Department: Electrical & Computer Engineering
Faculty Advisor(s): Shankar Subramaniam

Primary Student
Name: Tejaswini Narayanan
Email: tnarayan@ucsd.edu
Phone: 858-822-0960
Grad Year: 2012

Clinical evidence suggests that exposure to common allergens plays an important role in the development of childhood allergies and asthma. We present a machine learning approach for modelling the effect of allergen-specific exposure levels on predicting the likelihood of developing childhood allergies. Specifically, we examine the effects of dust-mite allergen levels at home on developing atopic sensitivity to mite. Our dataset is derived from a population-based birth cohort consisting of 1028 children, for which atopy tests such as Skin Prick Test (SPT) and Immunoglobulin E (IgE) tests have been performed at ages 1, 3, 5 and 8. Dust and allergen levels were also measured from 10 different "dust reservoirs" (such as child mattress, living room floor) at ages 1, 3, 5. As an extension to an existing approach, we build an augmented Hidden Markov Model (HMM) to cluster patient data into patterns of mite exposure and sensitivity over time. In particular, we incorporate exposure levels to dust-mite as a first-class model parameter, and investigate the relation between the emergent clusters and the ones produced from a modelling approach which did not account for exposure levels to mite. Inference was performed using Infer.NET, a library for large-scale Bayesian inference. An approximate Bayesian inference method was used to perform the inference in an efficient manner. We observe that the output parameters resulting from our modelling approach provide interesting insights into the effect of mite exposure on atopic sensitivity. A two-class dichotomous classification emerged from our model which accounts for exposure-levels to specific allergens. We compare this outcome to the classification produced by a model which is agnostic to allergen-specific exposure. In particular, we identify one cluster primarily constituted by children who were positively sensitised to dust-mite and hence corresponded to the atopic cluster, and the other cluster was non-atopic. While the emergent classes from both models were comparable in terms of cardinality and membership, the probabilistic model parameters were significantly different across the models. For example, we observe from our model that the probability of gaining sensitisation to dust-mite is markedly different for children with and without exposure to dust-mite; visibility into this association between exposure and sensitisation was not available from the model that did not account for exposure. Our results reaffirm the utility and value of machine learning techniques as effective tools in the domain of applied healthcare.

Related Links:

  1. http://research.microsoft.com/infernet

« Back to Posters or Search Results