Text

Clinical decision-support systems (CDSS) can support experts‘ decision-making by exploiting big data. For example, a classification integrated into the CDSS can provide a statistical proposition of which hearing device a patient would benefit from; or data-driven, unsupervised approaches can identify distinct patient groups in the data. The main challenge is to base such a CDSS on international, really “big data”, since local clinical-audiological databases comprise different audiological tests and test conditions, data structure and formats, expert knowledge, or patient populations. Speech recognition tests are widely employed to assess hearing device indication criteria or benefit, but differences in speech material, language, or outcome variable (discrimination loss or speech recognition thresholds) hinder integration of different data sets. To assess the comparability of these two interpretation modes, this study investigates the potential to estimate SRTs from clinical data collected to characterize the discrimination loss (1 – maximum word recognition score (WRS_max)). The two variables are conceptually linked via the psychometric function, which describes speech recognition as a function of speech level or SNR.

The relationship between SRT and WRS_max for the Freiburg monosyllabic speech test (FMST) was assessed based on retrospective analysis of the clinical database of Hannover Medical School, on a data set comprising about 27,000 patients with FMST in quiet and pure-tone audiogram conducted on the same day. The FMST word recognition score was measured for different test lists presented at fixed levels, with levels chosen depending on the patients" hearing thresholds and the obtained results at previous levels, until the maximum word recognition score was achieved. Depending on data availability, different SRT estimation procedures were based solely on measured data or, if sufficient data were unavailable, complemented by model-based information about the slope of the speech intelligibility index (SII). Psychometric function fits and linear fits covering only the assumed linear slope range were compared. Monte Carlo simulations were applied to estimate the SRT error from the intelligibility-dependent confidence intervals of the WRS.

Similar and plausible SRTs can be estimated using all procedures, with an error depending on the test-retest reliability of the FMST. The advantage of the psychometric function fit is that it takes into account all data points and their relative dependencies, while the linear fit is more controlled in the level range close to the SRT.

Depending on data availability at different speech levels, SRTs can be estimated on available measured data, with different limitations for some special (data) cases. If data close to the SRT are not available, this can be compensated for by assuming an SII-based slope for the psychometric function or linear fit. Both fit versions can be used interchangeably for most patients.