Cough Sound Analysis Can Rapidly Diagnose Childhood Pneumonia

Abstract

Pneumonia annually kills over 1,800,000 children throughout the world. The vast majority of these deaths occur in resource poor regions such as the sub-Saharan Africa and remote Asia. Prompt diagnosis and proper treatment are essential to prevent these unnecessary deaths. The reliable diagnosis of childhood pneumonia in remote regions is fraught with difficulties arising from the lack of field-deployable imaging and laboratory facilities as well as the scarcity of trained community healthcare workers. In this paper, we present a pioneering class of technology addressing both of these problems. Our approach is centred on the automated analysis of cough and respiratory sounds, collected via microphones that do not require physical contact with subjects. Cough is a cardinal symptom of pneumonia but the current clinical routines used in remote settings do not make use of coughs beyond noting its existence as a screening-in criterion. We hypothesized that cough carries vital information to diagnose pneumonia, and developed mathematical features and a pattern classifier system suited for the task. We collected cough sounds from 91 patients suspected of acute respiratory illness such as pneumonia, bronchiolitis and asthma. Non-contact microphones kept by the patient’s bedside were used for data acquisition. We extracted features such as non-Gaussianity and Mel Cepstra from cough sounds and used them to train a Logistic Regression classifier. We used the clinical diagnosis provided by the paediatric respiratory clinician as the gold standard to train and validate our classifier. The methods proposed in this paper could separate pneumonia from other diseases at a sensitivity and specificity of 94 and 75% respectively, based on parameters extracted from cough sounds alone. The inclusion of other simple measurements such as the presence of fever further increased the performance. These results show that cough sounds indeed carry critical information on the lower respiratory tract, and can be used to diagnose pneumonia. The performance of our method is far superior to those of existing WHO clinical algorithms for resource-poor regions. To the best of our knowledge, this is the first attempt in the world to diagnose pneumonia in humans using cough sound analysis. Our method has the potential to revolutionize the management of childhood pneumonia in remote regions of the world.

Appendix

Our method requires the computation of a number of mathematical features from cough sounds. This Appendix describes the features we computed from each sub-segment x _i, i = 1, 2, 3 of a recorded cough sound x.

The Bispectrum Score (BGS)

The 3rd order spectrum of a signal is known as the bispectrum.2 Unlike the power spectrum (the 2nd order spectrum based on the autocorrelation), the bispectrum preserves Fourier phase information. The bispectrum $B_{xi} \left( {\omega_{1} ,\omega_{2} } \right)$ of the segment x _i can be estimated from (5) as,

$$ B_{xi} \left( {\omega_{1} ,\omega_{2} } \right) = \mathop \sum \limits_{{\tau_{1 = - \infty } }}^{{\tau_{1 = + \infty } }} \mathop \sum \limits_{{\tau_{2 = - \infty } }}^{{\tau_{2 = + \infty } }} W\left( {\tau_{1} ,\tau_{2} } \right) \cdot C_{xi} \left( {\tau_{1} ,\tau_{2} } \right)e^{{ - j\left( {\tau_{1} \omega_{1} + \tau_{2} \omega_{2} } \right)}} $$

(5)

where $W\left( {\tau_{1} ,\tau_{2} } \right)$ is a bispectrum window function such as the minimum bispectrum-bias supremum window14 used in this paper, $C_{xi} \left( {\tau_{1} ,\tau_{2} } \right)$ is the third order cumulants of x _i estimated with (6), and, $\omega_{1} ,\omega_{2}$ denotes digital frequencies.

$$ C_{xi} \left( {\tau_{1} ,\tau_{2} } \right) = \frac{1}{L}\mathop \sum \limits_{k = 0}^{L - 1} x_{i} \left( t \right)x_{i} \left( {t + \tau_{1} } \right)x_{i} \left( {t + \tau_{2} } \right), \left| \quad {\tau_{1} } \right| \le Q,\left| {\tau_{2} } \right| \le Q $$

(6)

In (6), Q is the length of the 3rd order correlation lags considered and x _i is a zero-mean signal.

The bispectrum is a 2D signal. However, it can be proven2 that for linear signals, any 1D oblique slice of the bispectrum other than the slices parallel to the axes: ω₁ = 0, ω₂ = 0 and ω₁ + ω₂ = 0 carries sufficient information to characterise the entire 2D bispectrum within a phase factor. In this work, we capture the information available in the bispectrum via the diagonal slice P(ω) defined by ω₁ = ω₂ = ω, i.e. $P\left( \omega \right) = B_{xi} \left( {\omega , \omega } \right).$Then the Bispectrum Score (BSG) is computed as defined in (7). In (7) we used k ₁ = 90 Hz, k₂ = 5 kHz, k₃ = 6 kHz and k₄ = 10.5 kHz.

$$ BSG = \frac{{\mathop \int \nolimits_{k1}^{k2} |P\left( \omega \right)| \cdot d\omega }}{{\mathop \int \nolimits_{k3}^{k4} |P\left( \omega \right)| \cdot d\omega }} $$

(7)

Non-Gaussianity Score (NGS)

NGS score is a numerical measure of non-Gaussianity of a given segment of data x _i . The normal probability plot can be utilized to obtain a visual measure of the Gaussianity of a set of data, and the NGS score is a way of quantifying the non-Gaussianity based on regression analysis. We used (8) to estimate the NGS score, where p and q represents the normal probability plots of the reference normal data and the analysed data (x _i ). The symbol N is the number of data points used in the probability plot.

$$ NGS = 1 - \left( {\frac{{\mathop \sum \nolimits_{j = 1}^{N} \left( {q\left[ j \right] - p} \right)^{2} }}{{\mathop \sum \nolimits_{j = 1}^{N} \left( {q\left[ j \right] - \bar{q} } \right)^{2} }}} \right) $$

(8)

Formant Frequencies

In speech analysis, formants frequencies (FF) are referred to as the resonances of the vocal tract.16 In cough analysis, it is reasonable to expect that the resonances of the overall airway that contribute to the generation of a cough sound will be represented in the formant structure. One classic example for this is wheeze. Existence of mucous can also change acoustic properties of airways. We included the first four formants (F1, F2, F3, F4) in our candidate feature set. We computed F1–F4 by peak picking the Linear Predictive Coding (LPC) spectrum of cough segments x _i. For this work we used a 14th order LPC model with the parameters determined via the Levinson-Durbin recursive procedure.17

Log Energy (LogE)

The log energy for every sub-segment x _i was computed using (9):

$$ \text{LogE} = 10\log_{10} \left( {\varepsilon + \frac{1}{N}\mathop \sum \limits_{k = 1}^{K} \left( {x_{i} \left( t \right)^{2} } \right)} \right) $$

(9)

In (9) ε is an arbitrarily small positive constant added to prevent any inadvertent computation of the logarithm of 0.

Zero Crossing (Zcr)

The number of zero crossings was counted for each sub-segment x _i.

Kurtosis (Kurt)

The kurtosis is a measure of how peaky the probability density distribution of x _i is. It is the fourth central moment of x _i and can be computed using (10), where μ and σ respectively denote the mean and the standard deviation of x _i.

$$ {\text{Kurt}} = \frac{{E\{ \left( {x_{i} \left[ k \right] - \mu } \right)^{4} \} }}{{\sigma^{4} }} $$

(10)

Mel-Frequency Cepstral Coefficients (MFCC)

MFCCs have been widely used in speech recognition systems.7,9 MFCC provides some resilience to the non-linguistic sources of variance in speech signals. They also provide orthogonal features making facilitating the training of the classifier. The computation of MFCC involves the estimation of short-term power spectra, mapping to Mel frequency scale and then computing the cepstral coefficients. In our work, we included 12 MFCC coefficients in our feature set.