This paper introduces a new machine learning based comparator model to assess changes in the ease of breathing of COPD patients during loaded breathing. The comparator model is based on a random forest classifier that detects whether breathing becomes either more dificult, easier or remains stable. The designed model can accurately detect respiratory changes by comparing temporal segments of physiological signals measured during loaded breathing, with an 1 score of almost 80%, resp. 70% for the wearable solution. As the model is trained and tested with features derived from diferent signal modalities, such as respiratory flow, audio, bio-impedance and accelerometer data, we also did a systematic comparison of the signal modalities to assess their predictive power.
Chronic obstructive pulmonary disease (COPD) is a chronic inflammatory lung disease where the airflow adverse symptoms, such as breathing dificulty and shortness of breath, excess phlegm or sputum, and frequent coughing or wheezing. In the United States, it is estimated that 16 million Americans have this disease, whereas millions more people sufer from COPD but have not yet been diagnosed [1]. As a result, morbidity and mortality in COPD patients are considerably high. In order to diagnose COPD, a spirometer test is applied as a gold standard test for pulmonary function. The patient inhales and then exhales with a maximal efort through a mouthpiece, while the airflow going into and out of the lungs is measured and analyzed [2].
the predictive power of diferent physiological signals is
Time Time
still largely unaddressed.
This paper aims to address and explore this by proposing a new machine learning based comparator model that detects whether breathing becomes either more dificult, easier or remains stable. The patient’s ease of breathing is assessed by comparing pairs of segments of non-invasive physiological signals that were recorded during a loaded breathing test. By analysing the performance of the model, valuable new insights are obtained on the choice of signals that are most relevant to monitor changes in respiration. As such, the outcome of this study provides useful insights to decide which signal modalities should be measured by the wearable device, and forms a basis for the development of next-generation wearable respiratory monitoring technology. The availability of such a technology can possibly result in a faster intervention to prevent disease worsening, potentially leading to reduced health care costs, hospital (re)admissions and improved quality of life.
In summary, the main contributions of this paper are threefold: • A systematic analysis and comparative study of multiple physiological signals is performed, to identify which ones are most influential to correctly detect changes in the ease of breathing of a COPD patient during inspiratory loading. • Relevant features are automatically extracted from the signals, and a state-of-the-art classification algorithm for time series data is used to build a comparator model that accurately detects whether breathing becomes either more dificult, easier or remains stable. • The performance of the model is validated on data from a clinical study and the results are evaluated. Comparator model (load A – load B) ≥ t “easier breathing ” | load A – load B | < t “stable breathing ” (load B – load A) ≥ t “more difficult breathing ”
2.1. Data collection and setup
and enrolled for an inspiratory loaded breathing test in a clinical setting at Ziekenhuis Oost-Limburg (Belgium).
The data from six patients were excluded in this study due to one or more of the collected signal modalities being either unavailable or saturated. The resulting data set comprises 44 patients.
The study was approved by the local institutional medical ethics committee from Ziekenhuis Oost-Limburg with reference 18/0047U. The study followed the World Medical Association’s Declaration of Helsinki on Ethical Principles for Medical Research Involving Humans Subjects.
All patients provided written informed consent.
An incremental inspiratory threshold loading protocol was performed, where the patients were imposed to increasing inspiratory loads proportional to their maximal inspiratory pressure (MIP). The loads are quantified as 0%, 12%, 24%, 36%, 48% and 60% of the MIP that was measured before the start of the test [8]. As such, a higher load corresponds to an overall higher inspiratory efort that is required to breathe. Each load test was applied for a period of thirty breaths (i.e. a variable time length), with a two-minute resting period between the tests for each diferent load. The loads were sequentially applied in increasing order. The measurement of the MIP and the imposition of the loads was both done using an inspiratory muscle trainer (POWERbreathe KH2, POWERbreathe International Ltd, Southam, UK).
During the test, two systems were used to record several physiological signals simultaneously. The first system was a standard wired acquisition system (MP150, Biopac Systems, Inc., Goleta, CA, USA), used to record the accelerometer data and audio signals; the airflow was measured using Biopac, together with a pneumotach transducer (TSD107B, Biopac Systems, Inc.) connected to a diferential amplifier (DA100C, Biopac Systems, Inc.).
Given the fact that the lung sound signals have a bandwidth around 4000 Hz [9], the Biopac was configured to the bio-impedance signals are upsampled using cubic record at a sampling frequency of 10,000 Hz, and this was interpolation from 16 Hz to 100 Hz. applied to all the channels. The second system was a low- After filtering, a bidirectional moving average and power wearable device (imec the Netherlands, Eindhoven, moving variance filter is applied to remove artefacts in the Netherlands) with an injecting current of 100 uAp-p the data where the patient did not breathe into the pneuat 80 kHz that was used to record the bio-impedance sig- motach transducer correctly. When either of these values nal with a sampling frequency of 16 Hz [10]. The location in either direction is close to zero, this is considered an and configuration of the signals were as follows: artefact and removed.
After removal of artefacts, the signal is divided into • Respiratory flow (gold-standard, used as a base- multiple shorter signals. In order to define a commonline) length input signal for the machine learning model, each • Accelerometer: parasternal and diaphragm part of the signal that is not corrupted, is subdivided into (lower intercostal spaces), three axes several smaller segments having a predefined window • Microphone (audio): left lung, right lung, tracheal length . For each combination of patient and inspiratory • Bio-impedance (bioZ): 4 configurations load, one window is sampled from the longest stable breathing period without interruptions or artefacts. This
For the bio-impedance signals, the first of the four ensures that each segment has a consistent length. channels described in [10] is used in this study, because It was found that a window length of = 30 sec it was shown to have a more robust performance for es- is the most adequate choice, because it ensures that at timating respiratory volume changes when higher loads least 95% of the input space (i.e. all possible patient-load are imposed. The full details about this data collection combinations) is still included in the dataset after preprotocol for the respiratory flow, bio-impedance, and processing, while guaranteeing the inclusion of multiple accelerometer signals, as well as the set-up of the ex- breaths. This window length strikes the balance between periments are explained in [7, 11]. Regarding the audio including as much data as possible in the entire window signals, these were recorded using three microphones and including as many patient-load combinations as pos(TSD108, Biopac Systems, Inc) with a frequency response sible, avoiding biased results. of 35-3500 Hz. Two microphones (for both lungs) were positioned on the back, two to three centimeters below 2.3. Machine learning-based comparator the shoulder blades, at each side of the spinal cord. The model other microphone (for the tracheal sound) was positioned on the right side of the patient’s neck. After amplifying the sound 200 times, it was filtered with an analog lowpass filter of 5 kHz and a high-pass filter of 0.05 Hz. In the forthcoming sections of the paper, the respiratory flow is referred to as an obtrusive signal modality, whereas all others are considered as unobtrusive signal modalities.
Fig. 1 shows an overview of the designed comparator model. Pairs of segments from the same signal modalities of the same patient are used as an input for the model, and a three-way classification is calculated, depending on whether the load of the second signal is lower than, higher than, or comparable to the load of the first signal.
It is hypothesized that this method can then also be used 2.2. Preprocessing and segmentation to detect increased dificulty in breathing, which may be indicative for worsening of the respiratory condition.
The signals contain noise due to subject movement, elec- The threshold used to separate the categories is set to trical inference, measurement noise and other distur- 12%, to match the granularity of variations in the loads. bances. In order to extract all relevant information, all For each signal type, a comparator model is trained signals except audio are first filtered with a low-pass using pairs of signals recorded at diferent load combizero-phase Butterworth filter with the following orders nations, with the labels corresponding to an increase, and cut of frequencies per signal modality: a fith-order decrease, or stability in inspiratory load. Since the model 40 Hz low-pass filter for the respiratory flow signal, a is trained on data from all the patients in the training fourth-order 0.7 Hz low-pass filter for the bio-impedance set at once, a general model is obtained, rather than a signal, and a eighth-order 40 Hz low-pass filter for the patient-specific one. accelerometer data. The filter values are based on the As six diferent loads were considered in the breathing characteristics of the noise present in the signals, remov- test, 21 pairs of load combinations can be generated. The ing higher frequency noise while retaining the relevant swapped pairs are also included to train the model, resultinformation below the cut of frequencies. ing in 42 load pairs. As such, it would have been possible
All signals are resampled to a common sampling rate. to have 1848 pairs. However, one patient was able to The respiratory flow, accelerometer, and audio signals perform only 5 loads instead of 6. For this reason, 1836 are downsampled from 10,000 Hz to 100 Hz, whereas pairs were available for the development of the model. 2.3.1. Feature extraction 2.3.2. Model training and cross-validation The performance of the models is evaluated using 10fold cross-validation, ensuring no overlap of samples from patients between diferent folds. For each training and test step, features are extracted from the 9 training folds with the exact same parameters each time. The model performance is assessed for every pair of loads from patients in the test fold. Note that the evaluation is based on a weighted 1 score to take into account the small class imbalance, as instances from the stable condition are less common than from the worsening and improving conditions.
modalities is shown in Fig. 2.
A one-way ANOVA analysis is performed on all results of the unobtrusive signal modalities, with = 0.05 [15].
For this, all 1 scores are considered from each signal modality as a separate group, with the null hypothesis that the means for each group are sampled from the same distribution. The resulting p-value of 0.029 < and rejection of this null hypothesis confirms that there is a
1
the classification performance of the model. For examEasier 0.82 0.90 0.82 0.90 0.87 0.82 ple, the model has a 10% chance of missing an increase Stable 0.69 0.88 0.71 0.88 0.83 0.70 in dificulty and a 19% chance of being wrong when an More dif- 0.82 0.90 0.81 0.90 0.87 0.82 increase is detected. As such, it is important to interpret ficult the classification results carefully and weigh the potential consequences of false positives and false negatives, Table 1 as the acceptable margin of error may vary depending on Quality metrics of best performing model (sensitivity, speci- its application. For example, this error margin may be acficity, positive predictive value, negative predictive value, ac- ceptable for regular monitoring of stable COPD patients, curacy and 1 score) while it may not be acceptable for high-risk patients.
When considering other signals that can easily be acquired with wearable devices, a lower 1-score is obstatistically significant diference in the mean 1 scores served when compared to the use of the respiratory flow of the signal modalities. Amongst those signal modal- signal. Nevertheless, the actual performance of these ities, bioZ, accelerometer (parasternal) and audio (tra- models is less important from a clinical perspective, becheal) seem to provide the best predictive power to detect cause the comparator model in this study is only used changes in the ease of breathing. A paired t-test with to benchmark and rank the diferent signal modalities = 0.05 confirms that the 1 score for the tracheal audio according to their predictive power. Having these insignal is significantly higher than the left lung ( = 0.002 ) sights can be valuable, because an identification of the and right lung ( = 0.028 ) signals. Furthermore, the best-performing signal modalities can help to make an 1 score for parasternal accelerometer signal is signifi- informed choice of sensors during the design of a wearcantly higher than the diaphragm accelerometer signal able. ( = 0.025 ). There is no significant diference between Such a wearable can collect longitudinal data from pathe 1 score of the bioZ and accelerometer (parasternal) tients during normal daily activities, on which the comsignals ( = 0.821 ), the bioZ and audio (tracheal) sig- parator model can be retrained. Having more lengthy nals ( = 0.401 ), and the accelerometer (parasternal) and signals makes it possible to consider multiple window segaudio (tracheal) signals ( = 0.64 ). ments, which can further boost the model performance.
Fig. 3 shows the confusion matrix of the overall best Furthermore, the availability of more extensive data sets performing model, i.e. the random forest model based on creates new possibilities to apply advanced deep learntsfresh features, that considers the (obtrusive) respiratory ing techniques that have shown to be efective on similar lfow signal. It is seen that the majority of the instances problem settings with respiratory data [16], while also enare classified correctly, and misclassifications are more hancing the generalizability of the current feature-based common between neighbouring classes. Table 1 pro- model. vides an overview of the performance metrics calculated Future work will focus on the identification of an opto assess the quality of the model, including sensitivity, timal combination of unobtrusive signal modalities to specificity, positive predictive value (PPV), negative pre- avoid redundancy in the selection of signal modalities dictive value (NPV), accuracy and the 1 metric for each within a certain category. Longitudinal, clinical and exof the three classes. From both Fig. 3 and Table 1, it is ternal validation of the approach will also be performed. seen that changes in breathing dificulty can be identified Additionally, to increase trust in the model, interpretable more accurately than stability. This is not unexpected, as machine learning methods will be explored. Generating the stable class is more similar to the other two classes, explanations alongside classifications also allows for phythan they are to each other. isicans to incorporate the reasoning of the model within their own decisions.
pattern during the inspiratory load protocol applied to This paper presents a novel machine-learning based comCOPD patients can be assessed by our machine learning parator model that detects changes in the ease of breathbased comparator model. The best performance is ob- ing of COPD patients during inspiratory loaded breathing. tained when considering the respiratory flow signal, with Numerical results provide a comparison of diferent input an 1 score of 0.78. This demonstrates the ability of the signals and models. When applied to the respiratory flow, model to discriminate between an increase, remaining a weighted 1 score of 0.78 is obtained. When considering stable or decrease in the ease of breathing for patients. other signal modalities that are not as obtrusive, and can
However, there is still an error margin associated with be measured with wearable devices, the ones that ofer
The authors have no conflicts of interest to declare. This work was partially funded by the Flemish Government (AI Research Program). the best predictive performance are bioZ, accelerometer (parasternal) and audio (tracheal), with a weighted 1 score of 0.69, 0.69 and 0.68 respectively.
cross power spectral density autoregressive process coefLempel-Ziv complexity estiabsolute energy quantiles (10ℎ ) kurtosis Combinations of simple features 2 > standard error ⋅ 1 average over diferences longest subsequence above and below mean sum of reoccuring values (10ℎ ) energy ratio by chunks mean of largest values first and last location of maximum Complex features Benford correlation c3 statistic autocorrelation statistics binned entropy fourier entropy FFT statistics linear trend statistics symmetry index of of mass quantiles percentage of unique values number of zero-crossings median variance ( 2) (absolute) maximum root mean square skewness value and range count mean values below percentage of reoccuring ⋅ > minimum first and last location of time reversal asymmetry statistic tance number of (unique) peaks permutation entropy CWT coeficients ficient test Langevin coeficients augmented
Dickey-Fuller