This study pretends to evaluate the potential use of Hyfe Cough Tracker (Hyfe) to screen for, diagnose, and support the clinical management of patients with respiratory diseases, while enriching a dataset of disease-specific annotated coughs, for further refinement of similar systems.
This is an observational study that will take place in the two campuses of the Clínica Universidad de Navarra, located in Pamplona and Madrid (Spain). An Artificial-Intelligence system (AI) that detects and records explosive putative cough sounds and identifies human cough based on acoustic characteristics will be used to automatically monitor cough. Potential participants either attending the outpatient clinic or hospitalised with a complaint of cough will be invited by their treating physician, or a member of the research team, and included in the study by part of the research team. A researcher will instruct participants on how to install and use Hyfe Cough Tracker in their smartphones. Participants will be monitored for 30 days (outpatients) or until discharged from the hospital (inpatients). Participants will be asked to complete a daily, online, standardised 100 mm visual analogue scale (VAS) to register changes in the subjective intensity of their cough, while using Hyfe to objectively monitor changes in its frequency. In parallel, a dataset of annotated cough sounds will be constructed and retrospectively used to assess differences in acoustic patterns of cough, and to evaluate the performance of the system detecting them. A first subgroup of participants will be recruited outside the clinical setting and asked to provide a series of elicited sounds, including coughs, which will then be used to determine the system's performance accurately discriminating coughs from non-cough sounds, and compared to trained human listeners. A second subgroup of participants will be will be instructed to use Hyfe, and the related Hyfe Air wearable device continuously for a period between 6 and 24 hours, while they record themselves using a MP3 recorder connected to a lapel microphone. This group will be used to evaluate the performance of Hyfe and Hyfe Air in a real-life setting, with spontaneous coughs.
Study Type
OBSERVATIONAL
Enrollment
616
Hyfe Cough Tracker is a digital acoustic surveillance system that uses an artificial intelligence system to discriminate cough from non-cough sounds. Hyfe is an AI-enabled mobile app that records short snippets (\<0.5 seconds) of putative cough explosive sounds and then classifies them as cough or non-cough using a convolutional neural network (CNN) model. Briefly, the acoustic characteristics of recorded sounds are converted into an image file, which is then processed by an algorithm trained to identify graphical differences in images. This creates an adjustable prediction score, with values above it, resulting in a sound being classified as "cough", and those below being classified as "non-cough.
Hyfe Air is a wearable device with an incorporated wireless lapel microphone. The device´s recordings can be run through the same cough-detection algorithm used by Hyfe Cough Tracker, while its results are directly stored in a remote database and are not displayed to participants.
Clinica Universidad de Navarra
Pamplona, Navarre, Spain
Correlation between subjective perception of cough and objective frequency
The daily VAS score of participants will be compared to the cough frequency registered by the cough surveillance system. These data will be used to fit a linear regression model to compare self-reported VAS scores to daily cough frequency and calculate a correlation coefficient (r).
Time frame: 6 months.
Sensitivity of the system discriminating coughs
The sensitivity of Hyfe for the discrimination of coughs from other explosive sounds will be compared to that of trained human listeners. Sensitivity will be reported as the proportion of sounds correctly identified as coughs (true positives), from the total cough sounds produced (true positives + false negatives).
Time frame: 6 months.
Specificity of the system discriminating coughs
The specificity of Hyfe for the discrimination of coughs from other explosive sounds will be compared to that of trained human listeners. Specificity will be defined as the proportion of non-cough sounds correctly identified by the system (true negatives) from the total non-cough sounds produced (true negatives + false positives)
Time frame: 6 months.
Positive predictive value (PPV) of the system discriminating coughs
The PPV of Hyfe for the discrimination of coughs from other explosive sounds will be compared to that of trained human listeners. PPV will be defined as the proportion of cough sounds correctly identified by the system (true positives) from the total sounds labelled as coughs (true positives + false positives).
Time frame: 6 months.
Negative predictive value (NPV) of the system discriminating coughs
The NPV of Hyfe for the discrimination of coughs from other explosive sounds will be compared to that of trained human listeners. NPV will be defined as the proportion of non-cough sounds correctly identified by the system (true negatives) from the total of sounds labelled as non-coughs (true negatives+ false negatives).
Time frame: 6 months.
Construction of an annotated cough dataset
Cough registries of participants with an etiologic diagnosis will be annotated and stored to create a dataset that can be used for further algorithm training and refinement.
Time frame: 5 years.
Sensitivity of the system differentiating coughs caused by different conditions
The records obtained from participants for which an etiologic diagnosis is reached before the end of the study will be analysed to detect differential acoustic patterns, which will in turn be used to train the system's convolutional neural network to perform respiratory disease cough classification. The performance of this system will be retrospectively evaluated by determining its sensitivity for the diagnosis of different respiratory conditions, compared to clinical diagnoses made by a physician. Sensitivity will be defined as the proportion of participants in which Hyfe reaches a correct diagnoses based on cough acoustic patterns (true positives) from the total number of participants diagnosed with a certain condition (true positives + false negatives).
Time frame: 5 years.
Specificity of the system differentiating coughs caused by different conditions
The records obtained from participants for which an etiologic diagnosis is reached before the end of the study will be analysed to detect differential acoustic patterns, which will in turn be used to train the system's convolutional neural network to perform respiratory disease cough classification. The performance of this system will be retrospectively evaluated by determining its specificity for the diagnosis of different respiratory conditions, compared to clinical diagnoses made by a physician. Specificity will be defined as the proportion of participants in which Hyfe correctly identifies the absence of acoustic cough patterns associated to a certain disease (true negatives), from the total of participants without that specific condition (true negatives+ false positives).
Time frame: 5 years.
This platform is for informational purposes only and does not constitute medical advice. Always consult a qualified healthcare professional.