Optimal-design domain-adaptation for exposure prediction in two-stage epidemiological studies

BACKGROUND

In the first stage of a two-stage study, the researcher uses a statistical model to impute the unobserved exposures. In the second stage, imputed exposures serve as covariates in epidemiological models. Imputation error in the first stage operate as measurement errors in the second stage, and thus bias exposure effect estimates.

OBJECTIVE

This study aims to improve the estimation of exposure effects by sharing information between the first and second stages.

METHODS

At the heart of our estimator is the observation that not all second-stage observations are equally important to impute. We thus borrow ideas from the optimal-experimental-design theory, to identify individuals of higher importance. We then improve the imputation of these individuals using ideas from the machine-learning literature of domain adaptation.

RESULTS

Our simulations confirm that the exposure effect estimates are more accurate than the current best practice. An empirical demonstration yields smaller estimates of PM effect on hyperglycemia risk, with tighter confidence bands.

SIGNIFICANCE

Sharing information between environmental scientist and epidemiologist improves health effect estimates. Our estimator is a principled approach for harnessing this information exchange, and may be applied to any two stage study.

Statistical parsing of varieties of clinical Finnish

OBJECTIVES

In this paper, we study the development and domain-adaptation of statistical syntactic parsers for three different clinical domains in Finnish.

METHODS AND MATERIALS

The materials include text from daily nursing notes written by nurses in an intensive care unit, physicians' notes from cardiology patients' health records, and daily nursing notes from cardiology patients' health records. The parsing is performed with the statistical parser of Bohnet (http://code.google.com/p/mate-tools/, accessed: 22 November 2013).

RESULTS

A parser trained only on general language performs poorly in all clinical subdomains, the labelled attachment score (LAS) ranging from 59.4% to 71.4%, whereas domain data combined with general language gives better results, the LAS varying between 67.2% and 81.7%. However, even a small amount of clinical domain data quickly outperforms this and also clinical data from other domains is more beneficial (LAS 71.3-80.0%) than general language only. The best results (LAS 77.4-84.6%) are achieved by using as training data the combination of all the clinical treebanks.

CONCLUSIONS

In order to develop a good syntactic parser for clinical language variants, a general language resource is not mandatory, while data from clinical fields is. However, in addition to the exact same clinical domain, also data from other clinical domains is useful.