.282) for inpatient mortality. Patient history abstracted from the EHR has the potential for being used to stratify patients tested for COVID-19 in terms of utilization and mortality. The dominant EHR features for hospital admissions and inpatient outcomes are different. For the former, age, social indicators and previous utilization are the most important predictive features. For the latter, age and physiological summaries (pulse and blood pressure) are the main drivers.

SARS-CoV-2 infection triggers highly variable host responses and causes varying degrees of illness in humans. We sought to harness the peripheral blood mononuclear cell (PBMC) response over the course of illness to provide insight into COVID-19 physiology. We analyzed PBMCs from subjects with variable symptom severity at different stages of clinical illness before and after IgG seroconversion to SARS-CoV-2. Prior to seroconversion, PBMC transcriptomes did not distinguish symptom severity. In contrast, changes in chromatin accessibility were associated with symptom severity. Furthermore, single-cell analyses revealed evolution of the chromatin accessibility landscape and transcription factor motif occupancy for individual PBMC cell types. The most extensive remodeling occurred in CD14+ monocytes where sub-populations with distinct chromatin accessibility profiles were associated with disease severity. Our findings indicate that pre-seroconversion chromatin remodeling in certain innate immune populations is associated with divergence in symptom severity, and the identified transcription factors, regulatory elements, and downstream pathways provide potential prognostic markers for COVID-19 subjects.One sentence summaryChromatin accessibility in immune cells from COVID-19 subjects is remodeled prior to seroconversion to reflect disease severity.

Background: Peripheral artery disease (PAD) is underrecognized, undertreated, and understudied: each of these endeavors requires efficient and accurate identification of patients with PAD. Currently, PAD patient identification relies on diagnosis/procedure codes or lists of patients diagnosed or treated by specific providers in specific locations and ways. The goal of this research was to leverage natural language processing to more accurately identify patients with PAD in an electronic health record system compared with a structured data–based approach. Methods: The clinical notes from a cohort of 6861 patients in our health system whose PAD status had previously been adjudicated were used to train, test, and validate a natural language processing model using 10-fold cross-validation. The performance of this model was described using the area under the receiver operating characteristic and average precision curves; its performance was quantitatively compared with an administrative data–based least absolute shrinkage and selection operator (LASSO) approach using the DeLong test. Results: The median (SD) of the area under the receiver operating characteristic curve for the natural language processing model was 0.888 (0.009) versus 0.801 (0.017) for the LASSO-based approach alone (DeLong P Conclusions: Using a natural language processing approach in addition to partial cohort preprocessing with a LASSO-based model, we were able to meaningfully improve our ability to identify patients with PAD compared with an approach using structured data alone. This model has potential applications to both interventions targeted at improving patient care as well as efficient, large-scale PAD research. Graphic Abstract: A graphic abstract is available for this article.

Background/AimsTo develop a convolutional neural network (CNN) to detect symptomatic Alzheimer’s disease (AD) using a combination of multimodal retinal images and patient data.MethodsColour maps of ganglion cell-inner plexiform layer (GC-IPL) thickness, superficial capillary plexus (SCP) optical coherence tomography angiography (OCTA) images, and ultra-widefield (UWF) colour and fundus autofluorescence (FAF) scanning laser ophthalmoscopy images were captured in individuals with AD or healthy cognition. A CNN to predict AD diagnosis was developed using multimodal retinal images, OCT and OCTA quantitative data, and patient data.Results284 eyes of 159 subjects (222 eyes from 123 cognitively healthy subjects and 62 eyes from 36 subjects with AD) were used to develop the model. Area under the receiving operating characteristic curve (AUC) values for predicted probability of AD for the independent test set varied by input used: UWF colour AUC 0.450 (95% CI 0.282, 0.592), OCTA SCP 0.582 (95% CI 0.440, 0.724), UWF FAF 0.618 (95% CI 0.462, 0.773), GC-IPL maps 0.809 (95% CI 0.700, 0.919). A model incorporating all images, quantitative data and patient data (AUC 0.836 (CI 0.729, 0.943)) performed similarly to models only incorporating all images (AUC 0.829 (95% CI 0.719, 0.939)). GC-IPL maps, quantitative data and patient data AUC 0.841 (95% CI 0.739, 0.943).ConclusionOur CNN used multimodal retinal images to successfully predict diagnosis of symptomatic AD in an independent test set. GC-IPL maps were the most useful single inputs for prediction. Models including only images performed similarly to models also including quantitative data and patient data.

Identifying patients at higher risk for poor outcomes from nonalcoholic fatty liver disease (NAFLD) remains challenging. Metabolomics, the comprehensive measurement of small molecules in biological samples, has the potential to reveal novel noninvasive biomarkers. The aim of this study was to determine if serum metabolite profiles in patients with NAFLD associate with future liver-related events. We performed a retrospective single-center cohort study of 187 participants with biopsy-proven NAFLD. Metabolomic analysis was performed on serum using ultrahigh performance liquid chromatography/tandem mass spectrometry and gas chromatography/mass spectrometry. We identified liver-related events (variceal bleeding, ascites, spontaneous bacterial peritonitis, hepatic encephalopathy, hepatocellular carcinoma, hepatopulmonary or hepatorenal syndrome) by manual chart review between index biopsy (2007-2013) and April 1, 2018. Generalized linear models and Cox proportional hazards models were used to test the association of metabolites with liver-related events and time to first liver-related event, controlling for covariates and fibrosis stage. Over a mean ± SD follow-up of 6.9 ± 3.2 years, 11 participants experienced 22 liver-related events. Generalized linear models revealed 53 metabolites significantly associated with liver-related events (P < 0.05). In Cox proportional hazards modeling, 69 metabolites were significantly associated with time to future liver-related events (P < 0.05), seven of which met the false discovery rate threshold of 0.10: vitamin E metabolites gamma-carboxyethyl-hydroxychroman (gamma-CEHC) and gamma-CEHC glucuronide; primary bile acid metabolite taurochenodeoxycholate; serotonin metabolite 5-hydroxyindoleacetate; and lipid metabolites (i) 2-hydroxyglutarate, (ii) 3beta,17beta-diol disulfate 1, and (iii) eicosenoyl sphingomyelin. Conclusion: Metabolites of a primary bile acid, vitamin E, and serotonin were associated with future liver-related events. Our results suggest metabolite pathways may be useful for predicting which patients with NAFLD are at higher risk for hepatic decompensation.

T2DM increases risk for advanced fibrosis in NAFLD, however little is known about the impact of glycemic control on fibrosis severity. Objective: To assess whether glycemic control is associated with severity of fibrosis in NAFLD. Methods: We utilized the Duke NAFLD Clinical Database to examine the association between HbA1c and fibrosis in patients with biopsy (Bx)-proven NAFLD (n=941). A generalized linear regression model was used (adjusted for age, gender, race, BMI, HLD, eGFR) to assess the association of mean HbA1c (1 year pre- to 90d post-Bx) and fibrosis severity (stage 3-4=advanced vs. 0-2=mild). Patients with ≥3 HbA1c values (n=335) were examined by group-based trajectory analysis using HbA1c over 5 yrs pre- to 90d post-Bx. Results: Every 1% increase in mean HbA1c was associated with 20% higher odds of advanced fibrosis (OR 1.20, 95%CI 1.02, 1.39 p=0.02). Group-based trajectory analysis identified three glycemic patterns - good, moderate and poor (groups 1,2,3, respectively); proportion of advanced fibrosis increased with worse glycemic control. Group 2 had higher odds of advanced fibrosis than group 1 (OR 4.46, 95%CI 2.29,8.70, p Conclusion: Glycemic control preceding Bx impacts severity of liver fibrosis. We are currently analyzing 65 patients with paired Bx to assess impact of HbA1c trends on change in fibrosis severity. Disclosure A. Alexopoulos: None. M. Crowley: None. Y. Wang: None. C.D. Guy: Consultant; Self; CymaBay, HistoIndex, Madrigal, NGM. R. Henao: None. K.A. Seymour: Other Relationship; Self; Gore, Medtronic. R. Sudan: None. D. Portenier: Consultant; Self; Medtronic. Speaker’s Bureau; Self; Novo Nordisk Inc. A. Diehl: None. M.F. Abdelmalek: None. A.D. Coviello: Consultant; Self; GI Dynamics Inc., Novo Nordisk Inc. Research Support; Self; Allergan plc., GENFIT, Orthus Health. Speaker’s Bureau; Self; Novo Nordisk Inc. Funding Endocrine Fellows Foundation

INTRODUCTION AND OBJECTIVE:Peripheral blood CD14/CD2 RNA expression ratios have shown associations with prostate biopsy features. The objectives of this study are: 1) To develop a computational imm...

Background Pathogen-based diagnostics for acute respiratory infection (ARI) have limited ability to detect etiology of illness. We previously showed that peripheral blood-based host gene expression classifiers accurately identify bacterial and viral ARI in cohorts of European and African descent. We determined classifier performance in a South Asian cohort. Methods Patients ≥15 years with fever and respiratory symptoms were enrolled in Sri Lanka. Comprehensive pathogen-based testing was performed. Peripheral blood ribonucleic acid was sequenced and previously developed signatures were applied: a pan-viral classifier (viral vs nonviral) and an ARI classifier (bacterial vs viral vs noninfectious). Results Ribonucleic acid sequencing was performed in 79 subjects: 58 viral infections (36 influenza, 22 dengue) and 21 bacterial infections (10 leptospirosis, 11 scrub typhus). The pan-viral classifier had an overall classification accuracy of 95%. The ARI classifier had an overall classification accuracy of 94%, with sensitivity and specificity of 91% and 95%, respectively, for bacterial infection. The sensitivity and specificity of C-reactive protein (>10 mg/L) and procalcitonin (>0.25 ng/mL) for bacterial infection were 100% and 34%, and 100% and 41%, respectively. Conclusions Previously derived gene expression classifiers had high predictive accuracy at distinguishing viral and bacterial infection in South Asian patients with ARI caused by typical and atypical pathogens., Host biomarkers can identify bacterial versus viral acute respiratory infection (ARI). Previously, we derived host gene expression classifiers that accurately identify bacterial versus viral ARI. Here, we show that our classifiers had high predictive accuracy in a South Asian cohort.

$\textbf{Objective}$: To develop a multi-channel device event segmentation and feature extraction algorithm that is robust to changes in data distribution. $\textbf{Methods}$: We introduce an adaptive transfer learning algorithm to classify and segment events from non-stationary multi-channel temporal data. Using a multivariate hidden Markov model (HMM) and Fisher's linear discriminant analysis (FLDA) the algorithm adaptively adjusts to shifts in distribution over time. The proposed algorithm is unsupervised and learns to label events without requiring $\textit{a priori}$ information about true event states. The procedure is illustrated on experimental data collected from a cohort in a human viral challenge (HVC) study, where certain subjects have disrupted wake and sleep patterns after exposure to a H1N1 influenza pathogen. $\textbf{Results}$: Simulations establish that the proposed adaptive algorithm significantly outperforms other event classification methods. When applied to early time points in the HVC data the algorithm extracts sleep/wake features that are predictive of both infection and infection onset time. $\textbf{Conclusion}$: The proposed transfer learning event segmentation method is robust to temporal shifts in data distribution and can be used to produce highly discriminative event-labeled features for health monitoring. $\textbf{Significance}$: Our integrated multisensor signal processing and transfer learning method is applicable to many ambulatory monitoring applications.

Pretrained Language Models (PLMs) have improved the performance of natural language understanding in recent years. Such models are pretrained on large corpora, which encode the general prior knowledge of natural languages but are agnostic to information characteristic of downstream tasks. This often results in overfitting when fine-tuned with low resource datasets where task-specific information is limited. In this paper, we integrate label information as a task-specific prior into the self-attention component of pretrained BERT models. Experiments on several benchmarks and real-word datasets suggest that the proposed approach can largely improve the performance of pretrained models when fine-tuning with small datasets.

The primary goal of knowledge distillation (KD) is to encapsulate the information of a model learned from a teacher network into a student network, with the latter being more compact than the former. Existing work, e.g., using Kullback-Leibler divergence for distillation, may fail to capture important structural knowledge in the teacher network and often lacks the ability for feature generalization, particularly in situations when teacher and student are built to address different classification tasks. We propose Wasserstein Contrastive Representation Distillation (WCoRD), which leverages both primal and dual forms of Wasserstein distance for KD. The dual form is used for global knowledge transfer, yielding a contrastive learning objective that maximizes the lower bound of mutual information between the teacher and the student networks. The primal form is used for local contrastive knowledge transfer within a mini-batch, effectively matching the distributions of features between the teacher and the student networks. Experiments demonstrate that the proposed WCoRD method outperforms state-of-the-art approaches on privileged information distillation, model compression and cross-modal transfer., Comment: Accepted by CVPR 2021

The response to respiratory viruses varies substantially between individuals, and there are currently no known molecular predictors from the early stages of infection. Here we conduct a community-based analysis to determine whether pre- or early post-exposure molecular factors could predict physiologic responses to viral exposure. Using peripheral blood gene expression profiles collected from healthy subjects prior to exposure to one of four respiratory viruses (H1N1, H3N2, Rhinovirus, and RSV), as well as up to 24 h following exposure, we find that it is possible to construct models predictive of symptomatic response using profiles even prior to viral exposure. Analysis of predictive gene features reveal little overlap among models; however, in aggregate, these genes are enriched for common pathways. Heme metabolism, the most significantly enriched pathway, is associated with a higher risk of developing symptoms following viral exposure. This study demonstrates that pre-exposure molecular predictors can be identified and improves our understanding of the mechanisms of response to respiratory viruses., The response to respiratory virus exposure can currently not be predicted by pre- or early post-exposure molecular signatures. Here, the authors conduct a community-based analysis of blood gene expression from healthy individuals exposed to respiratory viruses and provide predictive models and biological insight into the physiological response.

The severity of hepatic fibrosis is the primary predictor of liver‐related morbidity and mortality in patients with nonalcoholic fatty liver disease (NAFLD). Unfortunately, noninvasive serum biomarkers for NAFLD‐associated fibrosis are limited. We analyzed baseline serum samples for 24 cytokines of 97 patients with biopsy‐proven NAFLD. These patients were prospectively enrolled in a clinical study (ClinicalTrials.gov NCT00794716) to identify cytokines associated with liver fibrosis in patients with nonalcoholic steatohepatitis. Patients were stratified according to severity of hepatic fibrosis (mild, stage 0‐1, n = 37; moderate, stage 2, n = 40; and advanced, stage 3‐4, n = 20) while controlling for age, race, sex, body mass index, and diabetes mellitus. Interleukin‐8 (IL‐8), osteopontin (OPN), and monocyte chemoattractant protein 1 (MCP1) were associated with liver fibrosis (P

Nearly a third of all surgeries performed in the United States occur for patients over the age of 65; these older adults experience a higher rate of postoperative morbidity and mortality. To improve the care for these patients, we aim to identify and characterize high risk geriatric patients to send to a specialized perioperative clinic while leveraging the overall surgical population to improve learning. To this end, we develop a hierarchical infinite latent factor model (HIFM) to appropriately account for the covariance structure across subpopulations in data. We propose a novel Hierarchical Dirichlet Process shrinkage prior on the loadings matrix that flexibly captures the underlying structure of our data while sharing information across subpopulations to improve inference and prediction. The stick-breaking construction of the prior assumes an infinite number of factors and allows for each subpopulation to utilize different subsets of the factor space and select the number of factors needed to best explain the variation. We develop the model into a latent factor regression method that excels at prediction and inference of regression coefficients. Simulations validate this strong performance compared to baseline methods. We apply this work to the problem of predicting surgical complications using electronic health record data for geriatric patients and all surgical patients at Duke University Health System (DUHS). The motivating application demonstrates the improved predictive performance when using HIFM in both area under the ROC curve and area under the PR Curve while providing interpretable coefficients that may lead to actionable interventions.

Data on patterns of weight change among adults with overweight or obesity are minimal. We aimed to examine patterns of weight change and associated hospitalizations in a large health system, and to develop a model to predict 2-year significant weight gain. Data from the Duke University Health System was abstracted from 1/1/13 to 12/31/16 on patients with BMI ≥ 25 kg/m2 in 2014. A regression model was developed to predict patients that would increase their weight by 10% within 2 years. We estimated the association between weight change category and all-cause hospitalization using Cox proportional hazards models. Of the 37,253 patients in our cohort, 59% had stable weight over 2 years, while 24% gained ≥ 5% weight and 17% lost ≥ 5% weight. Our predictive model had reasonable discriminatory capacity to predict which individuals would gain ≥ 10% weight over 2 years (AUC 0.73). Compared with stable weight, the risk of hospitalization was increased by 37% for individuals with > 10% weight loss [adj. HR (95% CI): 1.37 (1.25,1.5)], by 30% for those with > 10% weight gain [adj. HR (95% CI): 1.3 (1.19,1.42)], by 18% for those with 5–10% weight loss [adj. HR (95% CI): 1.18 (1.09,1.28)], and by 10% for those with 5–10% weight gain [adj. HR (95% CI): 1.1 (1.02,1.19)]. In this examination of a large health system, significant weight gain or loss of > 10% was associated with increased all-cause hospitalization over 2 years compared with stable weight. This analysis adds to the increasing observational evidence that weight stability may be a key health driver.

Key Points Question Can noninvasive, wrist-worn wearable devices detect acute viral respiratory infection and predict infection severity before symptom onset? Findings In a cohort study of 31 participants inoculated with H1N1 and 18 participants with rhinovirus, infection detection and severity prediction models trained using data on wearable devices were able to distinguish between infection and noninfection with 92% accuracy for H1N1 and 88% accuracy for rhinovirus and were able to distinguish between mild and moderate infection 24 hours prior to symptom onset with 90% accuracy for H1N1 and 89% accuracy for rhinovirus. Meaning This study suggests that the use of wearable devices to identify individuals with presymptomatic acute viral respiratory infection is feasible; because wearable devices are common in the general population, using them for infection screening may help limit the spread of contagion., Importance Currently, there are no presymptomatic screening methods to identify individuals infected with a respiratory virus to prevent disease spread and to predict their trajectory for resource allocation. Objective To evaluate the feasibility of using noninvasive, wrist-worn wearable biometric monitoring sensors to detect presymptomatic viral infection after exposure and predict infection severity in patients exposed to H1N1 influenza or human rhinovirus. Design, Setting, and Participants The cohort H1N1 viral challenge study was conducted during 2018; data were collected from September 11, 2017, to May 4, 2018. The cohort rhinovirus challenge study was conducted during 2015; data were collected from September 14 to 21, 2015. A total of 39 adult participants were recruited for the H1N1 challenge study, and 24 adult participants were recruited for the rhinovirus challenge study. Exclusion criteria for both challenges included chronic respiratory illness and high levels of serum antibodies. Participants in the H1N1 challenge study were isolated in a clinic for a minimum of 8 days after inoculation. The rhinovirus challenge took place on a college campus, and participants were not isolated. Exposures Participants in the H1N1 challenge study were inoculated via intranasal drops of diluted influenza A/California/03/09 (H1N1) virus with a mean count of 106 using the median tissue culture infectious dose (TCID50) assay. Participants in the rhinovirus challenge study were inoculated via intranasal drops of diluted human rhinovirus strain type 16 with a count of 100 using the TCID50 assay. Main Outcomes and Measures The primary outcome measures included cross-validated performance metrics of random forest models to screen for presymptomatic infection and predict infection severity, including accuracy, precision, sensitivity, specificity, F1 score, and area under the receiver operating characteristic curve (AUC). Results A total of 31 participants with H1N1 (24 men [77.4%]; mean [SD] age, 34.7 [12.3] years) and 18 participants with rhinovirus (11 men [61.1%]; mean [SD] age, 21.7 [3.1] years) were included in the analysis after data preprocessing. Separate H1N1 and rhinovirus detection models, using only data on wearble devices as input, were able to distinguish between infection and noninfection with accuracies of up to 92% for H1N1 (90% precision, 90% sensitivity, 93% specificity, and 90% F1 score, 0.85 [95% CI, 0.70-1.00] AUC) and 88% for rhinovirus (100% precision, 78% sensitivity, 100% specificity, 88% F1 score, and 0.96 [95% CI, 0.85-1.00] AUC). The infection severity prediction model was able to distinguish between mild and moderate infection 24 hours prior to symptom onset with an accuracy of 90% for H1N1 (88% precision, 88% sensitivity, 92% specificity, 88% F1 score, and 0.88 [95% CI, 0.72-1.00] AUC) and 89% for rhinovirus (100% precision, 75% sensitivity, 100% specificity, 86% F1 score, and 0.95 [95% CI, 0.79-1.00] AUC). Conclusions and Relevance This cohort study suggests that the use of a noninvasive, wrist-worn wearable device to predict an individual’s response to viral exposure prior to symptoms is feasible. Harnessing this technology would support early interventions to limit presymptomatic spread of viral respiratory infections, which is timely in the era of COVID-19., This cohort study evaluates the feasibility of using noninvasive, wrist-worn wearable devices to detect presymptomatic viral infection after exposure and predict infection severity in patients exposed to H1N1 influenza or human rhinovirus.

Background Host gene expression has emerged as a promising diagnostic strategy to discriminate bacterial and viral infection. Multiple gene signatures of varying size and complexity have been developed in various clinical populations. However, there has been no systematic comparison of these signatures. It is also unclear how these signatures apply to different clinical populations. This meta-analysis examined 19 published signatures, validated in 49 publicly available datasets for a total of 4750 patients. The objectives were to understand how the signatures compared to each other with respect to composition and performance, and to evaluate their performance in different patient subgroups. Methods Signatures were characterized with respect to size, platform, and discovery population. For each of the 19 signatures, we ran leave-one-out cross-validation to generate AUCs for bacterial classification and viral classification. We then applied dataset-specific thresholds to generate accuracies, pooling patients across datasets. Results Signature performance varied significantly with a median AUC across all validation datasets ranging from 0.55 to 0.94 for bacterial classification and 0.79 to 0.96 for viral classification. Signature size varied (1- 341 genes) with smaller signatures generally performing more poorly for both bacterial classification (P < .001) and for viral classification (P = 0.02). Viral infection was easier to diagnose than bacterial infection (85% vs. 80% overall accuracy, respectively; P < .001). Host gene expression classifiers performed more poorly in children < 12-years compared to those older than 12-years for both bacterial infection (77% vs. 83%, respectively; P < .001) and for viral infection (82% vs. 89%, respectively; P < .001). We did not observe differences based on illness severity as defined by ICU care for either bacterial or viral infections. Conclusion We observed significant differences among gene expression signatures for bacterial/viral discrimination, though these were not due to variations in the discovery methods or populations. Signature size directly correlated with test performance, which was generally better for the diagnosis of viral infection and in populations >12-years. Disclosures Ephraim L. Tsalik, MD, MHS, PhD, Predigen (Shareholder, Other Financial or Material Support, Founder)

Background Difficulty distinguishing bacterial and viral infections contributes to excess antibiotic use. A host response strategy overcomes many limitations of pathogen-based tests, but depends on a functional immune system. This approach may therefore be limited in immunocompromised (IC) hosts. Here, we evaluated a host response test in IC subjects, which has not been extensively studied in this manner. Methods An 81-gene signature was measured using qRT-PCR in previously enrolled IC subjects (chemotherapy, solid organ transplant, immunomodulatory agents, AIDS) with confirmed bacterial infection, viral infection, or non-infectious illness (NI). A regularized logistic regression model estimated the likelihood of bacterial, viral, and noninfectious classes. Clinical adjudication was the reference standard. Results A host gene expression model trained in a cohort of 136 immunocompetent subjects (43 bacterial, 41 viral, and 52 NI) had an overall accuracy of 84.6% for the diagnosis of bacterial vs. non-bacterial infection and 80.8% for viral vs. non-viral infection. The model was validated in an independent cohort of 134 IC subjects (64 bacterial, 28 viral, 42 NI). The overall accuracy was 73.9% for bacterial infection (p=0.03 vs. training cohort) and 75.4% for viral infection (p=0.27). Test utility could be improved by reporting probability ranges. For example, results divided into probability quartiles would allow the highest quartile to be used to rule in infection and the lowest to rule out infection. For IC subjects in the lowest quartile, the test had 90.1% and 96.4% sensitivity for bacterial and viral infection, respectively. For the highest quartile, the test had 91.4% and 84.0% specificity for bacterial and viral infection, respectively. The type or number of immunocompromising conditions did not impact performance. Illness Etiology Probabilities Conclusion A host gene expression test discriminated bacterial, viral, and non-infectious etiologies at a lower overall accuracy in IC patients compared to immunocompetent patients, though this difference was only significant for bacterial vs non-bacterial disease. With modified interpretive criteria, a host response strategy may offer clinically useful and complementary diagnostic information for IC patients. Disclosures Thomas W. Burke, PhD, Predigen, Inc (Consultant) Geoffrey S. Ginsburg, MD PhD, Predigen, Inc (Shareholder, Other Financial or Material Support) Christopher W. Woods, MD, MPH, FIDSA, Predigen, Inc (Shareholder, Other Financial or Material Support) Ephraim L. Tsalik, MD, MHS, PhD, FIDSA, Predigen, Inc (Scientific Research Study Investigator, Shareholder, Other Financial or Material Support)

Objective Distinguishing bacterial, viral, or other etiologies of acute illness is diagnostically challenging with significant implications for appropriate antimicrobial use. Host gene expression offers a promising approach, although no clinically useful test has been developed yet to accomplish this. Here, Qvella’s FAST HR (Richmond Hill, Ontario, Canada) process was developed to quantify previously identified host gene expression signatures in whole blood in Method Whole blood was collected from 128 human subjects (mean age 47, range 18–88) with clinically adjudicated, microbiologically confirmed viral infection, bacterial infection, noninfectious illness, or healthy controls. Stabilized mRNA was released from cleaned and stabilized RNA-surfactant complexes using e-lysis, an electrical process providing a quantitative real-time reverse transcription polymerase chain reaction-ready sample. Threshold cycle values (CT) for 10 host response targets were normalized to hypoxanthine phosphoribosyltransferase 1 expression, a control mRNA. The transcripts in the signature were specifically chosen to discriminate viral from nonviral infection (bacterial, noninfectious illness, or healthy). Classification accuracy was determined using cross-validated sparse logistic regression. Results Reproducibility of mRNA quantification was within 1 cycle as compared to the difference seen between subjects with viral versus nonviral infection (up to 5.0 normalized CT difference). Classification of 128 subjects into viral or nonviral etiologies demonstrated 90.6% overall accuracy compared to 82.0% for procalcitonin (P = .06). FAST HR achieved rapid and accurate measurement of the host response to viral infection in less than 45 minutes. Conclusions These results demonstrate the ability to translate host gene expression signatures to clinical platforms for use in patients with suspected infection. Clinical Trials Registration NCT00258869.

Summary Background Early and accurate identification of individuals with viral infections is crucial for clinical management and public health interventions. We aimed to assess the ability of transcriptomic biomarkers to identify naturally acquired respiratory viral infection before typical symptoms are present. Methods In this index-cluster study, we prospectively recruited a cohort of undergraduate students (aged 18–25 years) at Duke University (Durham, NC, USA) over a period of 5 academic years. To identify index cases, we monitored students for the entire academic year, for the presence and severity of eight symptoms of respiratory tract infection using a daily web-based survey, with symptoms rated on a scale of 0–4. Index cases were defined as individuals who reported a 6-point increase in cumulative daily symptom score. Suspected index cases were visited by study staff to confirm the presence of reported symptoms of illness and to collect biospecimen samples. We then identified clusters of close contacts of index cases (ie, individuals who lived in close proximity to index cases, close friends, and partners) who were presumed to be at increased risk of developing symptomatic respiratory tract infection while under observation. We monitored each close contact for 5 days for symptoms and viral shedding and measured transcriptomic responses at each timepoint each day using a blood-based 36-gene RT-PCR assay. Findings Between Sept 1, 2009, and April 10, 2015, we enrolled 1465 participants. Of 264 index cases with respiratory tract infection symptoms, 150 (57%) had a viral cause confirmed by RT-PCR. Of their 555 close contacts, 106 (19%) developed symptomatic respiratory tract infection with a proven viral cause during the observation window, of whom 60 (57%) had the same virus as their associated index case. Nine viruses were detected in total. The transcriptomic assay accurately predicted viral infection at the time of maximum symptom severity (mean area under the receiver operating characteristic curve [AUROC] 0·94 [95% CI 0·92–0·96]), as well as at 1 day (0·87 [95% CI 0·84–0·90]), 2 days (0·85 [0·82–0·88]), and 3 days (0·74 [0·71–0·77]) before peak illness, when symptoms were minimal or absent and 22 (62%) of 35 individuals, 25 (69%) of 36 individuals, and 24 (82%) of 29 individuals, respectively, had no detectable viral shedding. Interpretation Transcriptional biomarkers accurately predict and diagnose infection across diverse viral causes and stages of disease and thus might prove useful for guiding the administration of early effective therapy, quarantine decisions, and other clinical and public health interventions in the setting of endemic and pandemic infectious diseases. Funding US Defense Advanced Research Projects Agency.

Patient management relies on diagnostic information to identify appropriate treatment. Standard evaluations of diagnostic tests consist of estimating sensitivity, specificity, positive/negative predictive values, likelihood ratios, and accuracy. Although useful, these metrics do not convey the tests’ clinical value, which is critical to informing decision-making. Full appreciation of the clinical impact of a diagnostic test requires analyses that integrate sensitivity and specificity, account for the disease prevalence within the population of test application, and account for the relative importance of specificity vs sensitivity by considering the clinical implications of false-positive and false-negative results. We developed average weighted accuracy (AWA), representing a pragmatic metric of diagnostic yield or global utility of a diagnostic test. AWA can be used to compare test alternatives, even across different studies. We apply the AWA methodology to evaluate a new diagnostic test developed in the Rapid Diagnostics in Categorizing Acute Lung Infections (RADICAL) study.

RationaleAsthma exacerbations often occur due to infectious triggers, but determining whether infection is present and whether it is bacterial or viral remains clinically challenging. A diagnostic strategy that clarifies these uncertainties could enable personalized asthma treatment and mitigate antibiotic overuse.ObjectivesTo explore the performance of validated peripheral blood gene expression signatures in discriminating bacterial, viral, and noninfectious triggers in subjects with asthma exacerbations.MethodsSubjects with suspected asthma exacerbations of various etiologies were retrospectively selected for peripheral blood gene expression analysis from a pool of subjects previously enrolled in emergency departments with acute respiratory illness. RT-PCR quantified 87 gene targets, selected from microarray-based studies, followed by logistic regression modeling to define bacterial, viral, or noninfectious class. The model-predicted class was compared to clinical adjudication and procalcitonin.ResultsOf 46 subjects enrolled, 7 were clinically adjudicated as bacterial, 18 as viral, and 21 as noninfectious. Model prediction was congruent with clinical adjudication in 15/18 viral and 13/21 noninfectious cases, but only 1/7 bacterial cases. None of the adjudicated bacterial cases had confirmatory microbiology; the precise etiology in this group was uncertain. Procalcitonin classified only one subject in the cohort as bacterial. 47.8% of subjects received antibiotics.ConclusionsOur model classified asthma exacerbations by the underlying bacterial, viral, and noninfectious host response. Compared to clinical adjudication, the majority of discordances occurred in the bacterial group, due to either imperfect adjudication or model misclassification. Bacterial infection was identified infrequently by all classification schemes, but nearly half of subjects were prescribed antibiotics. A gene expression-based approach may offer useful diagnostic information in this population and guide appropriate antibiotic use.

Purpose To accurately segment organs from 3D CT image volumes using a 2D, multi-channel SegNet model consisting of a deep Convolutional Neural Network (CNN) encoder-decoder architecture. Method We trained a SegNet model on the extended cardiac-torso (XCAT) dataset, which was previously constructed based on patient Chest–Abdomen–Pelvis (CAP) Computed Tomography (CT) studies from 50 Duke patients. Each study consists of one low-resolution (5-mm section thickness) 3D CT image volume and its corresponding 3D, manually labeled volume. To improve modeling on such small sample size regime, we performed median frequency class balancing weighting in the loss function of the SegNet, data normalization adjusting for intensity coverage of CT volumes, data transformation to harmonize voxel resolution, CT section extrapolation to virtually increase the number of transverse sections available as inputs to the 2D multi-channel model, and data augmentation to simulate mildly rotated volumes. To assess model performance, we calculated Dice coefficients on a held-out test set, as well as qualitative evaluation of segmentation on high-resolution CTs. Further, we incorporated 50 patients high-resolution CTs with manually-labeled kidney segmentation masks for the purpose of quantitatively evaluating the performance of our XCAT trained segmentation model. The entire study was conducted from raw, identifiable data within the Duke Protected Analytics Computing Environment (PACE). Result We achieved median Dice coefficients over 0.8 for most organs and structures on XCAT test instances and observed good performance on additional images without manual segmentation labels, qualitatively evaluated by Duke Radiology experts. Moreover, we achieved 0.89 median Dice Coefficients for kidneys on high-resolution CTs. Conclusion 2D, multi-channel models like SegNet are effective for organ segmentations of 3D CT image volumes, achieving high segmentation accuracies.

Many researchers in the field of machine learning have addressed the problem of detecting anomalies within Computed Tomography (CT) scans. Training these machine learning algorithms requires a dataset of CT scans with identified anomalies (labels), usually, in specific organs. This represents a problem, since it requires experts to review thousands of images in order to create labels for these data. We aim to decrease human burden at labeling CT scans by developing a model that identifies anomalies within plain-text-based reports that then could be further used as a method to create labels for models based on CT scans. This study contains more than 4800 CT reports from Duke Health System, for which we aim to identify organ specific abnormalities. We propose an iterative active learning approach that consists of building a machine learning model to classify CT reports by abnormalities in different organs and then improving it by actively adding reports sequentially. At each iteration, clinical experts review the report that provides the model with highest expected information gain. This process is done in real time by using a web interface. Then, this datum is used by the model to improve its performance. We evaluated the performance of our method for abnormalities in kidneys and lungs. When starting with a model trained on 99 reports, the results show the model achieves an Area Under the Curve (AUC) score of 0.93 on the test set after adding 130 actively labeled reports to the model from an unlabeled pool of 4,000. This suggests that a set of labeled CT scans can be obtained with significantly reduced human work by combining machine learning techniques and clinical experts' knowledge.

We consider machine-learning-based thyroid-malignancy prediction from cytopathology whole-slide images (WSI). Multiple instance learning (MIL) approaches, typically used for the analysis of WSIs, divide the image (bag) into patches (instances), which are used to predict a single bag-level label. These approaches perform poorly in cytopathology slides due to a unique bag structure: sparsely located informative instances with varying characteristics of abnormality. We address these challenges by considering multiple types of labels: bag-level malignancy and ordered diagnostic scores, as well as instance-level informativeness and abnormality labels. We study their contribution beyond the MIL setting by proposing a maximum likelihood estimation (MLE) framework, from which we derive a two-stage deep-learning-based algorithm. The algorithm identifies informative instances and assigns them local malignancy scores that are incorporated into a global malignancy prediction. We derive a lower bound of the MLE, leading to an improved training strategy based on weak supervision, that we motivate through statistical analysis. The lower bound further allows us to extend the proposed algorithm to simultaneously predict multiple bag and instance-level labels from a single output of a neural network. Experimental results demonstrate that the proposed algorithm provides competitive performance compared to several competing methods, achieves (expert) human-level performance, and allows augmentation of human decisions.

Machine learning models for radiology benefit from large-scale data sets with high quality labels for abnormalities. We curated and analyzed a chest computed tomography (CT) data set of 36,316 volumes from 19,993 unique patients. This is the largest multiply-annotated volumetric medical imaging data set reported. To annotate this data set, we developed a rule-based method for automatically extracting abnormality labels from free-text radiology reports with an average F-score of 0.976 (min 0.941, max 1.0). We also developed a model for multi-organ, multi-disease classification of chest CT volumes that uses a deep convolutional neural network (CNN). This model reached a classification performance of AUROC greater than 0.90 for 18 abnormalities, with an average AUROC of 0.773 for all 83 abnormalities, demonstrating the feasibility of learning from unfiltered whole volume CT data. We show that training on more labels improves performance significantly: for a subset of 9 labels - nodule, opacity, atelectasis, pleural effusion, consolidation, mass, pericardial effusion, cardiomegaly, and pneumothorax - the model's average AUROC increased by 10% when the number of training labels was increased from 9 to all 83. All code for volume preprocessing, automated label extraction, and the volume abnormality prediction model will be made publicly available. The 36,316 CT volumes and labels will also be made publicly available pending institutional approval., Comment: 20 pages, 3 figures, 5 tables (appendices additional). Published in Medical Image Analysis (October 2020)

Background and objectives Although the burden of non-alcoholic fatty liver disease (NAFLD) continues to increase worldwide, genetic factors predicting progression to cirrhosis and decompensation in NAFLD remain poorly understood. We sought to determine whether gene expression profiling was associated with clinical decompensation and death in patients with NAFLD, and to assess whether altered DNA methylation contributes to these changes in gene expression. Methods We performed a retrospective analysis of 86 patients in the Duke NAFLD Clinical Database and Biorepository with biopsy-proven NAFLD whose liver tissue was previously evaluated for gene expression and DNA methylation using array based technologies. We assessed the prospective development of liver and cardiovascular disease related outcomes, including hepatic decompensation as identified by the development of ascites, hepatic encephalopathy, hepatocellular carcinoma, or variceal bleeding as well as stroke and myocardial infarction via medical chart review. Results Of the 86 patients, 47 had F0-F1 fibrosis and 39 had F3-F4 fibrosis at index liver biopsy. Gene expression probe sets (n = 54,675) were analyzed; 42 genes showed significant differential expression (p

Many deep learning architectures have been proposed to model the compositionality in text sequences, requiring a substantial number of parameters and expensive computations. However, there has not been a rigorous evaluation regarding the added value of sophisticated compositional functions. In this paper, we conduct a point-by-point comparative study between Simple Word-Embedding-based Models (SWEMs), consisting of parameter-free pooling operations, relative to word-embedding-based RNN/CNN models. Surprisingly, SWEMs exhibit comparable or even superior performance in the majority of cases considered. Based upon this understanding, we propose two additional pooling strategies over learned word embeddings: (i) a max-pooling operation for improved interpretability; and (ii) a hierarchical pooling operation, which preserves spatial (n-gram) information within text sequences. We present experiments on 17 datasets encompassing three tasks: (i) (long) document classification; (ii) text sequence matching; and (iii) short text tasks, including classification and tagging. The source code and datasets can be obtained from https:// github.com/dinghanshen/SWEM., To appear at ACL 2018 (code: https://github.com/dinghanshen/SWEM)

OBJECTIVE: To find and validate generalizable sepsis subtypes using data-driven clustering. DESIGN: We used advanced informatics techniques to pool data from 14 bacterial sepsis transcriptomic datasets from 8 different countries (N=700). SETTING: Retrospective analysis. SUBJECTS: Persons admitted to the hospital with bacterial sepsis. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: A unified clustering analysis across 14 discovery datasets revealed three subtypes, which, based on functional analysis, we termed Inflammopathic, Adaptive, and Coagulopathic. We then validated these subtypes in 9 independent datasets from 5 different countries (N=600). In both discovery and validation data, the Adaptive subtype is associated with a lower clinical severity and lower mortality rate, and the Coagulopathic subtype is associated with higher mortality and clinical coagulopathy. Further, these clusters are statistically associated with clusters derived by others in independent single sepsis cohorts. CONCLUSIONS: The three sepsis subtypes may represent a unifying framework for understanding the molecular heterogeneity of the sepsis syndrome. Further study could potentially enable a precision-medicine approach of matching novel immunomodulatory therapies with septic patients most likely to benefit.

Improved risk stratification and prognosis prediction in sepsis is a critical unmet need. Clinical severity scores and available assays such as blood lactate reflect global illness severity with suboptimal performance, and do not specifically reveal the underlying dysregulation of sepsis. Here, we present prognostic models for 30-day mortality generated independently by three scientific groups by using 12 discovery cohorts containing transcriptomic data collected from primarily community-onset sepsis patients. Predictive performance is validated in five cohorts of community-onset sepsis patients in which the models show summary AUROCs ranging from 0.765–0.89. Similar performance is observed in four cohorts of hospital-acquired sepsis. Combining the new gene-expression-based prognostic models with prior clinical severity scores leads to significant improvement in prediction of 30-day mortality as measured via AUROC and net reclassification improvement index These models provide an opportunity to develop molecular bedside tests that may improve risk stratification and mortality prediction in patients with sepsis., Sepsis is characterized by deregulated host response to infection. Efficient therapies are still needed but a limitation for sepsis treatment is the heterogeneity in patients. Here Sweeney et al. generate prognostic models based on gene expression to improve risk stratification classification and prediction for 30-day mortality of patients.

Word embeddings are effective intermediate representations for capturing semantic regularities between words, when learning the representations of text sequences. We propose to view text classification as a label-word joint embedding problem: each label is embedded in the same space with the word vectors. We introduce an attention framework that measures the compatibility of embeddings between text sequences and labels. The attention is learned on a training set of labeled samples to ensure that, given a text sequence, the relevant words are weighted higher than the irrelevant ones. Our method maintains the interpretability of word embeddings, and enjoys a built-in ability to leverage alternative sources of information, in addition to input text sequences. Extensive results on the several large text datasets show that the proposed framework outperforms the state-of-the-art methods by a large margin, in terms of both accuracy and speed., Comment: Published in ACL 2018; Code: https://github.com/guoyinwang/LEAM

Background The heart is a metabolically active organ, and plasma acylcarnitines are associated with long-term risk for myocardial infarction. We hypothesized that myocardial ischemia from cardiac stress testing will produce dynamic changes in acylcarnitine and amino acid levels compared to levels seen in matched control patients with normal stress tests. Methods We analyzed targeted metabolomic profiles in a pilot study of 20 case patients with inducible ischemia on stress testing from an existing prospectively collected repository of 357 consecutive patients presenting with symptoms of Acute Coronary Syndrome (ACS) in an Emergency Department (ED) observation unit between November 2012 and September 2014. We selected 20 controls matched on age, sex, and body-mass index (BMI). A peripheral blood sample was drawn

Improved risk stratification and prognosis in sepsis is a critical unmet need. Clinical severity scores and available assays such as blood lactate reflect global illness severity with suboptimal performance, and do not specifically reveal the underlying dysregulation of sepsis. Here three scientific groups were invited to independently generate prognostic models for 30-day mortality using 12 discovery cohorts (N=650) containing transcriptomic data collected from primarily community-onset sepsis patients. Predictive performance was validated in 5 cohorts of community-onset sepsis patients (N=189) in which the models showed summary AUROCs ranging from 0.765-0.89. Similar performance was observed in 4 cohorts of hospital-acquired sepsis (N=282). Combining the new gene-expression-based prognostic models with prior clinical severity scores led to significant improvement in prediction of 30-day mortality (p

We introduce a novel dynamic model for discrete time-series data, in which the temporal sampling may be nonuniform. The model is specified by constructing a hierarchy of Poisson factor analysis blocks, one for the transitions between latent states and the other for the emissions between latent states and observations. Latent variables are binary and linked to Poisson factor analysis via Bernoulli-Poisson specifications. The model is derived for count data but can be readily modified for binary observations. We derive efficient inference via Markov chain Monte Carlo, that scales with the number of non-zeros in the data and latent binary states, yielding significant acceleration compared to related models. Experimental results on benchmark data show the proposed model achieves state-of-the-art predictive performance. Additional experiments on microbiome data demonstrate applicability of the proposed model to interesting problems in computational biology where interpretability is of utmost importance.