A Multidimensional Framework for Operational AI Dataset Readiness: An Evaluation of 120 Video, Image, Audio, and Numerical Datasets

The effective use of datasets in artificial intelligence (AI) depends not only on data availability but also on documentation, legal reusability, technical organisation, annotation characteristics, and preprocessing requirements. This study proposes a multidimensional framework for assessing the operational readiness and reusability of heterogeneous AI datasets and applies it to a portfolio of 120 datasets comprising 30 video, 30 image, 30 audio, and 30 numerical datasets. A harmonised metadata schema and controlled vocabularies were used to derive indicators describing dataset identification, legal and reuse information, documentation, common and modality-specific metadata completeness, annotation characteristics, technical consistency, estimated preprocessing burden, and resource scale. Descriptive analysis was complemented by multiplicity-adjusted non-parametric group comparisons, Dunn–Holm post-hoc tests, categorical association tests, Spearman correlation analysis, and modality-stratified sensitivity analyses of the significant pooled correlations. Common metadata fields were almost completely populated across the portfolio, whereas legal and modality-specific technical information showed substantially greater variation. Under the original scoring specification, Holm-adjusted omnibus comparisons identified significant group differences in legal-documentation completeness, legal-rights clarity, modality-specific metadata completeness, and technical consistency, while documentation completeness remained non-significant. However, sensitivity analyses indicated that the statistical conclusions for technical consistency and estimated preprocessing burden depended on certain alternative scoring specifications. Identification completeness also differed among groups. Annotation granularity and dataset storage size showed pronounced between-group structural differences. Annotation coverage, however, showed no systematic ordinal shift despite a significant difference in the distribution of coverage categories. In the pooled portfolio, higher technical consistency was associated with lower estimated preprocessing burden (Spearman’s ρ=−0.350, Holm-adjusted p=0.001), while greater dataset storage size was associated with greater estimated burden (ρ=0.296, Holm-adjusted p=0.011). Neither association remained statistically significant within individual modality-specific groups after multiplicity correction. These pooled correlations should therefore be interpreted as portfolio-level findings rather than evidence of relationships that necessarily hold across data modalities. Given the purposive sampling and confounding between modality and collecting institution, the observed between-group differences cannot be attributed solely to data modality or generalised beyond the analysed portfolio. The results show that high generic metadata completeness does not by itself establish operational dataset readiness and support the use of separate legal, technical, annotation, preprocessing, and resource dimensions when screening heterogeneous datasets before AI model development.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-19
DOI
https://doi.org/10.3390/app162211477
Primary Topic
Research Data Management Practices
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A Multidimensional Framework for Operational AI Dataset Readiness: An Evaluation of 120 Video, Image, Audio, and Numerical Datasets

Luka Batistić, Primož Podržaj, Janko D. Jovanović, Mirza Pašić et al.
Applied Sciences
Research Data Management Practices
article

A Multidimensional Framework for Operational AI Dataset Readiness: An Evaluation of 120 Video, Image, Audio, and Numerical Datasets

Luka Batistić, Primož Podržaj, Janko D. Jovanović, Mirza Pašić, Mugdim Pašić, Tomaž Požrl, Dražen Drašković, Sandi Ljubić, Jan Pleterski, Jelena Sakovic Jovanovic
article en

Abstract

The effective use of datasets in artificial intelligence (AI) depends not only on data availability but also on documentation, legal reusability, technical organisation, annotation characteristics, and preprocessing requirements. This study proposes a multidimensional framework for assessing the operational readiness and reusability of heterogeneous AI datasets and applies it to a portfolio of 120 datasets comprising 30 video, 30 image, 30 audio, and 30 numerical datasets. A harmonised metadata schema and controlled vocabularies were used to derive indicators describing dataset identification, legal and reuse information, documentation, common and modality-specific metadata completeness, annotation characteristics, technical consistency, estimated preprocessing burden, and resource scale. Descriptive analysis was complemented by multiplicity-adjusted non-parametric group comparisons, Dunn–Holm post-hoc tests, categorical association tests, Spearman correlation analysis, and modality-stratified sensitivity analyses of the significant pooled correlations. Common metadata fields were almost completely populated across the portfolio, whereas legal and modality-specific technical information showed substantially greater variation. Under the original scoring specification, Holm-adjusted omnibus comparisons identified significant group differences in legal-documentation completeness, legal-rights clarity, modality-specific metadata completeness, and technical consistency, while documentation completeness remained non-significant. However, sensitivity analyses indicated that the statistical conclusions for technical consistency and estimated preprocessing burden depended on certain alternative scoring specifications. Identification completeness also differed among groups. Annotation granularity and dataset storage size showed pronounced between-group structural differences. Annotation coverage, however, showed no systematic ordinal shift despite a significant difference in the distribution of coverage categories. In the pooled portfolio, higher technical consistency was associated with lower estimated preprocessing burden (Spearman’s ρ=−0.350, Holm-adjusted p=0.001), while greater dataset storage size was associated with greater estimated burden (ρ=0.296, Holm-adjusted p=0.011). Neither association remained statistically significant within individual modality-specific groups after multiplicity correction. These pooled correlations should therefore be interpreted as portfolio-level findings rather than evidence of relationships that necessarily hold across data modalities. Given the purposive sampling and confounding between modality and collecting institution, the observed between-group differences cannot be attributed solely to data modality or generalised beyond the analysed portfolio. The results show that high generic metadata completeness does not by itself establish operational dataset readiness and support the use of separate legal, technical, annotation, preprocessing, and resource dimensions when screening heterogeneous datasets before AI model development.

Applied SciencesVol. 16(22)
University of Sarajevo (BA), University of Ljubljana (SI), University of Rijeka (HR), University of Belgrade (RS), University of Montenegro (ME)
Research Data Management Practices
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon