Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005
Classification of Impulse Oscillometric Patterns of Lung Function in Asthmatic Children using Artificial Neural Networks Miroslava Barúa, Homer Nazeran, Patricia Nava, Bill Diong, and Michael Goldman* Department of Electrical and Computer Engineering, University of Texas at El Paso El Paso, TX 79968, USA * Geffen School of Medicine, University of California at Los Angeles, Los Angeles, CA 90034, USA Email addresses:
[email protected],
[email protected] Abstract -- Impulse Oscillometry (IOS) is an innovative patient-friendly pulmonary testing technique which measures the respiratory system impedance (Z) by using the spectral components of pressure to flow ratio which yields resistance and reactance values at different frequencies. The high dimensionality of IOS measurement data makes the analysis of this information difficult. Artificial Neural Networks (ANNs) are mathematical models composed of a large number of highly interconnected neurons that are able to learn and generalize from data. An ANN-based approach to the analysis of IOS data can potentially provide an efficient and automatic method to recognize and classify pulmonary diseases. This would help characterize major respiratory illnesses such as asthma based on IOS measurements. Asthma can be difficult to diagnose, because the symptoms are sometimes similar to other lung conditions. A data set composed of 361 impulse oscillometric patterns from asthmatic children was used in this study. The ANN was capable of distinguishing between relatively constricted and nonconstricted airway conditions in these patients. Using all of the 361 patterns during training as well as in the feed-forward stage, a classification accuracy of 95.01% was obtained for validation. When the ANN was presented with only 60% of the original 361 patterns in the data set during training and with the remaining 40% as unseen patterns, the generalization stage, a classification accuracy of 98.61% was achieved. These results show that ANNs can successfully be trained with the IOS data, enabling them to generalize the IOS parameter relationships to classify previously unseen pulmonary patterns, such as in asthma. The next step is to obtain expert rules by extracting them from the knowledge acquired by the neural network and develop a fully automated classification system to aid physicians in classifying and characterizing pulmonary diseases based on the patient-friendly IOS measurements. Keywords - Impulse oscillometric measurements; lung function; artificial neural networks; pulmonary disease classification; asthma; children.
I. INTRODUCTION Impulse Oscillometry (IOS) is an innovative pulmonary testing technique that measures the respiratory system impedance (Z). By obtaining the pressure to flow ratio, it yields a complex number (Z= R + jX) composed of respiratory system resistance (R) and reactance (X) [1]. IOS offers several advantageous over conventional techniques lung testing methods because it is non-invasive
0-7803-8740-6/05/$20.00 ©2005 IEEE.
and unobtrusive, and requires only ‘passive cooperation’ from the patient. IOS provides closely reproducible results when care is taken to ensure technically satisfactory tracings, and requires only a brief examination time, during which three or four 20-second replicate tests are performed. As well as offering greater sensitivity to efforts of treatment which may prove to be more cost effective for patient testing, IOS provides both the patient and operator a userfriendly testing environment. An Artificial Neural Network (ANN) is a parallel distributed processing system capable of solving complex classification problems. By their remarkable ability to derive meaning from complicated or imprecise data, ANNs can become excellent classifiers and be able to learn and generalize from data by using different learning processes or learning rules to acquire their knowledge. Currently there are no robust algorithms to classify the respiratory system airway impedance from impulse oscillometric patterns into pulmonary disease classes. An ANNbased approach to the analysis of pulmonary data can potentially provide an efficient and automatic method to recognize and classify pulmonary diseases, as well as characterize the features of respiratory illnesses such as asthma in a manner that contributes to the physician’s understanding of disease processes and their changes with treatment. In this paper, we first present some background information on the IOS technique followed by the concepts necessary to understand ANNs operation and their learning process as well as an overview of asthma. We then define the problem of classifying patterns in asthmatic children as measured by IOS and present an ANN that can solve this classification problem with good accuracy. We finally present our results and conclude the paper with the future directions for this research. II. MATERIALS AND METHODS A. Impulse Oscillometry The IOS testing technique consists of applying a rectangular electrical impulse through an external loudspeaker (as illustrated in Figure 1) and driving it to move abruptly for-
ward within 30-40 msec, followed 200 msec later by a similarly abrupt backward movement. Speaker motion causes rapid pulses of pressure change in the mouth, which are analyzed by Fast Fourier Transform (FFT) into spectral components between 5-35 Hz. The resulting pressure waveform is applied to the patient’s mouth and superimposed on the normal breathing waveform. Flow and pressure waveforms are analyzed by FFT separately for the amplitudes of spectral components in phase with each other (resistance) and out of phase with each other (reactance). Therefore, the calculated respiratory impedance consists of graphs of respiratory resistance and reactance as functions of oscillation frequency (Figure 2). These plots provide information regarding the nature of the frequency dependent impedance characteristics that describe airway function at different anatomical levels within the respiratory tree.
more sensitive reactance values indicate the back pressure, or “echo” that relates to elastic and inertial properties of the respiratory system. As IOS mouth pressure pulses move towards the lung periphery, damping of the pressure pulse is greater for high frequencies than for low frequencies, as shown in studies of high frequency ventilation [2]. The pulmonary impedance calculations are continuous in the frequency domain, as the FFT in IOS uses a Fourier integral, in contrast to the Fourier series used in other forms of forced oscillation (pseudorandom noise). Numerical values at intervals from the graphs described above are provided at six different frequencies, from 5Hz to 35 Hz, at 5 Hz intervals, generating R5, R10, R15, R20, R25, R35, X5, X10, X15, X20, X25, and X35 for IOS resistance and reactance values, respectively. B. Artificial Neural Networks Just like the brain, an artificial neural network (ANN) is a parallel distributed information processing structure composed of many neurons, which are heavily interconnected. The model of the artificial neuron and its connections captures just the most essential features of its biological counterpart.
Figure 1. IOS Testing Technique. Figure 2. Respiratory impedance and its components.
ANNs are model-free estimators that consist of many simple processing elements called neurons, units, cells, or nodes [3]. Each artificial neuron is connected to other neurons via unidirectional signal channels called links, each with an associated weight (W’s) representing the information being used to solve the problem. Each neuron has an internal state or activation function, where the inputs received are processed. The neuron will only “fire” or activate if the threshold level has been met, and propagation of this response occurs. ANNs can be configured for a specific application and are able to “learn” and generalize from training data by using different learning processes or learning rules to achieve their “knowledge.” The most well known and highly implemented type of training process is the backpropagation algorithm. Backpropagation is an efficient and exact method for calculating all the derivatives of a single target quantity (e.g., pattern classification error) with respect to a large set of input quantities (i.e., parameters or weights in a classification rule) [4]. C. Asthma
After approximately 30 sec of normal tidal breathing, the IOS provides measurements from the pressure to flow ratio, which yields a complex airway impedance number (Z = R + jX) that is most profitably analyzed with respect to each of its two parts, respiratory system resistance (R) and respiratory system reactance (X). Resistance values reflect the forward pressure needed to push flow, while the
Asthma is a chronic disease that causes the lining of the airways to become inflamed and swollen, often associated with extra mucus production. Episodes of asthmatic constriction of the airways are manifested when sufficient narrowing of the airway occurs. This is manifested both at the level of large-diameter airways due to smooth muscle contraction, and at the level of smaller, more peripheral airways, due to accumulation of fluid and mucus within the small
airway walls, making it difficult to breathe [5]. Symptoms include wheezing, tight chest, shortness of breath, cough and obstruction of the airways. These symptoms vary in their severity and duration in adults and in children. Asthma can be difficult to diagnose, because the symptoms are sometimes similar to other conditions, including allergic rhinitis (‘hay fever’), lung infection, and even cardiac problems. Asthma affects 3% to 5% of adults and 7% to 10% of children. Half of the people with asthma develop it before age 10 and most develop it before age 30. The rate of asthma has increased by about 60% since 1979 among all ages, races, and gender groups – however, it is more common in blacks and Hispanics than in whites [6]. Although a variety of causes, both allergic and non-allergic, as well as environmental factors contribute to asthma, early detection and proper treatment can ease the patient’s struggle to live with this disorder. III. CLASSIFICATION OF PULMONARY CONDITIONS BASED ON IOS MEASUREMENTS
A. The Problem Identifying changes in airway impedance is essential to early detection of pathological impairments of the small airways [7], and is not recognized by conventional lung function tests. IOS provides an attractive test alternative to traditional methods. In fact, many recent reports have described IOS as being more sensitive than conventional spirometry in assessing airway abnormalities in asthmatics and in children at risk for developing persistent asthma [8, 9]. IOS has been reported in children about 3-4 years old, an age when the usefulness of conventional spirometry is questionable [1016]. A major drawback with using IOS is that although obtaining respiratory impedance values is easy, the resulting values are difficult to understand by clinicians as they are based on an electrical equivalent circuit model of the respiratory system. Categorization of pulmonary diseases by looking at the plotted curves of respiratory impedance can prove a difficult task for the untrained pulmonologist. Since the IOS measurements provide six values of resistance and six values of reactance, the high degree of dimensionality makes diagnostic analysis of this information complicated. Using conventional techniques to solve this problem is not straight forward. Finding a function that relates these variables to one another is not immediately obvious. Development of an algorithmic solution and finding a pattern in the resistance and reactance measurements could prove challenging. B. The Solution Since this application involves the analysis and classification of experimental data composed of many different variables, the deployment of artificial neural net-
works becomes an attractive proposition to finding solutions to such complex problems. In this study a data set composed of IOS respiratory impedance measurements from children diagnosed with asthma and evaluated either in a condition of clinical asthmatic airway constriction (constricted), or relative freedom from asthmatic airway constriction (non-constricted) was available to us. The IOS data set comprised of 361 patterns acquired from 41 asthmatic children along with their demographic attributes and corresponding diagnostic information determined by forced expiration test results. The 361 patterns in the data set came from multiple tests obtained from both male and female children (age: 2 - 8 years, height range: 0.88 - 1.4 m, and weight: 12 – 32 Kg). On average each patient provided 8.8 tests. Some patients seemed to follow a routine check up of multiple tests per day. In other words, a patient would get measured in his/her asthmatic state for an average of 3 good measurements (providing 3 asthmatic patterns). Then the patient took the medicine/treatment and after 30 and 60 minutes the patients were measured again providing an average of 3 good measurements (yielding 3 non-asthmatic patterns.) Some of the patients would return for up to 4 different dates in a 1-month period and repeat the measurement routine. Some others just got tested in one date or two. Therefore one patient might have provided up to 24 tests while others just yielded 6 tests. The 361 data set contained two different classes: 168 airway constriction patterns and 193 patterns considered relatively free from airway constriction. Tables 1 and 2 summarize the children data set characteristics. TABLE I SUBJECT DEMOGRAPHICS
Age Height (m) Weight (Kg)
MIN 2 0.88 12
MAX 8 1.4 32.7
Mode 5 0.97 19.1
Avg. 3.96 1.07 19.48
TABLE 2 ASTHMATIC DATA SET CHARACTERISTICS
Female Male Total
Airway Constricted 114 54 168
Airway Non-constricted 127 66 193
Total 241 120 361
The different patterns in the data set were mixed at random so that the patterns from all categories were distributed throughout the training data set. The categorization of data by an ANN requires the extraction of all possible key features to be considered from the data, and these are then used to undertake the classification process.
Each record in the data set comprises of an input vector with 16 attributes or variables properly normalized: age; gender (female = 1, male = 2); height (meters); weight (kilograms); and the IOS measurements: R5, R10, R15, R20, R25, R35, X5, X10, X15, X20, X25, and X35. An output of (1, 0) represents constricted airway, and (0, 1) denotes non-constricted airway. IV. RESULTS The ANN used was simulated using the C language to create a fully connected feed-forward network with a binary sigmoid activation function. The allocation of neurons required 16 nodes in the input layer, 16 in the hidden layer and 2 for the output layer as illustrated in Fig. 3.
achieved notably high classification accuracy when categorizing new, previously unseen asthmatic and non-asthmatic patterns. V. CONCLUSIONS IOS measures respiratory impedance in a user-friendly manner, but simple visual inspection of just the resistance and reactance values makes it hard for the untrained pulmonologist to come up with a classification of pulmonary conditions. Since neural networks are computing models with the amazing ability to learn from examples, it makes them a perfect candidate to provide an accurate and automatic classification of the IOS data. The simulated ANN produced notably high classification accuracy when categorizing new, previously unseen relatively constricted and non-constricted patterns in asthmatic children. This shows that ANNs can successfully be trained with the IOS data, making them capable of generalizing the IOS parameter relationships to classify previously unseen pulmonary patterns.
Figure 3. Architecture of the ANN.
The next step is to extract expert rules from the knowledge acquired by the network. By incorporating the strengths of fuzzy theory to obtain fuzzy logic decision rules, a more powerful hybrid neuro-fuzzy classifier can be created. Since the network’s ‘knowledge’ is represented at a sub-symbolic level in terms of connections and weights which provide little insight into how decisions are made [17], unraveling the meaning contained in these values is highly desirable in order to obtain humanly understandable representation of the network’s knowledge.
The network was trained using the backpropagation algorithm for up to 3000 iterations at a time. To avoid local minima, the backpropagation learning rule was modified by using momentum and learning rate variables that were dynamically changed throughout the training process from 0.1 to 0.8 while allowing a 0.01 maximum total error.
The obtained results are encouraging and further research is needed in order to provide a fully automated system to aid clinicians in classifying and characterizing pulmonary diseases based on the patient-friendly IOS measurements of lung function.
The results for the simulated network are plotted in Fig. 4 as a function of the (DP set. The ANN was capable of distinguishing between relatively constricted and nonconstricted asthmatic states. Using all the 361 patterns during training as well as in the feed-forward stage, a classification accuracy of 95.01% was obtained in the validation stage. When the ANN was presented with only 60% of the original 361 patterns in the data set during training and with the remaining 40% as unseen patterns in what is called generalization, the classification accuracy was 94.44%.
ACKNOWLEDGEMENT
Further investigation of the generalization result of 94.44% accuracy was performed. By using the optimum values of DP set to (0.8, 0.3), and training for 322 epochs, the ANN was capable of a generalization result with 98.61% classification accuracy. These results show that the network
This research was funded in part by National Science Foundation Award # EIA-0325024 to Dr Patricia Nava.
REFERENCES [1] Vogel, J., Schmidt, U., “Impulse Oscillometry. Analysis of lung mechanics in general practice and the clinic, epidemiological and experimental research,” pml Verlagsgruppe Gmbh, Frankfurt am Main, 1995; 6-16, 76-81. [2] Frantz, I.D., Close, R., “Alveolar pressure swings during high frequency ventilation in rabbits,” Pediatr. Res., 1985; 19:162-166. [3] Nazeran, H., Behbehani, K. “Neural Networks in Processing and Analysis of Biomedical Signals,” in Nonlinear Biomedical Signal Processing, Edited by M. Akay, IEEE Press, Piscataway, USA, 2000; 69-97. [4] Werbos P. J., “Backpropagation through time: What it does and how to do it”, Proc. IEEE, vol.78, no.10, 1990; 1550-1560.
[5] Badash,
Michelle, “Living with asthma,” http://www.medtropolis.com/healthcontent.asp? Page = /CID/ asthma/index, 2004. [6] Allergic and Environmental Asthma, eMedicine Journal. November 26, 2001. [7] Jaeger, E., “MasterScreen IOS - Follow the Impulse,” Erich JAEGER GmbH. Hoechberg, http://www.jaeger-toennies.com/english.html, 2000. [8] Marotta, A., Klinnert, M.D., Price, M.R., Larsen, G.L., Liu, A.H.: “Impulse oscillometry provides an effective measure of lung dysfunction in 4-year-old children at risk for persistent asthma,” J Allergy Clin Immunol, 2003; 112(2):317-22. [9] Goldman, M., Carter, R, Klein, R, Fritz, G, Carter, B, Pachucki, P., “Within- and between day variability of respiratory impedance, using impulse oscillometry in adolescent asthmatics,” Pediatric Pulmonology, 2002; 34:312–19. [10] Nielsen K, Bisgaard H. The effect of inhaled budesonide on symptoms, lung function, and cold air and methacholine responsiveness in 2- to 5-year-old asthmatic children. Am J Respir Crit Care Med. 2000; 162:1500-1506. [11] Nielsen K, Bisgaard H. Lung function response to cold air challenge in asthmatic and healthy children of 2-5 years of age. Am J Respir Crit Care Med. 2000; 161:1805-1809. [12] Nielsen K, Bisgaard H. Discriminative Capacity of Bronchodilator
Response Measured with Three Different Lung Function Techniques in Asthmatic and Healthy Children Aged 2 to 5 Years. Am J Respir Crit Care Med 2001; 164:554-559. [13] Goldman M, Carter R, Klein R, Fritz G, Carter B, Pachucki P. Withinand between-day variability of respiratory impedance using impulse oscillometry in Adolescent Asthmatics. Pediatric Pulmonology, 2002; 34:312-319. [14] Bisgaard H, Klug B. Lung function measurement in awake young children. Eur Respir J. 1995; 8:2067-2075. [15] Klug B, Bisgaard H. Measurement of Lung Function in Awake 2-4 year-old Asthmatic Children During Methacholine Challenge and Acute Asthma: A Comparison of the Impulse Oscillation Technique, the Interrupter Technique, a Transcutaneous Measurement of Oxygen Versus Whole-Body Plethysmography. Pediatr Pulmonol. 1996; 21:290-300. [16] Goldman M, Crim C, Lee B, VanderMeer A. Salmeterol Xinafoate Improves Pulmonary Function as Assessed by Forced Oscillation (FO) in 2-4 Year Old Children with Asthma. Am J Respir Crit Care Med 2003; 167:A512. [17] Huang, H. S., Xing, H., “Extract intelligible and concise fuzzy rules from neural networks,” Fuzzy Sets and Systems, Elsevier Science, 2002; 132, 233-243.
Classification Performance Results 100 90 80
Accuracy %
70 60 50 40 30 20 10 0 (0.1,0.1) (0.1,0.5) (0.2,0.1) (0.2,0.5) (03.,0.1) (0.3,0.5) (0.4,0.1) (0.4,0.5) (0.5,0.1) (0.5,0.5) (0.6,0.1) (0.6,0.5) (0.7,0.1) (0.7,0.5) (0.8,0.1) (0.8,0.5)
(Learning Rate, Momentum) Validation
Generalization
Figure 4. The Classification accuracy of the artificial neural network during Validation and Generalization stages.
