Oldenburg Hearing Health Record (OHHR)

• OHHR is a large-scale, open-access clinical–audiological dataset featuring 1,127 adult profiles that capture diverse auditory abilities via comprehensive audiometric, psychoacoustic, and cognitive tests.
• It employs rigorous data acquisition and preprocessing protocols—including z-scoring, imputation, and dimensionality reduction—to ensure reproducibility and robust evaluation of auditory profiling methods.
• The dataset supports precision audiology research by facilitating systematic comparisons of clustering indices and enabling advanced phenotyping for individualized hearing diagnostics.

The Oldenburg Hearing Health Record (OHHR) is an openly available, large-scale clinical–audiological dataset designed to enable systematic comparison and development of auditory profiling frameworks and to advance phenotyping approaches in precision audiology. The extended OHHR comprises standardized, multivariate data from 1,127 adults with a broad spectrum of hearing abilities and incorporates a comprehensive test battery spanning audiometric, psychoacoustic, speech-in-noise, cognitive, and self-reported health measures. All procedures, variables, and preprocessing steps adhere to rigorous protocols, supporting reproducibility and robust cross-methodological evaluation (Xu et al., 7 Jan 2026).

1. Cohort Composition and Data Acquisition

The extended OHHR features 1,127 adult participants (mean age 67.2 years, SD 12.0; 55.7% male, 44.3% female) recruited from routine diagnostic and research evaluations at the University of Oldenburg Medical Physics & Acoustics clinic. Inclusion criteria encompass the full range from clinically normal hearing to moderately-severe sensorineural hearing loss. The recruitment protocol captures a diverse cross-section of hearing profiles, facilitating generalizability of data-driven analyses.

Audiometric and speech-in-noise testing were conducted in IEC 60645-1–compliant sound-treated booths by certified hearing-aid acousticians. Adaptive categorical loudness scaling (ACALOS) was administered with randomized frequency and level sequences to mitigate response bias, following Oetting et al. (2014). Cognitive assessments and self-report instruments were performed under standard laboratory conditions, and all raw data were subjected to plausibility checks and outlier review. Exclusion was rare, with <1% of the dataset omitted due to questionable test adherence.

2. Test Battery, Variables, and Feature Matrix

Each participant completed a standardized battery from which 37 summary parameters were derived:

• Pure-tone air conduction audiometry at 11 frequencies per ear (0.25–8 kHz), with thresholds captured in dB HL. Pure-tone average (PTA4) was computed across 0.5–4 kHz.
• ACALOS measures at 1, 1.5, and 4 kHz provided per-frequency parameters: transition level (LCUT), medium loudness level (L25), uncomfortable loudness level (L50), and slope at low levels (MLOW), totaling 15 supra-threshold loudness-scaling variables.
• Speech-in-noise proficiency assessed via Goettingen Sentence Test (SRT_GÖSA) and Digit Triplet Test (SRT_DTT), both yielding SRTs in dB SNR.
• Cognitive/psychosocial measures, including:
  • DemTect cognitive screening score
  • Verbal intelligence estimate
  • SF-12 health-related quality of life: physical (PCS) and mental (MCS) composite scores
  • A brief hearing-history questionnaire (self-reported hearing difficulties, hearing-aid use)

The data are organized as an 1127 × 37 feature matrix, where each row corresponds to a participant and columns represent numeric or integer-coded summary variables. File structure includes a comma-separated values (.csv) file and a JSON data dictionary describing variable names, units, ranges, and types.

Block	Variables (example names)	Data type (units)
Audiometry	HL_L_0.5kHz, HL_R_4kHz, PTA4	Numeric (dB HL)
Loudness scaling	LCUT_4kHz, L25_1.5kHz, MLOW_1.5kHz	Numeric (loudness units)
Speech-in-noise	SRT_GOSA, SRT_DTT	Numeric (dB SNR)
Cognitive/survey	DemTect, SF12_PCS, IQ_verbal	Numeric (std. scores)
Self-report	Questionnaire items	Integer/ordinal

3. Preprocessing and Standardization

Comprehensive preprocessing ensures analytic reliability and comparability:

• All 37 variables are z-scored to zero mean and unit variance prior to multivariate methods including PCA and clustering.
• Missing data (<2% of entries) are imputed via column-wise median.
• Outliers exceeding ±3 z are winsorized to ±3.
• For dimensionality reduction, PCA (FactoMineR in R) retains five components, with the first two commonly visualized; t-SNE (Rtsne in R, perplexity=30, θ=0.0) is also employed for nonlinear embeddings.

This preprocessing protocol provides a unified input representation for downstream machine learning and profiling frameworks.

4. Clustering and Intrinsic Evaluation Metrics

The OHHR enables rigorous unsupervised evaluation of auditory-profile assignments using three standard intrinsic clustering indices, each applied to clusterings over the full 1127 × 37 feature matrix:

• Davies–Bouldin Index (DB):

$$\mathrm{DB} = \frac{1}{K} \sum_{i=1}^K \max_{j \neq i} \frac{S_i + S_j}{M_{ij}}$$

where $K$ is cluster count, $S_i$ is within-cluster scatter, $M_{ij}$ is centroid distance. Lower values denote superior separation and compactness.

• Calinski–Harabasz Score (CH):

$$\mathrm{CH} = \frac{\mathrm{trace}(B)/(K-1)}{\mathrm{trace}(W)/(N-K)}$$

Trace terms represent between- and within-cluster sums of squares; larger CH indicates more distinct clustering.

• Silhouette Index (SI):

$$s(i) = \frac{b(i)-a(i)}{\max\{a(i),\,b(i)\}}, \quad \mathrm{SI} = \frac{1}{N}\sum_{i=1}^N s(i)$$

$a(i)$: average distance to cluster mates; $b(i)$: lowest mean distance to another cluster. SI near +1 reflects strong clustering.

All indices are normalized with respect to cluster number: for $K$ profiles, each metric is divided or multiplied by $\log_2(K)$ as appropriate. This standardization addresses class-count confound and allows direct comparison across varied profiling approaches.

5. Data Access, Structure, and Citation

The extended OHHR is fully open access and archived on Zenodo (https://doi.org/10.5281/zenodo.16919812). Each release includes the primary data matrix (CSV), a comprehensive JSON data dictionary, and supporting metadata. Researchers are instructed to cite:

• Jafri, S. et al. (2025). OHHR – The Oldenburg Hearing Health Record (v1.3.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.16919812

Standardization of structure (fixed feature set, explicit dictionary) and open-source example code (R, Python) facilitate rapid application of a broad suite of analytic methods.

6. Research Utility and Applications

The breadth and standardization of the OHHR make it uniquely suited for:

• Direct, controlled benchmarking of auditory-profiling or clustering algorithms
• Multivariate phenotyping incorporating audiometric thresholds, supra-threshold psychoacoustics, speech-in-noise recognition, cognitive status, and self-report
• Supporting individualized diagnostics and research in precision audiology, including profile-based hearing-aid fitting and stratified medicine

The OHHR was the common analysis resource in (Xu et al., 7 Jan 2026), which enabled assessment and comparison of eight established profiling frameworks using manifold learning and intrinsic metrics. Notably, among audiogram-based frameworks, the Bisgaard profiles showed superior performance, while the Hearing4All auditory profiles—incorporating supra-threshold measures and 13 profile classes—achieved the highest clustering quality as indicated by a low Davies–Bouldin index.

7. Significance in Auditory Profiling and Phenotyping

The extended OHHR is positioned as a gold standard resource for empirical method development and comparison in auditory profile research. Its rigorous acquisition, broad variable coverage, and standardized structure allow for reproducible, scalable application of statistical and machine learning techniques. By harmonizing data at the participant and feature level, the resource underpins fair, quantitative evaluation of clustering validity, dimensionality reduction, and emergent phenotyping methods, supporting future advances in precision audiology (Xu et al., 7 Jan 2026).