Bottom-up Subspace Clustering Based Occupational Hearing Loss Signal Detection

Occupational hearing loss (OHL) is a common and serious health problem, so early hearing hazard signal detection is important in occupational health management. As some people may be more vulnerable to certain specific occupational exposures, it is necessary to provide them better protection by adaptively detecting constrained background based signals from occupational health examination data (OHED). In this paper we propose a bottom-up subspace clustering-based signal detection approach, which consists of three steps: subgroup generation, signal extraction, and signal selection. Bottom-up subspace clustering is utilized to generate background subgroups, which serve as constrained backgrounds in disproportionality analysis. To avoid missing important subgroups, we apply various algorithms and validity indices to construct initial subgroups. As subgroups are constructed by iteratively intersecting existing subgroups, most of them are small in size. As a result, traditional proportional reporting ratio (PRR) will lead to an unacceptable fake positive rate in signal extraction. Here a simulation based approach is proposed for evaluating small samples' disproportionality. Yet, the number of extracted signals remains very large because there are many similar signals. So a hybrid score, which combines the simulation based score and shrinkage PRR, is presented for selecting representative signals. We have analysed an OHED of 2051 records. Experimental results suggest the algorithm can effectively detect potential occupational hearing hazard signals.