Objectives: Pediatric asthma is a leading cause of emergency department (ED) utilization and hospitalization.Earlier identification of need for hospital-level care could triage patients more efficiently to high-or low-resource ED tracks. Existing tools to predict disposition for pediatric asthma use only clinical data, perform best several hours into the ED stay, and are static or score-based. Machine learning offers a population-specific, dynamic option that allows real-time integration of available nonclinical data at triage. Our objective was to compare the performance of four common machine learning approaches, incorporating clinical data available at the time of triage with information about weather, neighborhood characteristics, and community viral load for early prediction of the need for hospital-level care in pediatric asthma.Methods: Retrospective analysis of patients ages 2 to 18 years seen at two urban pediatric EDs with asthma exacerbation over 4 years. Asthma exacerbation was defined as receiving both albuterol and systemic corticosteroids. We included patient features, measures of illness severity available in triage, weather features, and Centers for Disease Control and Prevention influenza patterns. We tested four models: decision trees, LASSO logistic regression, random forests, and gradient boosting machines. For each model, 80% of the data set was used for training and 20% was used to validate the models. The area under the receiver operating characteristic (AUC) curve was calculated for each model.Results: There were 29,392 patients included in the analyses: mean (AESD) age of 7.0 (AE4.2) years, 42% female, 77% non-Hispanic black, and 76% public insurance. The AUCs for each model were: decision tree 0.72 (95% confidence interval [CI] = 0.66-0.77), logistic regression 0.83 (95% CI = 0.82-0.83), random forests 0.82 (95% CI = 0.81-0.83), and gradient boosting machines 0.84 (95% CI = 0.83-0.85). In the lowest decile of risk, only 3% of patients required hospitalization; in the highest decile this rate was 100%. After patient vital signs and acuity, age and weight, followed by socioeconomic status (SES) and weather-related features, were the most important for predicting hospitalization.Conclusions: Three of the four machine learning models performed well with decision trees preforming the worst. The gradient boosting machines model demonstrated a slight advantage over other approaches at predicting need for hospital-level care at the time of triage in pediatric patients presenting with asthma exacerbation. The addition of weight, SES, and weather data improved the performance of this model.

The analysis of lung sounds, collected through auscultation, is a fundamental component of pulmonary disease diagnostics for primary care and general patient monitoring for telemedicine. Despite advances in computation and algorithms, the goal of automated lung sound identification and classification has remained elusive. Over the past 40 years, published work in this field has demonstrated only limited success in identifying lung sounds, with most published studies using only a small numbers of patients (typically N<;20) and usually limited to a single type of lung sound. Larger research studies have also been impeded by the challenge of labeling large volumes of data, which is extremely labor-intensive. In this paper, we present the development of a semi-supervised deep learning algorithm for automatically classify lung sounds from a relatively large number of patients (N=284). Focusing on the two most common lung sounds, wheeze and crackle, we present results from 11,627 sound files recorded from 11 different auscultation locations on these 284 patients with pulmonary disease. 890 of these sound files were labeled to evaluate the model, which is significantly larger than previously published studies. Data was collected with a custom mobile phone application and a low-cost (US$30) electronic stethoscope. On this data set, our algorithm achieves ROC curves with AUCs of 0.86 for wheeze and 0.74 for crackle. Most importantly, this study demonstrates how semi-supervised deep learning can be used with larger data sets without requiring extensive labeling of data.

Pulmonary and respiratory diseases (e.g. asthma, COPD, allergies, pneumonia, tuberculosis, etc.) represent a large proportion of the global disease burden, mortality, and disability. In this context of creating automated diagnostic tools, we explore how the analysis of voluntary cough sounds may be used to screen for pulmonary disease. As a clinical study, voluntary coughs were recorded using a custom mobile phone stethoscope from 54 patients, of which 7 had COPD, 15 had asthma, 11 had allergic rhinitis, 17 had both asthma and allergic rhinitis, and four had both COPD and allergic rhinitis. Data were also collected from 33 healthy subjects. These patients also received full auscultation at 11 sites, given a clinical questionnaire, and underwent full pulmonary function testing (spirometer, body plethysmograph, DLCO) which culminated in a diagnosis provided by an experienced pulmonologist. From machine learning analysis of these data, we show that it is possible to achieve good classification of cough sounds in terms of Wet vs Dry, yielding an ROC curve with AUC of 0.94, and show that voluntary coughs can serve as an effective test for determining Healthy vs Unhealthy (sensitivity=35.7% specificity=100%). We also show that the use of cough sounds can enhance the performance of other diagnostic tools such as a patient questionnaire and peak flow meter; however voluntary coughs alone provide relatively little value in determining specific disease diagnosis.

No abstract

The remote measurement of heart rate (HR) and heart rate variability (HRV) via a digital camera (video plethysmography) has emerged as an area of great interest for biomedical and health applications. While a few implementations of video plethysmography have been demonstrated on smart phones under controlled lighting conditions, it has been challenging to create a general scalable solution due to the large variability in smart phone hardware performance, software architecture, and the variable response to lighting parameters. In this context, we present a selfcontained smart phone implementation of video plethysmography for Android OS, which employs both stochastic and deterministic algorithms, and we use this to study the effect of lighting parameters (illuminance, color spectrum) on the accuracy of the remote HR measurement. Using two different phone models, we present the median HR error for five different video plethysmography algorithms under three different types of lighting (natural sunlight, compact fluorescent, and halogen incandescent) and variations in brightness. For most algorithms, we found the optimum light brightness to be in the range 1000-4000 lux and the optimum lighting types to be compact fluorescent and natural light. Moderate errors were found for most algorithms with some devices under conditions of low-brightness (<;500 lux) and highbrightness (>4000 lux). Our analysis also identified camera frame rate jitter as a major source of variability and error across different phone models, but this can be largely corrected through non-linear resampling. Based on testing with six human subjects, our real-time Android implementation successfully predicted the measured HR with a median error of -0.31 bpm, and an inter-quartile range of 2.1bpm.