Objective Comparison of Auditory Profiles Using Manifold Learning and Intrinsic Measures.
Chen Xu, Birger Kollmeier, Lena Schell-Majoor
Among eight hearing-loss classification systems tested on the same 1,127-patient dataset, the Bisgaard audiogram profiles separated patients most cleanly and the audiometric phenotype least; adding speech-in-noise and loudness measures did not beat simple audiogram-based groupings.
Key findings
1The Bisgaard audiogram profiles achieved the lowest normalized Davies-Bouldin score (best clustering), while the audiometric phenotype had the highest (worst); the remaining six frameworks were statistically comparable.
2Vector quantization (used in the Bisgaard profiles) produced significantly lower Davies-Bouldin scores than the Gaussian Mixture Model (used in the general phenotype) when both were set to 10 profiles on the same data.
3The two comprehensive profiles that include supra-threshold measures (BEAR and Hearing4all) did not outperform the simpler audiogram-based profiles on any of the normalized intrinsic measures, PCA, or t-SNE.
Still to come
How it was doneWhat they found
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
All eight systems were tested on a single dataset of primarily older adults (mean age 67.2) with hearing impairments, so the ranking may not hold for younger or more diverse populations. Under the audiometric phenotype framework, 704 of 1,127 participants could not be assigned to any class, and under the general phenotype 545 were unclassified; these large 'unidentified' groups were treated as an extra cluster, which inflates their apparent separability. The two-dimensional PCA and t-SNE plots may hide separations that exist in the full 37-dimensional space; the authors acknowledge this. The logarithmic normalization used to make the scores comparable across different numbers of groups is described by the authors as a heuristic, not a theoretically derived correction. The paper compares how well the systems cluster data statistically; it does not test whether any of them leads to better hearing-aid fitting, speech-in-noise performance, or patient outcomes. The authors note that the Hearing4all profiles lack expert validation and have low interpretability, and that the correspondence between the different frameworks remains unclear.
Declared interests
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy (Project ID 390895286). The authors declared no conflicts of interest.
The easy way to misread this
Do not read 'Bisgaard profiles cluster best' as 'Bisgaard profiles are the right way to classify your patient for treatment.' The paper measures statistical separability in a 37-dimensional space, not clinical utility. The authors themselves stress that a framework's usefulness should be judged by the clinical or interpretative value of the resulting profiles, not just by how cleanly the groups separate, and they explicitly state that the practical applications (hearing-aid fitting, individualized treatment) are not addressed in this study.
Summarised by AI from the full paper, without a clinician reviewing it. Check it against the source before it changes what you do. Read it on PubMed →