Validation of the LEOSound Cough Detection Algorithm

Background Cough is an important respiratory symptom being of great interest to many researchers. Up to now, most knowledge about cough has been collected through standardized questionnaires. Objective, and reliable detection of cough assessed by automated lung sound monitoring are becoming increasingly important. The aim of this study is to validate the LEOSound lung sound monitor by using previously determined and investigated COPD datasets (1,2). Methods Based on multiple recordings of 48 patients with stable COPD II-IV, we validated the cough detection algorithm of LEOSound by using a contingency table. Sensitivity, specificity, positive and negative predictive values were used as quantitative measures. We found the overall accuracy to be 87.3% with sensitivity and specificity of 98.7% and 80.2%, respectively. Major reasons for midsections in descending order were throat cleaning, snoring and movement artifacts.


Abstract
Background Cough is an important respiratory symptom being of great interest to many researchers.
Up to now, most knowledge about cough has been collected through standardized questionnaires. Objective, and reliable detection of cough assessed by automated lung sound monitoring are becoming increasingly important. The aim of this study is to validate the LEOSound lung sound monitor by using previously determined and investigated COPD datasets (1,2).

Methods
Based on multiple recordings of 48 patients with stable COPD II-IV, we validated the cough detection algorithm of LEOSound by using a contingency table. Sensitivity, specificity, positive and negative predictive values were used as quantitative measures.

Results
We found the overall accuracy to be 87.3% with sensitivity and specificity of 98.7% and 80.2%, respectively. Major reasons for midsections in descending order were throat cleaning, snoring and movement artifacts.

Conclusion
In comparison to other full-automated cough monitoring systems, the LEOSound performs the best in sensitivity, but shows slightly poor specificity. Misdetections were mainly caused due to morphological similar noises and can be withdrawn while scanning through the recording manually.

Background
The acoustic symptom cough is defined as a characteristic explosive sound. The typical noise arises due to turbulence of the outflowing air, vibrations of the tissue and movement 3 of liquid through the respiratory tract. The volume of the sound depends on the airflow rate that is achieved while coughing, the density of respiratory air, and of the dimension of the airways. (3,4) Cough as a major symptom of Chronic Obstructive Pulmonary Disease (COPD), one of the leading causes of death worldwide(5,6) and many other diseases(4) is of great interest to a variety of different research fields. Up to now, research in cough mostly relied on questionnaires such as the Leicester Cough Questionnaire (LCQ) (7,8), Cough-specific quality of life questionnaire (CQLQ) (9) or in the case of COPD the COPD Assessment Test (CAT). Over the last decade, systems that record and automatically detect cough have been developed and used by several research groups. Such systems have enabled researches and physicians to objectively analyze the symptom cough, which might be showing more different aspects of cough, compared to the subjectively assessed information by questionnaires (10). Consequently, the European Respiratory Society, as well as other cough researchers have recommended to consider both subjective and objective measures in the field of cough research(4,11). For the objective evaluation of cough measurement equipment is necessary Currently available cough monitors include e.g. the Leicester Cough Monitor (LCM)(12), VitaloJAK(13), the Hull automated cough counter (HACC) (14) and the LEOSound Lung-Sound-Monitor (15). LEOSound was used in a variety of works addressing multiple acute and chronic diseases in different patient groups (1,2,16,17). Even though, LEOSound has been validated before in smaller settings in children (18), we would evaluate the embedded cough detection algorithm utilizing a larger cohort with chronic disease. In this regard, the LEOSound cough detection algorithm is validated in a cohort of stable COPD patients. Germany) is a certified medical device with the purpose of lung sound monitoring during nighttime. The device ships with a software called Lung-Sound-Analyzer, which includes the display and replay functions, as well as the detection algorithms for cough and wheezing. In its default setting, LEOSound is applied using three bioacoustics sensors attached to the neck and back of the patient (see Fig 1).. Additionally the system includes an ambient sound microphone fixed into the body of the LEOSound to distinguish between patient generated and external noises. (15) LEOSound is established in therapy monitoring in clinical trials and is used by established paediatricians, pneumologists and in sleep medicine (16,17).

Dataset
We have used a COPD dataset, familiar to the authors (1,2) for the purpose of this validation study. All recordings were acquired in an outpatient environment. The dataset consists of data from two consecutive nights from 48 different stable COPD patients, resulting in 96 recordings. The recordings started about 1-2 hours before patients usually go to bed and lasted approx. 10 hours in average. All three bioacoustics sensors of the LEOSound were applied at default positions on the neck and back of the patients (see Fig.   1).. An overview of the demographics of the used dataset can be found in table 1.

Validation and statistics
LEOSound recordings are recorded in 30-second time segments, which we used as the basis of our validation. Medical experts scanned through every segment and assigned those to one of the categories of a contingency table. Rating was performed by using a self-implemented rating tool (implemented in MATLAB R2015b). Afterwards we pooled the ratings together to calculate the classification parameters (i.) accuracy, (ii.) sensitivity, (iii.) specificity, and (iv.) positive and negative predictive values as objective measures of the tested algorithm. Due to a largely unbalanced dataset, we randomly selected a similar amount of cough containing and non-cough containing segments to increase the accuracy of our objective measures. We further analysed false positive detections to identify typical sound patterns, which could lead to misdetections.

Discussion And Conclusion
Sensitivity and specificity of the LEOSound Lung Sound Monitor are showing an accurate and sufficient detection performance of the introduced algorithm. PPV and NPV depict the tendency of overshooting the actual amount cough. As shown by the analysis of false 6 positive results, throat cleaning is the major reason for misdetections. However, throat cleaning, also referred to as harrumph or slight cough, is fulfilling a similar purpose as cough, while also sharing the same signal pathway with cough, and can be a sign for secretions in the upper airways (19). Therefore, we think it is up to debate whether throat cleaning misdetections should be treated the same way as the other categories like snoring. Depending on the diagnosis or hypothesis in question, physicians can decide whether throat cleaning is a useful parameter in their respective case. Other reasons for false positive results in decreasing order were snoring, movement artifacts and moaning.
Movement artifacts are usually recorded as sudden rises of the amplitude, probably caused by scratching on the backside of the microphones. Those spikes in amplitude are similar to those caused by cough, but according to our analysis should be shorter. We think the algorithm in LEOSound can be improved by looking for length of a higher amplitude, rather than just for a sudden increase. Snoring can also incorporate sudden amplitude spikes, but has a different frequency pattern due to its harmonic properties (20). Therefore, spectral features can be used to distinguish cough and snoring. For moaning similar solutions might be applicable.
In comparison to other available cough detection devices and algorithms, it is evident that the LEOSound cough detection achieves a better sensitivity than all other validated devices, except for the VitaloJAK system. However, a comparison to the VitaloJAK system is difficult, since it is only a semi-automated cough detection device(21), whereas LEOSound is fully-automated. The main difference between VitaloJAK and LEOSound, or other fully-automated devices, is that the final counting of cough events for the VitaloJAK system is done by a rater, rather than an algorithm (7,22). Thus, data with reasonable sensitivity can be delivered with significantly less time and personnel expenditure in LEOSound. When compared to the LCM the LEOSound algorithm outperforms its sensitivity 7 by 13 % (85.7 % compared to 98.7 %), while also underperforming by 19 % (99.9 % compared to 80.3 %) in terms of specificity. Additional refinements of the detection results are necessary in both devices (7). The used dataset in the LCM validation is comparable to our dataset, but involves a more heterogeneous patient collective (12).
Comparisons to the HACC are also difficult, because the HACC is only an algorithm, which requires an one channel audio input (14). Since the LEOSound algorithm has a tendency to overshoot the actual amount of coughing, comparisons with the HACC lead to similar conclusions as the comparison to LCM. The HACC is better in distinguishing cough from other noises, but is less effective in detecting cough. Also the validation dataset of the HACC research group is much smaller in recording time and number of patients included (14). Overall, we conclude that the LEOSound is having another approach with trying to achieve a maximum sensitivity, while others try to maximize their detection specificity. Therefore, no cough events are missed, while the false positive results can be further improved. Considering the uncommon use of cough detection devices in daily healthcare practice, one could assume that such rare recording would be validated by medical experts anyways. Assuming this, a quick dismiss of false positive results would be more manageable than scanning through the whole recording for potentially missed cough events or episodes.