An official website of the United States government
Cough detection can provide an important marker to monitor chronic respiratory conditions. However, manual techniques which require human expertise to count coughs are both expensive and time-consuming. Recent Automatic Cough Detection Algorithms (ACDAs) have shown promise to meet clinical monitoring requirements, but only in recent years they have made their way to non-clinical settings due to the required portability of sensing technologies and the extended duration of data recording. More precisely, these ACDAs operate at high sampling frequencies, which leads to high power consumption and computing requirements, making these difficult to implement on a wearable device. Additionally, reproducibility of their performance is essential. Unfortunately, as the majority of ACDAs were developed using private clinical data, it is difficult to reproduce their results. We, hereby, present an ACDA that meets clinical monitoring requirements and reliably operates at a low sampling frequency. This ACDA is implemented using a convolutional neural network (CNN), and publicly available data. It achieves a sensitivity of 92.7%, a specificity of 92.3%, and an accuracy of 92.5% using a sampling frequency of just 750 Hz. We also show that a low sampling frequency allows us to preserve patients’ privacy by obfuscating their speech, and we analyze the trade-off between speech obfuscation for privacy and cough detection accuracy. Clinical relevance—This paper presents a new cough detection technique and preliminary analysis on the trade-off between detection accuracy and obfuscation of speech for privacy. These findings indicate that, using a publicly available dataset, we can sample signals at 750 Hz while still maintaining a sensitivity above 90%, suggested to be sufficient for clinical monitoring [1].
more »
« less
Audio-Based Cough Detection in Clinic Waiting Rooms
Automated cough detection has significant applications for the surveillance of diseases and supports medical decisions, as cough sounds can be a useful biomarker. However, the implementation and evaluation of robust cough detection models can be challenging due to the lack of real-world data. This paper introduces and makes available a collection of 2,883 coughs and 3,074 non-cough sounds recorded in clinic waiting rooms that we hope will become a baseline for this task. Using this dataset, we evaluate different convolutional network architectures for classifying short audio segments as cough or non-cough. An ensemble model of convolutional neuronal networks provides the most robust performance and has a ROC AUC of $$98.1\%$$. Equally important, we construct a cough counter that incorporates the ensemble model to compute the number of coughs per day. Then, a simple linear model estimates the number of visits in which the patients report cough symptoms from the cough counts. This simple regression model can predict the number of cough visits in the clinic with an absolute mean error of 4.26 cough visits per day. Using additional information about when patients are in the clinic helps a similar regression model reach a mean absolute error of 3.65 cough visits per day. These results demonstrate the feasibility of using cough detection as a biomarker for the spread of respiratory viruses within the community.
more »
« less
- PAR ID:
- 10337862
- Date Published:
- Journal Name:
- IEEE International Conference on Healthcare Informatics
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Background:The Glycemia Risk Index (GRI) was developed in adults with diabetes and is a validated metric of quality of glycemia. Little is known about the relationship between GRI and type 1 diabetes (T1D) self-management habits, a validated assessment of youths’ engagement in habits associated with glycemic outcomes. Method:We retrospectively examined the relationship between GRI and T1D self-management habits in youth with T1D who received care from a Midwest pediatric diabetes clinic network. The GRI was calculated using seven days of continuous glucose monitor (CGM) data, and T1D self-management habits were assessed ±seven days from the GRI score. A mixed-effects Poisson regression model was used to evaluate the total number of habits youth engaged in with GRI, glycated hemoglobin A1c (HbA1c), age, race, ethnicity, and insurance type as fixed effects and participant ID as a random effect to account for multiple clinic visits per individual. Results:The cohort included 1182 youth aged 2.5 to 18.0 years (mean = 13.8, SD = 3.5) comprising 50.8% male, 84.6% non-Hispanic White, and 64.8% commercial insurance users across a total of 6029 clinic visits. Glycemia Risk Index scores decreased as total number of habits performed increased, suggesting youth who performed more self-management habits achieved a higher quality of glycemia. Conclusions:In youth using CGMs, GRI may serve as an easily obtainable metric to help identify youth with above target glycemia, and engagement/disengagement in the T1D self-management habits may inform clinicians with suitable interventions for improving glycemic outcomes.more » « less
-
Introduction:The Virtual Diabetes Specialty Clinic (VDiSC) study demonstrated the feasibility of providing comprehensive diabetes care entirely virtually by combining virtual visits with continuous glucose monitoring support and remote patient monitoring (RPM). However, the financial sustainability of this model remains uncertain. Methods:We developed a financial model to estimate the variable costs and revenues of virtual diabetes care, using visit data from the 234 VDiSC participants with type 1 or type 2 diabetes. Data included virtual visits with certified diabetes care and education specialists (CDCES), endocrinologists, and behavioral health services (BHS). The model estimated care utilization, variable costs, reimbursement revenue, gross profit, and gross profit margin per member, per month (PMPM) for privately insured, publicly insured, and overall clinic populations (75% privately insured). We performed two-way sensitivity analyses on key parameters. Results:Gross profit and gross profit margin PMPM (95% confidence interval) were estimated at $−4 ($−14.00 to $5.68) and −4% (−3% to −6%) for publicly insured patients; $267.26 ($256.59-$277.93) and 73% (58%-88%) for privately insured patients; and $199.41 ($58.43-$340.39) and 67% (32%-102%) for the overall clinic. Profits were primarily driven by CDCES visits and RPM. Results were sensitive to insurance mix, cost-to-charge ratio, and commercial-to-Medicare price ratio. Conclusions:Virtual diabetes care can be financially viable, although profitability relies on privately insured patients. The analysis excluded fixed costs of clinic infrastructure, and securing reimbursement may be challenging in practice. The financial model is adaptable to various care settings and can serve as a planning tool for virtual diabetes clinics.more » « less
-
Background:Despite recent advances, patients with heart failure (HF) often experience repeat hospitalizations and worsening clinical trajectories from inadequate decongestion. Evidence-based approaches for optimizing interventions in the acute hospital setting for patients with decompensated HF are needed. We evaluated whether machine learning (ML) models can accurately predict next-day levels for decongestion surrogates in hospitalized HF patients. Hypothesis:ML can accurately predict body weight, hematocrit, creatinine, and potassium values in the next 24 hours in hospitalized HF patients. Methods:We utilized national Veterans Affairs (VA) databases to study all patients admitted with HF from January 2014 to July 2022. Records including at least one value for at least one biomarker of interest (body weight, hematocrit, creatinine, and potassium) were included. Patients were randomly split into training (80%), validation (10%), and test (10%) datasets. We trained a recurrent neural network to predict each biomarker’s value on admission day n+1 using data until day n, simulating a scenario where a clinician monitors response to treatment (e.g., diuresis) over a 24-hour cycle. The model that performed best on the validation set was evaluated on the test set. The R2, mean absolute error (MAE), and feature importance were determined. Results:We identified 589,114 admissions involving 124,163 unique patients. The mean (SD) age on admission was 72 (10) years; 98% were male, 69% were white, and 25% were Black. The performance (R2, MAE) for each biomarker model was as follows: body weight (0.94, 6.15 lb.), creatinine (0.92, 0.21 mg/dL), hematocrit (0.86, 1.7%), and potassium (0.53, 0.27 mmol/L). The top predictive features across all models were intravenous or oral diuretic use, patient age, and diastolic blood pressure. The predicted 24-hour change in each biomarker based on total daily diuretic dose for five representative patients is demonstrated in the Figure. Conclusions:ML can accurately predict the 24-hour body weight, hematocrit, creatinine, and potassium values in hospitalized HF patients, suggesting the potential for AI to guide acute in-hospital management.more » « less
-
Foot traffic is a business term to describe the number of customers that enter a point of interest (POI). This work aims to predict future foot traffic: the number of people from each census block group (CBG) that will visit each POI of a study region with potential applications in marketing and advertising. Existing techniques for spatiotemporal prediction of foot traffic use location-based social network data that suffer from sparsity, capturing only a handful of visits per day. This study utilizes highly granular foot traffic data from SafeGraph, a data company that collects mobility data regarding hundreds of millions of visits per day in the United States alone. Using this data, we explore solutions to predict weekly foot traffic data at the POI level. We propose a collaborative filtering approach using tensor factorization on the (POIs x CBGs x Weeks) data tensor. This approach provides us with a de-noised estimation of visits in previous weeks for all POI-CBG pairs. Using this tensor, we explore various time series prediction models: weekly rolling average, weighted weekly rolling average, univariate linear regression, polynomial regression, and long short-term memory (LSTM) recurrent neural networks. Our results show that of all the prediction models, the collaborative filtering step consistently improves prediction results. We also found that a simple weighted average consistently performed better than the more sophisticated approaches. Given this abundance of foot traffic data, this result shows that we can improve the spatiotemporal prediction of foot traffic data by harnessing collaborative filtering.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
