Chest Email Content Delivery
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Guest Access | Sign In via User Name/Password
This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF) Free
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Article Archive
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via ISI Web of Science (10)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Bibi, H.
Right arrow Articles by Nutman, J.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Bibi, H.
Right arrow Articles by Nutman, J.
(Chest. 2002;122:1627-1632.)
© 2002 American College of Chest Physicians

Prediction of Emergency Department Visits for Respiratory Symptoms Using an Artificial Neural Network*

Haim Bibi, MD; Amir Nutman, BSc; David Shoseyov, MD; Mendel Shalom, MSc; Ronit Peled, MPH; Shmuel Kivity, MD and Jacob Nutman, MD

* From the Barzilai Medical Center (Drs. Bibi and Shoseyov, and Mr. Peled), Ashkelon, Israel; Tel-Aviv Technical College (Mr. A. Nutman and Mr. Shalom), Tel-Aviv University, Tel-Aviv, Israel; Allergy Clinic (Dr. Kivity), Soraski Medical Center, Tel Aviv, Israel; and Department of Pediatrics (Dr. J. Nutman), The Children’s Medical Center of Overlook, Summit, NJ.

Correspondence to: Haim Bibi, MD, Barzilai Medical Center, Ashkelon, Israel 78306; e-mail: haim_76407{at}yahoo.com


   Abstract
TOP
 Abstract
Introduction
 Materials and Methods
 Results
 Discussion
 References
 
Study objectives: Accurate prediction of the effect of atmospheric changes, including pollutants, on emergency department (ED) visits for respiratory symptoms would be useful, but has proven difficult. The main difficulty is the limitation of the classical linear models and logistic regression with multiple variables to handle the multifactorial effect.

Design and setting: To predict ED visits, we have created a computer-based model called an artificial neural network (ANN) using a back-propagation training algorithm and genetic algorithm optimization. This ANN was fed meteorologic and air pollution input variables and trained to predict the number of patients admitted to the ED with respiratory symptoms of asthma, COPD, and acute and chronic bronchitis on the corresponding day. One thousand twenty data sets were extracted from an ED admittance database at the Barzilai Medical Center (Ashkelon, Israel), and randomized to a network training set (n = 816) and a test set (n = 204).

Results: The neural network performed best when the predictor variables used were temperature, relative humidity, barometric pressure, SO2, and oxidation products of nitric oxide, and the data presented as peak value 24 h prior to ED admission and the average during the 7 days before the ED visit. The neural network was able to predict the test set with an average error of 12%.

Conclusion: Based on meteorologic and pollution data, the use of an ANN can assist in the prediction of ED visits related to respiratory conditions.

Key Words: artificial neural networks • emergency department • respiratory symptoms


   Introduction
TOP
 Abstract
 Introduction
Materials and Methods
 Results
 Discussion
 References
 
Symptoms of asthma and COPD are common complaints of emergency department (ED) patients. These symptoms vary with "triggers" in the environment. The triggers can be air pollution,1 allergens,2 3 4 and atmospheric changes.5 6 Triggers actuate inflammatory processes in the airways, which ultimately result in bronchospasm and shortness of breath7 8 that may prompt an ED visit.

Because the exacerbations are caused by multiple stimuli, accurate prediction of an exacerbation is difficult. To our knowledge, there is no reliable way to predict these exacerbations. We attempted to develop a mathematical model for predicting respiratory symptom exacerbations using neural network technology.

Artificial neural networks (ANNs) are computer-based algorithms inspired by the structure and behavior of real neurons. Like the brain, they can recognize patterns, reorganize data and, most appealing, and learn from experience. The ANN consists of a set of processing units that simulate neurons and are interconnected via a set of weights in a way that allows signals to travel in parallel as well as serially.9 Of several possible ways to train an ANN, one of the most successful is the back-propagation algorithm.10 11

The purpose of this study was to model the relationship between atmospheric changes, including air pollution, and ED visits for respiratory conditions using an ANN. The ANN was trained to predict the number of ED visits for respiratory symptoms due to asthma, COPD, and acute and chronic bronchitis per day based on the input of meteorologic and air pollution data.


   Materials and Methods
TOP
 Abstract
 Introduction
 Materials and Methods
Results
 Discussion
 References
 
Patient Records
Ashkelon is located in the southwest of Israel. The region has approximately 400,000 inhabitants, with most living in villages and cities with a population of 15,000 to 100,000. Largely, a working middle-class population inhabits these cities and villages, and each has at least one light industrial area. A large power plant is situated on the extreme edge of the main urban center of Ashkelon. All the residents live within a 15-km radius of the Ashkelon city center, where the Barzilai Medical Center is located.

The Barzilai Medical Center has a large database of ED admittance records collected prospectively between January 1992 and December 1995. The records include demographic data: age, sex and address, and medical information, including symptoms and diagnosis defined by the ED physician and later confirmed by a medical secretary. The diagnosis was based on symptoms, physical examination findings, laboratory data including chest radiography, oxygen saturation measured by pulse oximetry, ECG, and response to treatment.

Group diagnoses can be classified into six main categories: (1) cardiac, including coronary artery disease (myocardial infarction, anginal pains), myocarditis, endocarditis, arterial hypertension, congestive heart failure, arrhythmias, and congenital cardiac malformations; (2) allergy, any kind of allergic response to food, animal, pollen, insect bite, or venom; (3) ear, nose, and throat disease, including tonsillitis, chronic ear nose and throat, otitis media or externa, sinusitis, rhinitis, peritonsillar abscess, and epistaxis; (4) asthma, including known disease, wheezing, and dyspnea responsive to bronchodilators and corticosteroids; (5) other respiratory disease, including pneumonia, COPD, acute and chronic bronchitis, upper respiratory tract infections, bronchiectasis, and pleural disease, and cough not defined as asthma; and (6) other, including all other visits, such as accidents, trauma, gynecology, etc. This database includes > 220,000 patient records (Table 1 ).


View this table:
[in this window]
[in a new window]

  		Table 1.. Distribution of Patients According to Demographics and Diagnosis

 
We selected the records of patients examined in the ED because of respiratory symptoms due to asthma, COPD, and acute or chronic bronchitis. Usable records were limited by the availability of meteorologic and air pollution data. Since we were interested in predicting the number of ED visits for respiratory symptoms (asthma, COPD, and acute and chronic bronchitis per day), we created a database of data sets representing 1,020 days of ED activity. The days accounted for in the study were those days that included all the meteorologic and air pollution data. The days that did not have all the meteorologic and air pollution data were excluded. Those days not included in the study were distributed evenly throughout the study period. Each data set included an input of meteorologic and air pollution data and an output of the number of patients visits to the ED for respiratory symptoms: asthma, COPD, and acute and chronic bronchitis on the corresponding day. These were randomized to 816 network-training data sets (80%) and 204 test data sets (20%).

Atmospheric and Air Pollution Data
Data collected on air pollution (CO, oxidation products of nitric oxide [NOx], SO2, total suspended particles [TSP], and O3) and meteorologic conditions (temperature, relative humidity, precipitation, barometric pressure, solar radiation, and wind speed and direction) were recorded by seven monitoring stations in 30-min intervals. The monitoring stations were scattered evenly throughout the Ashkelon area. The input data used for prediction was the average of the recordings of all the stations, after we determined that there were no significant differences among them.

ANNs
ANNs are gross models of actual neurons. Like the brain, they can recognize patterns and organize data, but most ominous is an inherent ability to learn. ANNs are typically composed of interconnected objects called units, which represent neurons. The units are connected by links, which act as the axons and dendrites. The link multiplies the output from a unit by a weighting factor, a value analogous to the connection strength at a synapse. The link then passes the weighted output value to another unit, which sums up the values passed to it by all other incoming links (Fig 1 , left). If the total input value exceeds a designated threshold value, the unit fires. Modifications in the firing patterns constitute the learning. This occurs as the weighting factors on the links change.



View larger version (21K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1.. Left: a unit output (Y) is a function (F) of the sum of weighted inputs (W1, W2, W3), which is derived from the preliminary processed inputs (x1, x2, x3) created from the raw inputs (X1, X2, X3). Right: a schematic diagram of a three-layer ANN; unit outputs (y1, yj) are a function of processed raw inputs (I1, I2, I3, Ik).

 
An ANN is typically composed of layers that are classified into three groups: a layer of input units, representing the information that is fed into the network; a layer of output units, representing the network output; and one or more hidden layers of units between them (Fig 1 , right). In order to train an ANN to perform a certain task, we must first determine how the units are connected (called the network architecture). Training involves presenting the neural network with specific examples, called data sets. These consist of a pattern of input activities for the input units, together with the desired pattern of activities for the output units. The data sets are presented repeatedly to the network, and the weights are updated after each presentation to constitute the learning.

There are several algorithms for ANN training. They can be classified into two groups: supervised and unsupervised learning. In supervised learning, the network is with input and output pairs and the algorithm is set to change the weight on the connections so that the output units give a better approximation of the desired output. In unsupervised learning, the network is presented with input data only and the algorithm is set to modify weights so that the data are categorized into groups based on similarities. We chose a supervised training algorithm called back propagation for network training.

Back Propagation
One of the most successful supervised training methods is the back-propagation algorithm. The basic concept is to use the derivative of an error function in order to find the direction that minimizes the error of the network and updating the weights accordingly.10 The error function most commonly used is the sum-squared error of the output units. The algorithm attempts to minimize the mean error over the entire training set, much as statisticians calculate a best-fit line for scattered points on a graph. We chose the back-propagation algorithm after using it successfully in a previous study.12

Our neural network application was custom programmed based on the conventional back-propagation equations using an Intel C++ Compiler; (Intel; Santa Clara, CA), rather then using commercially available software. This allowed us better control and flexibility during training.

Genetic Algorithms
Construction of an ANN begins with arbitrary decisions concerning the network architecture and other network parameters. These decisions could have substantial effect on network performance. For instance, as more units are added to the hidden layers, the ANN performance on the training data sets will improve; however, if the network is tried on independent data sets, its performance will first improve, then get worse. This phenomenon is called overfitting. It occurs when the ANN models the noise as well as the data in the training data sets. To avoid this and other training dilemmas, the search for the optimal network configuration can be automated using a genetic algorithm (GA),13 a method loosely based on the Darwinian principle of natural selection. The exact number of neurons was decided by the GA, which built the architecture automatically and chose the architecture that worked best.

The algorithm starts with a large random population of data strings called chromosomes representing ANNs. Each locus in the chromosome describes one of the ANN parameters. The population of ANNs undergoes a process of evolution through three steps: selection, crossing over, and mutation. In the selection step, each ANN is ranked according to its performance based on a fitness function. The better performing chromosomes are copied to form a new population, while the weaker ones are eliminated to maintain a constant population size. This equates to survival of the fittest. In the crossing-over step, combining parameters of the fittest ANNs generates new chromosomes. This represents mating between individuals. In the mutation step, random alterations of genes are introduced. This allows new parameter options to be evaluated.

This sequence of events, called a generation, is repeated until the optimal ANN configuration is found. There are many variations on the classical GA. We used a steady-state GA with a population size of 100 chromosomes.

We chose a fitness function that would take into account the ANN performance on both the training data sets (indicating successful training) and independent test data sets (indicating successful generalization). This was done by calculating the average output error for the training data sets multiplied by a factor of 0.6, and adding the average output error for the test data sets multiplied by a factor of 0.4. The reason for using a factor of 0.6 and 0.4, respectively, was to give the ability to correctly learn the training data sets a slight advantage.

Using the GA also allowed selecting which of the monitored meteorologic and air pollution variables are most predictive of ED visits for respiratory symptoms due to asthma, COPD, and acute and chronic bronchitis. This was done by incorporating the input variables used in ANN training into the chromosomes. The input variables were presented in several possible ways: (1) peak value during the 24 h prior to the ED visit, (2) average of values 1 day before ED admittance, (3) average of values during the 3 days before ED admittance, and (4) average of values during the 7 days before ED admittance.

We were interested in predicting the number of ED visits for respiratory symptoms on a particular day. In order to eliminate trends in the total number of visits to the ED, we used a relative index derived by dividing the number of visits for respiratory symptoms by the total number of ED visits for that day.


   Results
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
Discussion
 References
 
The GA was allowed to run and process ANNs until a plateau was reached in the performance of generated networks. The algorithm produced and evaluated 1,128 chromosomes representing neural networks. Figure 2 shows the structure of the ANN with the best performance. This was the 1,009th network generated by the GA. The network was trained with 816 training data sets, and then tested on 204 unlearned test data sets. It was able to predict the previously unlearned data sets with an average error of 12%.



View larger version (41K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2.. The structure of the best performing neural network produced by the GA.

 
Since the GA was allowed to select for the input variables used in prediction, we examined the 10 best performing neural networks in order to see which variables were most predictive. In all 10 networks, the monitored meteorologic and air pollution variables chosen were temperature, relative humidity, and barometric pressure of the meteorologic data, and SO2 and NOx of the air pollution data. In the neural network that performed best, these data were presented as both the peak value 24 h prior to the ED visit and the average of values during the 7 days before ED admittance.

No significant correlation between ED visits and atmospheric and air pollution data were determined despite using statistical methods as time-series analysis with trend and seasonal component and generalized additive model on 3-month cohorts. Seasonal variations do exist; however, no correlation with air pollution was found. Although the ANN does not provide a distribution of visits per day, the long period of observation intended to even out seasonal variations in the final analysis.


   Discussion
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
References
 
Many factors influence the development of bronchospasm in hyperactive airways. The factors that have been shown to affect the airways are air pollution (NOx, SO2, O3, and particles such as TSP, particulate matter of 10 µm, and particulate matter of 2.5 µm) as well as atmospheric and temperature changes.

NO2 is produced indoors by cooking with gas stoves. NO2 levels outdoors are mainly produced by cars.14 It has been shown that exposure to traffic exhaust has been related to wheezing.15 Exposure to NO2 exacerbated preexisting asthma in a panel of asthmatic volunteers and increased ED admissions for asthma in the acute phase16 and in chronic exposure.17 18

Exposure to SO2 also increases ED visits among patients with COPD.19 Field studies have shown prolonged clinical deterioration during and following episodes of air pollution, especially when patients were exposed to relatively high levels of SO2.20 Other oxidant pollutants such as O3 also increase ED visits due to acute asthma exacerbation.21 Studies have shown that temperature and O3, each separately and combined, worsen asthma symptoms.22 Symptoms become more severe 1 to 3 days after a decline in peak expiratory flow following exposure to changes in temperature.23

Particles, by themselves and in combination with oxidant pollutants, and changes in temperature have been reported to promote an increase in respiratory visits after a 24- to 48-h lag period, especially for asthma attacks.24 Temperature changes and combinations of gaseous pollutants were associated with increased asthma admissions to hospitals.5 25

Weather changes26 have been shown to affect airways, sometimes additionally with an increase of aeroallergens (such as pollens). The effect of acute or chronic exposure to a cold environment has been shown to induce responses in the airways, mainly congestion and bronchoconstriction. Bronchoconstriction, airway congestion, secretions, and decreased mucociliary clearance compromise pulmonary mechanics.27 These responses destabilize the airways, and induce shortness of breath, resulting in the need for medical assistance.

After taking the above-mentioned information into account, we are able to apply it using an ANN. ANNs have been used successfully in medical applications previously.28 29 Moseholm et al30 used neural networks to analyze the effect of weather and air pollution on asthmatic patients. We have already demonstrated that ANN technology can be used for modeling the effect of air pollution on ED visits for asthma, although with limited accuracy.12 Our new study is more accurate compared with the previous study published by our group. The previous study was short term and the analysis was on a smaller group of patients, compared to the longer term, larger population, and more advanced ANN technology with GA optimization of this study.

Most of the above-mentioned studies describing the effect of air pollutants and weather on respiratory symptoms found an effect while accounting for only a few variables. Such limitations are related to the lack of sensitivity of statistical analysis for multiple variables. In an attempt to overcome these limitations, we chose an ANN system.

Our study has investigated the predictive value of meteorologic and air pollution data on ED visits for respiratory symptoms of asthma and cough. The most important variables found were atmospheric temperature, barometric pressure, and relative humidity of the meteorologic data, and the levels of SO2 and NOx.

Changes in precipitation, solar radiation, wind speed and direction, O3 levels, and TSP levels were not found to be predictive by the GA. Input meteorologic and air pollution data were represented as the peak value within 24 h prior to ED admittance, and the average of values during the 7 days before ED admittance. There are many factors, such as indoor pollutants and indoor and outdoor allergens not recorded in our study, that may cause or aggravate the wheezing and shortness of breath and lead to referral to the ED.

The disadvantage in using ANN analysis is that the "black box" nature of neural networks provides little insight into the relative importance of the various input variables used in the model. Also, it is not possible to infer whether the predictors have a positive or negative impact on the output.

Another pitfall of ANNs is overfitting. This occurs when the neural network performs with great accuracy on the training data sets, but poorly on unlearned test data sets. We avoided this problem by including the ANN performance on independent (previously unlearned) data sets in the fitness function of the GA.

The ANN is a nonconventional "flexible" method, and the flexibility can be a disadvantage. However, this method is appealing for the following reasons: (1) it "learns" easily without supervision, (2) the worker does not need computer or mathematical experience, and (3) software is now readily available.

The ANN can assist in detecting trends of meteorologic and air pollutant influence on the airways. These are reflected by ED visits due to respiratory symptoms.

We have shown that an ANN can be used to predict ED visits for respiratory conditions based on meteorologic and air pollution data. The next logical step would be to use neural networks to predict exacerbations in individual patients. Neural network technology is readily available in numerous software packages that physicians could use and custom tailor for their patients.


   Footnotes
 
Abbreviations: ANN = artificial neural network; ED = emergency department; GA = genetic algorithm; NOx = oxidation products of nitric oxide; TSP = total suspended particles

Received for publication March 7, 2001. Accepted for publication May 21, 2002.


   References
TOP
 Abstract
 Introduction
 Materials and Methods
 Results
 Discussion
 References
 

  1. Peters, A, Goldstein, IF, Beyer, U, et al (1996) Acute health effects of exposure to high levels of air pollution in Eastern Europe. Am J Epidemiol 144,570-581[Abstract/Free Full Text]
  2. Delfino, RJ, Coate, BD, Zeiger, RS, et al Daily asthma severity in relation to personal ozone exposure and out door fungal spores. Am J Respir Crit Care Med 1996;154,633-641[Abstract]
  3. Lintner, TJ, Brame, KA The effects of season, climate, and air-conditioning on the prevalence of Dermatophagoides mite allergens in house hold dust. J Allergy Clin Immunol 1993;91,862-867[CrossRef][ISI][Medline]
  4. Subiza, J, Masiello, JM, Subiza, JL, et al Prediction of annual variations in atmospheric concentrations of grass pollen: a method based on meteorological factors and grain crop estimates. Clin Exp Allergy 1992;22,540-546[CrossRef][ISI][Medline]
  5. Rossi, OV, Kinnula, VL, Tienari, J, et al Association of severe asthma attacks with weather, pollen, and air pollutants. Thorax 1993;48,244-248[Abstract/Free Full Text]
  6. Forsberg, B, Stjernberg, N, Falk, M, et al Air pollution levels, meteorological conditions and asthma symptoms. Eur Respir J 1993;6,1109-1115[Abstract]
  7. Beasley, R, Roche, WR, Roberts, JA, et al Cellular events in the bronchi in mild asthma and after bronchial provocation. Am Rev Respir Dis 1989;139,806-817[ISI][Medline]
  8. Ollerenshaw, SL, Woolcock, AJ Characteristics of the inflammation in biopsies from large airways of subjects with asthma and subjects with chronic airflow limitation. Am Rev Respir Dis 1992;145,922-927[ISI][Medline]
  9. Fausett, L Fundamentals of neural networks: architectures, algorithms and applications. 1994 Prentice-Hall International Englewood Cliffs, NJ.
  10. Camp, DV Neurons for computers. Sci Am 1992;267,125-127
  11. Hilton, GE How neuronal networks learn from experience? Sci Am 1992;267,105-109
  12. Nutman, A, Solomon, Y, Mendel, S, et al The use of a neural network for studying the relationship between air pollution and asthma-related emergency room visits. Respir Med 1998;92,1192-1202
  13. Miller, G, Jodd, P, Hegde, S Designing neural networks using genetic algorithms. Schaffer, JD eds. Proceedings of the 3rd International Conference on Genetic Algorithms. 1989,379-384 Morgan Kaufmann San Mateo, CA.
  14. Studnika, M, Hackl, E, Pischinger, J, et al Traffic-related NO2 and the prevalence of asthma and respiratory symptoms in seven year olds. Eur Respir J 1997;10,2275-2278[Abstract]
  15. Oosterlee, A, Drijver, M, Lebret, E, et al Chronic respiratory symptoms in children and adults living along street with high traffic density. Occup Environ Med 1996;53,241-247[Abstract/Free Full Text]
  16. Castellsague, J, Sunyer, M, Saez, JM Short-term association between air pollution and emergency room visits for asthma in Barcelona. Thorax 1995;50,1051-1056[Abstract/Free Full Text]
  17. Pearlman, ME, Finklea, JF, Creason, JP, et al Nitrogen dioxide and lower respiratory illness. Pediatrics 1971;47,391-398[Abstract/Free Full Text]
  18. Neas, LM, Dockrey, DW, Ware, JH, et al Association of indoor nitrogen dioxide with respiratory symptoms and pulmonary function in children. Am J Epidemiol 1991;134,204-219[Abstract/Free Full Text]
  19. Sunyer, J, Saez, M, Murrillo, C, et al Air pollution and emergency room admissions for chronic obstructive pulmonary disease: a five-year study. Am J Epidemiol 1995;137,701-705[Abstract/Free Full Text]
  20. Stebbing, JH, Fogelman, DG Identifying a susceptible subgroup: effects of the Pittsburgh air pollution episode upon school children. Am J Epidemiol 1979;110,27-40[Abstract/Free Full Text]
  21. Romieu, I, Meneses, F, Sienra-Monge, JJ, et al Effects of urban air pollutants on emergency room visits for childhood asthma in Mexico City. J Epidemiol 1995;141,546-553
  22. Lebowitz, MD, Holberg, CJ, Boyer, B, et al Respiratory symptoms and peak flow associated with indoor and out door air pollution in the southwest. J Air Pollut Control Assoc 1985;35,1154-1158[ISI][Medline]
  23. Lioy, PJ, Vollmuth, TA, Lippmann, M Persistence of peak flow decrement in children following ozone exposure exceeding the national air quality standard. J Air Pollut Control Assoc 1985;35,1068-1071
  24. Bates, DV, Sizto, R Relationship between air pollution levels and hospital admissions in Southern Ontario. Can J Public Health 1983;74,117-133[ISI][Medline]
  25. Ponka, A Asthma and low level air pollution in Helsinki. Arch Environ Health 1991;46,262-270[ISI][Medline]
  26. Celenza, A, Fothergill, J, Kupek, E, et al Thunderstorm associated asthma: a detailed analysis of environmental factors. BMJ 1996;312,604-607[Abstract/Free Full Text]
  27. Giesbrecht, G The respiratory system in cold environment. Aviat Space Environ Med 1995;66,890-902[Medline]
  28. Baxt, WG The application of the artificial neural networks in pathology and medical laboratories. Lancet 1995;346,1135-1138[CrossRef][ISI][Medline]
  29. Dybowski, R, Weller, P, Chang, R, et al Prediction of outcome in critically ill patients using artificial neural network synthesised by genetic algorithm. Lancet 1996;347,1146-1149[CrossRef][ISI][Medline]
  30. Moseholm, L, Taudorf, E, Frosting, A Pulmonary function changes in asthmatics associated with low-level SO2 and NO2 air pollution, weather, and medicine intake. Allergy 1993;48,334-344[ISI][Medline]




This Article
Right arrow Abstract Freely available
Right arrow Full Text (PDF) Free
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in ISI Web of Science
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Add to My Personal Article Archive
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via ISI Web of Science (10)
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Bibi, H.
Right arrow Articles by Nutman, J.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Bibi, H.
Right arrow Articles by Nutman, J.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS