Models that read a sequence of video frames classified simulated neonatal intensive care interventions far more accurately than a comparator given only one frame in a new benchmark. TSF achieved 93.97% top-1 accuracy on the reported test set, compared with 23.17% for the framewise system; the authors reported a 70.80% performance gap. The result comes from 4,144 videos across 12 simulated intervention classes built around a manikin. It is therefore a benchmark result under the study's recording design, not evidence that the same performance has been shown in an actual NICU.
A benchmark built around a manikin
ICVD, the name given to the dataset, is introduced for automated documentation development. Its privacy-preserving design uses a manikin and shows only the clinician's hands; the paper presents this as enabling distribution without privacy constraints. The choice also sets the boundary of the evidence, because the benchmark is built around a simulated care scene.
The recording plan added defined variation. It specified two background conditions, five hand-covering types, two performers and eight camera perspectives. That gives the benchmark a stated range of visual setups, while the evidence remains tied to the recorded choices rather than to real NICU footage.
Post-production produced 4,144 clips from an initial target of 4,160. Quality control accepted 99.6% of the material and rejected 16 clips, or 0.4%. These figures describe how the benchmark was assembled, not how accurately the models classified the interventions.
The paper reports 4,144 total videos and lists 2,892 for training, 624 for validation and 631 for testing. The headline metrics were calculated on the reported test partition, making the result specific to the split described by the study rather than to a separate set of real-world footage.
The comparison was designed to test time
Two transformer approaches - AI models that compare patterns across an input - supplied the temporal comparison. TSF and MF were pretrained on Kinetics-400 and then fine-tuned on the training set. TSF used divided space-time attention, while MF used trajectory attention, giving the two systems different ways to organize information across a video.
Input preparation also differed slightly. TSF used 20 frames per clip and MF used 16 frames per clip, with a target frame rate of 30 FPS. For evaluation, frames were resized to 224 by 224 pixels. The comparison was therefore between particular model-and-input setups, not an abstract test of every possible video system.
The evaluation used class-balanced macro-averaged top-1 and top-5 accuracy, precision, recall and F1. Top-1 asks whether the correct label is the model's first choice; top-5 counts it if it appears among five choices. Macro-averaging means the classes contribute equally to the summary, while F1 combines precision and recall into one score. The ablation compared full TSF with a framewise model using space-only attention and a single-frame input.
The gap was large, but not uniform
TSF's test-set numbers were high across the reported measures: 93.97% top-1 accuracy and 100.00% top-5 accuracy, with precision of 94.11%, recall of 93.29% and F1 of 92.92%. The authors did not report a confidence interval or other uncertainty estimate for these results.
MF produced a very similar pattern: 93.17% top-1 accuracy, 99.84% top-5 accuracy, precision of 93.73%, recall of 92.32% and F1 of 91.29%. The two temporal architectures therefore delivered similar reported top-1 results in this benchmark. No confidence intervals or other uncertainty estimates were reported for either temporal model.
The framewise comparator told a different story. Its top-1 accuracy was 23.17%, with 65.87% top-5 accuracy; precision was 23.08%, recall 23.00% and F1 19.45%. Because it received a single-frame input and used space-only attention, FW did not have the same temporal input setup as TSF or MF. The comparison shows a large measured difference, but it cannot isolate every reason for that difference.
The paper summarizes that contrast as a 70.80% performance gap between the framewise and temporal approaches. The result shows that the two input setups performed very differently on this classification task under the recorded conditions. The paper does not report uncertainty around the gap, and it does not establish that the same difference would appear with other footage or in an actual NICU.
Looking beyond the overall average, TSF was reported at 100% accuracy in six classes. The listed non-perfect results included wiping-body at 98.91%, blanket-cover and pacifier-take at 97.96% each, and feeding at 97.92%. These figures help explain the high overall test score while showing that the average can conceal differences between intervention labels.
The weaker part of the analysis involved composite diaper-related actions. Diaper-apply reached 85.42% accuracy, while diaper-change was reported at 41.30%; the analysis recorded six confusions involving diaper-change and one involving diaper-remove. MF's diaper-change accuracy was lower still at 28.26%. The pattern points to a harder boundary between closely related action labels than the near-perfect scores in several other classes.
A strong benchmark with a narrow reach
Taken together, the scores describe a benchmark result with a narrow scope. ICVD is built from manikin videos, and the authors identify a representativeness gap between those recordings and real NICU settings. They call for validation on ethically approved footage from actual NICU care, which is the step needed to determine whether the reported performance travels beyond the simulated scene.
The class results underline why that question matters. TSF was near-perfect on many reported labels but much weaker on composite diaper actions; the benchmark also used a specified set of backgrounds, hand coverings, performers and camera perspectives. A new setting could test whether closely related actions remain difficult and whether the model handles visual conditions beyond those recorded.
Within its own boundaries, the dataset provides a concrete benchmark for automated documentation development. It specifies the intervention classes, recording variation, quality-control process and model comparison, then reports scores on the test partition described in the paper. Its central contribution is a defined question about whether video models can distinguish the recorded interventions; the representativeness question remains open until actual NICU footage is tested.
The supplied front matter identifies the work as arXiv:2608.23838v1, dated 24 Aug 2026, and labels it a preprint. Its reported accuracy is therefore best understood as a result from the submitted benchmark, while validation in real NICU settings remains open.
Paper data and sources
Original title: Infant Care Video Dataset for Classification of Interventions Using Transformers
Authors: Igor Bogdanov, James Green
Journal/Repository: 2025 IEEE 49th Annual Computers, Software, and Applications Conference (COMPSAC), pp. 2130-2135, 2025
Status: Preprint, not yet peer-reviewed
First online: 2026-08-24
DOI: 10.1109/compsac65507.2025.00299
Original paper · Full text