BEYOND ATOPY: MULTIPLE PATTERNS OF SENSITIZATION IN RELATION
TO ASTHMA IN A BIRTH COHORT STUDY
ONLINE DATA SUPPLEMENT
Angela Simpson, Vincent Y. F. Tan, John Winn, Markus Svensén, Christopher M. Bishop,
David E. Heckerman, Iain Buchan, Adnan Custovic
1
METHODS
Study populations The Manchester Asthma and Allergy Study is an unselected, population-based
prospective study which follows the development of asthma and other atopic disorders in
a cohort of children. The setting is the maternity catchment area of Wythenshawe and
Stepping Hill Hospitals, comprising of 50 square miles of South Manchester and
Cheshire, UK, a stable mixed urban-rural population. Study was approved by the Local
Research Ethics Committee. Informed consent was obtained from all parents.
Screening & Recruitment
ththAll pregnant women were screened for eligibility at 'Booking' antenatal visits (8-10
week of pregnancy). The study was explained to the parents, and informed consent for
initial questionnaires and skin prick testing was obtained. Both parents completed a
questionnaire about their and their partner‟s history of asthma and allergic diseases and
smoking habits.
If the pregnant woman‟s partner was not present at the antenatal clinic visit, an invitation
was sent for him to attend an open-access evening clinic for skin prick testing and
questionnaire. Once both parents had completed questionnaires and skin prick testing,
a full explanation of the proposed future follow-up for the child was given.
Of the 1499 couples who met the inclusion criteria (<10 weeks of pregnancy, maternal
age >18 years, questionnaire and skin test data available for both parents), 288 declined to take part in the study.
Of the 1186 participants with any evaluable data, 133 who were randomized into the
primary prevention study were excluded from the analysis of the association between
clinical outcomes and inferred sensitization class.
2
Follow-up
The children have been followed prospectively, and attended review clinics at ages 1, 3,
5 and 8 years (?4 weeks). At age 1 year, only high and low risk children were invited to
attend for clinical follow up. At all other time points for all other measures all children
were invited to participate.
Definitions of exposures and outcomes
Atopic sensitization
Atopic sensitization was ascertained by skin prick testing at age 1, 3, 5 and 8 years (D
pteronyssinus, cat, dog, grasses, moulds, milk, egg [Bayer, Elkahrt, IN, USA]). We
defined sensitization as a mean weal diameter 3mm greater than negative control to at
least one of the allergens tested. We also measured specific serum IgE to mite, cat, dog,
TMgrasses, milk, egg and peanut by ImmunoCAP (Phadia, Uppsala, Sweden) collected at the four time points. We defined allergen-specific sensitization as mean wheal
diameter at least 3mm greater than the negative control and/or specific IgE?0.35kU/l.
Conventional definition considered a child to be atopic if he/she had allergen-specific
sensitization to at least one allergen. Children with any positive test (skin test or IgE) at
any time point were considered to be atopic ever, and those with no evidence of
sensitisation as never atopic.
Wheeze
A validated ISAAC questionnaire was administered by a trained interviewer to collect
information on parentally reported symptoms, physician-diagnosed illnesses and
treatments received.
Lung function
At age 3 and 5 years we carried out measurements of specific airway conductance (sG) aw
to assess airway function in all children who were willing to cooperate. Measurements
were made using a constant volume whole body plethysmograph (Masterscreen Body
3
4.34; Jaeger, Würzburg, Germany). Flow and volume were measured with a heated
-1?s differential pressure screen-type pneumotachograph with a resistance of 0.036 kPaand a dead space of 160mls. Pressure measurements were made with a pressure transducer (Nr.660.99007; Hube Control AG, Wuerenlos, Switzerland) with an input range of ?100 Pa, a resolution of 0.05 Pa and a linear response up to 10 Hz. The
plethysmograph was calibrated daily. Sensors in an ambient unit supplied with the plethysmograph recorded ambient data on temperature, humidity and barometric pressure. The pneumotachograph was volume calibrated according to the American Thoracic Society recommendations using a 2 L syringe at flow rates of 0–1.5, 1.5–5
and >5 l/s. The half value period was calibrated to ensure a specific leakage in the box of 4-7 seconds.
The pressure transducer was calibrated using a 50 mL motor driven piston pump to generate sinusoidal variations of plethysmographic pressure. Electronic body temperature, pressure, and saturation (BTPS) compensation was applied throughout, using a time-shift of 60 ms.
sG is measured by a single-step procedure from the simultaneously measured aw
changes of respiratory flow and changes of plethysmographic pressure, omitting the measurement of TGV. Measurements were carried out during tidal breathing using a facemask, which was adapted by fitting a standard paediatric facemask with a non-compressible mouthpiece made from silicone tubing. The end of the tubing was made rigid with an aluminium splint. The purpose of this was to maintain stable airway opening, prevent nose breathing and support the cheeks. The procedure was explained to the accompanying adult and the use of the facemask demonstrated to the child. The children were encouraged to sit in the plethysmograph alone but if they refused, the accompanying adult, usually a parent, accompanied the child in the plethysmograph
4
cabinet with the child seated on their knee. The door of the plethysmograph was closed
and the subject asked to breathe through the facemask.
Children were encouraged to breathe at a rate of 30-45 breaths per minute. If a parent
accompanied the child, the adult was asked to inhale and hold their breath for
measurements were made once a stable breathing approximately 20 seconds. sGawpattern had been re-established. Once a stable breathing pattern was established, at
least three measurements of sG were performed, and each was calculated from the awmeans of 5 consecutively measured technically acceptable loops (each child performed
at least 15 loops). The median of these 3 measurements of effective sG was used in aw
the analysis. The measured values of sG were corrected for the influence of the aw
pneumotachograph screen and for the volume displacement caused by the subject (or
subject + parent).
Children were asymptomatic at the time of assessment of lung function.
Airway reactivity - Methacholine Challenge
Airway reactivity was assessed using a 5 step protocol performed according to ATS
guidelines. The methacholine (acetyl-β-methylcholine chloride) solutions were prepared
with sterile normal saline (Stockport Pharmaceuticals, UK). Quadrupling doses of
methacholine (0.0625 – 16.0 mg/mL) were delivered to subjects via a DeVilbiss 646 nebuliser (Sunrise Medical HHG, Somerset, PA) and a KoKo dosimeter (Pulmonary Data
Services, Doylestown, PA) calibrated to deliver 0.009 mL per 0.6s actuation. The dosing
schedule is described in Table E1. The test was explained to the subject and the best
baseline FEV measurement performed in the wedge bellow spirometer was recorded. 1
The predicted FEV was calculated and if the measured values was <1.0 l or less than 1
60% predicted the test was not performed. If the child was unable to produce
reproducible FEV measurements the procedure was abandoned. Assuming the child 1
met the criteria to continue, the 20% drop from the child‟s baseline value was calculated
5
so that the operator would know when to stop the test. After normal tidal expiration to
FRC (functional residual capacity) the dosimeter was triggered at the onset of inspiration,
and the subject asked to inhale slowly and deeply over 6 s. Subjects were instructed to
was measured 30 and 1hold their breath for 5 s, followed by slow exhalation for 5 s. FEV
90 seconds after 5 inhalations of each dose of methacholine. The challenge was
stopped when either a 20% fall in FEVwas observed, or the maximum methacholine 1 concentration had been administered with a fall of less than 20% in FEV. 1
Children were categorized as having a positive or a negative challenge based on
whether or not they reached a 20% fall in FEV by the final dose of the challenge 1
(16mg/ml).
Hospital admission for asthma/wheeze:: The UK health care system ensures that a single medical record is held by the primary care physician which provides a full record
of all encounters with health professionals. GPs are legally required to maintain accurate
records of all medical encounters of their patients, including retention of all records of
hospital encounters. A trained physician reviewed the written and computerized primary
care medical records and extracted the data on hospitalisations for wheeze or asthma.
Data analysis
We took a machine learning approach to the data analysis. Using a Hidden Markov
Model, the available physiological measurements of skin prick tests (SPTs) and serum
specific IgE tests (SITs) to a panel of allergens were used to infer one multinomial latent
variable per child to cluster the children in an un-supervised manner into different
sensitization classes (the model is shown in Figure 1).
At the core of the model are the 4 binary latent variables for each allergen labelled
„Acquired Sensitization‟ and these are linked together in a Markov chain across the 4
time points. We inferred time-dependent transition probabilities (i.e. the probabilities of
6
gaining and losing sensitization at each age) which were assumed to be shared by all
children in each sensitization class, but differing between classes. In our model, for
ease of inference, we placed conjugate priors on all the variables that were to be inferred
- using beta priors as the variables of interest were binary and beta is conjugate to the
E1. We also observed that our results were insensitive to the choice binomial distribution
of hyperparameters (the parameters that define the prior distributions).
Inference
Inference was performed using Infer.NET (http://research.microsoft.com/en-
us/um/cambridge/projects/infernet/), a Microsoft-owned library of statistical algorithms for
large-scale Bayesian inference. We inferred the false and true positive rates of the SPT
and IgE tests, missing SPT and IgE values, the state-transition and observation
(emission) probabilities, the acquired sensitization variable and finally, the sensitization
class. An approximate Bayesian inference method (Variational Message Passing-
18VMP) was used to perform the inference in an efficient manner.
Handling the missing data
Variables corresponding to missing data values were included in the model but treated
as unobserved. Distributions over these missing data values were also computed using
VMP based on the available measurements. This approach assumes that the missing
values are missing completely at random (MCAR).
Training and validation data sets with multiple imputations and assessment of the
robustness of the clustering
To determine the appropriate number of classes, differing numbers of clusters were
tested as to their ability to predict the sensitization state of children where that state was
artificially made missing. The process starts by randomly dividing the data so that 80%
7
formed a training set and the remaining 20% formed a validation set. Using the training set, a clustering is learned by computing posterior distributions over the parameters of the sensitisation HMM using the variational message passing (VMP) inference algorithm. Hence, for each cluster, distributions were learned over the probability of initial sensitisation and the probabilities of gaining and retaining sensitisation for each cluster. In addition, common distributions were learned over probabilities of positive tests given sensitisation or lack of sensitisation.
This learned clustering was validated using an imputation experiment, where removed data values were predicted under the learned clustering. In each run of the experiment, the posterior distributions learned in the initial clustering (using the training data) were used as corresponding prior distributions in a new clustering model, used to cluster the validation data. 20% of the values in the validation data were removed at random and then predicted using the posterior distributions of the new clustering model.
Because VMP can be sensitive to its random initialisation, the training process was repeated for 10 different such initialisations and the clustering with the best score selected. To avoid bias due to the training/validation splits itself, the entire process was repeated for 10 different random training/validation data splits. The imputation score was computed as the sum of the log probability of the removed values under their inferred posterior distributions, averaged across the 10 runs. Results for models with 1–7 clusters
are shown in Figure E3; note that the baseline in the figure has been adjusted so that the (unique) single cluster model has a score of zero.
The robustness and repeatability of the clustering process to small changes in the data set was assessed by comparing the clusterings given by the 10 random
training/validation splits. For each clustering with a given number of clusters, a confusion matrix was computed indicating how frequently children were assigned to the
8
same cluster in the other nine clusterings. The clustering with the best confusion matrix (defined as the matrix with the largest sum of diagonal elements) was selected as the reference clustering for the given number of clusters. The confusion matrices for the reference clusterings with 2 and 5 clusters is shown in Tables E4 and E5, demonstrating no confusion between clusters in the 2-class case and very little confusion between the clusters in the 5-class case.
Sensitization Class
This is a multinomial variable indicating to which sensitization class each child belongs (out of between 2 and 5 classes). The model assumes that each child belongs to one of these classes. We investigated a two-class and a five-class model. The number of clusters was chosen as the maximum that contained sufficient observations for a statistically credible inference, and had a structure that was plausible. For example, if the number of clusters was set to 6, the number of children in some clusters will be very small. In some cases, some clusters may even not contain a single individual.No assumptions were made about the nature of each class. We assumed only that children in different classes have different state-transition probabilities, but they have the same observation probabilities across time. During inference, a distribution was computed for each child giving the probability of their belonging in each class. For further analysis, we assumed the child belonged to the highest probability class - a maximum a-posteriori approach.
We then investigated the association between the clinical outcomes and the classes which had been inferred in a completely unsupervised manner. The relation between each class and relevant clinical outcomes were tested in models that adjusted for known confounding factors, effect modifiers and multiple testing.
9
TABLES
Table E1. Dosing schedule for methacholine challenge
Step Methacholine Methacholine dose Cumulative
concentration (mg/ml) (mg) methacholine dose (mg) 1 0.0625 0.003 0.003 2 0.25 0.011 0.014 3 1.0 0.045 0.059 4 4.0 0.180 0.239 5 16.0 0.720 0.959
10