NRC Dataset: Multimodal Stroke Rehabilitation

• NRC dataset is a multimodal rehabilitation benchmark designed for domiciliary upper-limb stroke assessment with clinically scored movement quality.
• It comprises 10 ADL and 5 ROM tasks captured via RGB-D video-derived skeleton sequences and detailed physiotherapist annotations.
• The dataset supports deep learning models using ST-GCN and transformer attention, achieving significant improvements in quantitative score prediction.

Searching arXiv for papers mentioning the “NRC dataset” in rehabilitation and related usages of the acronym. First, I’ll search for the rehabilitation dataset paper. The NRC dataset is a multimodal rehabilitation dataset introduced for domiciliary stroke assessment in the context of an AI-based home rehabilitation system. It includes 10 upper-limb activities of daily living (ADL) and 5 range-of-motion (ROM) tasks collected from stroke and non-disabled participants, with Score annotations provided by licensed physiotherapists. In the reported study, the dataset supports quantitative assessment of movement quality rather than only action identification; although RGB-D video and IMU data were recorded, for model development in this study, only skeleton sequences derived from RGB-D videos were used (Lim et al., 27 Sep 2025).

1. System context and dataset role

The dataset was constructed as part of a home-based rehabilitation exercise and feedback system. That system consists of (1) hardware setup with RGB-D camera and wearable sensors to capture Stroke movements, (2) a mobile application for exercise guidance, and (3) an AI server for assessment and feedback. When a Stroke user exercises following the application guidance, the system records skeleton sequences, which are then assessed by the deep learning model RAST-G@ (Lim et al., 27 Sep 2025).

Within that pipeline, the NRC dataset functions as the principal supervised resource for movement-quality assessment. The model described in the same work employs a spatio-temporal graph convolutional network (ST-GCN) to extract skeletal features and integrates transformer-based temporal attention to figure out action quality. This situates the dataset at the intersection of skeleton-based action modeling, rehabilitation scoring, and domiciliary monitoring. A plausible implication is that the dataset was designed not merely as a generic pose corpus, but as a clinically annotated benchmark for quantitative upper-limb rehabilitation assessment.

2. Acquisition setup and data modalities

The NRC dataset is explicitly described as a multimodal rehabilitation dataset. The acquisition hardware comprised an RGB-D camera: Intel RealSense D435i, providing video sequences and depth maps, and IMU sensors: Movella Xsens Dot attached to both wrists, capturing acceleration and gyroscope data ($x, y, z$) (Lim et al., 27 Sep 2025).

Despite that multimodal acquisition, the benchmarked representation in the study is narrower. The paper states that the NRC dataset used for model development consisted of skeleton trajectories extracted by applying human pose estimation to the video sequences. The extracted skeletons are in the 25-joint format compatible with OpenPose/COCO WholeBody standards. This distinction is important: the dataset is multimodal at collection time, but the reported modeling results are based on skeleton sequences.

Pose extraction was performed with HRNet-DarkPose from MMPose library, trained on COCO-WholeBody. The resulting representation places the NRC dataset within the line of work that uses pose-estimated skeletal trajectories as the primary substrate for downstream graph-based sequence modeling, while preserving a link to the original RGB-D acquisition protocol.

3. Participants, activities, and partitioning

The corpus contains 1,142 total action-score pairs. It includes 6 non-disabled (ND) subjects and 11 stroke participants. The stroke subgroup focused on subacute patients (within six months of onset), age $\ge 19$, upper-limb hemiplegia (FMA $\ge 30$, Brunnstrom's stage 4–4.5), MOCA $\ge 22$, able to understand and communicate regarding the experiment (Lim et al., 27 Sep 2025).

The participant-wise split reported in the paper is as follows:

Group	Train / Validation / Test	Total
Non-disabled (ND)	293 / 32 / –	325
Stroke	633 / 70 / 114	817
Total	916 / 112 / 114	1,142

The dataset contains 15 well-annotated movement classes, divided into 10 Activities of Daily Living (ADL) and 5 Range of Motion (ROM) tasks, targeting upper-limb function. The ADL classes are LiftCupHandle, HairBrush, BrushTeeth, Remotecon, MovingCan, Writing, FoldingPaper, FoldupTowel, WashFace, and Smartphone. The ROM classes are RightShoulderFrontal, LeftShoulderFrontal, RightShoulderSide, LeftShoulderSide, and LateralRotation.

The activity taxonomy is further organized by hand-use structure. UNI denotes Unimodal hand use (single hand); BIA denotes Bimodal and asymmetric hand use (both hands, different actions); and BIS denotes Bimodal and symmetric hand use (both hands, same action). Every movement was confirmed in consultation with physiotherapy faculty. This organization is clinically relevant because it distinguishes one-handed, asymmetric bimanual, and symmetric bimanual motor demands within the same evaluation framework.

4. Annotation protocol and scoring design

Evaluation and annotation were conducted by three licensed physiotherapists (5+ years clinical experience), with validation by two Ph.D. physical therapy faculty. Each trial (an action instance) received an assessment score using a custom set of 10 evaluation items on a 0–5 Likert scale per item, yielding a total max score: 50 per action instance (Lim et al., 27 Sep 2025).

The paper gives examples of evaluation items, including: Is the main objective of the action achieved?; Is the action performed smoothly and continuously?; Are the force, speed, and direction appropriately controlled?; and Is the task performed without dropping/missing the tool? These examples indicate that the annotations target execution quality, task completion, and control characteristics.

The stated rationale is functional rather than impairment-centric: This approach emphasizes function (ability to perform in context) over impairment-based assessment. That design choice aligns the dataset with rehabilitation settings in which performance in ADL-like scenarios is clinically salient. A plausible implication is that the annotation protocol aims to capture ecologically valid movement quality rather than isolated kinematic impairment markers alone.

5. Preprocessing, normalization, and benchmark usage

Before modeling, the skeleton sequences underwent temporal normalization. All skeleton sequences standardized to 288 frames using a frame dropping and uniform sampling technique, maintaining action dynamics while handling variability in action length/duration. The data were segmented by exercise instance and split by participant to avoid data leakage. The reported split is approximately 8:1:1, and the test set only includes stroke data, reflecting the stated real application need for evaluating stroke rehabilitation progress (Lim et al., 27 Sep 2025).

The benchmark task in the paper is quantitative score prediction. The reported evaluation metrics are MAD (Mean Absolute Deviation), RMSE (Root Mean Square Error), and MAPE (Mean Absolute Percentage Error):

$$\text{MAD} = \frac{1}{n} \sum_{i=1}^{n} \left| y_i - \hat{y}_i \right|$$

$$\text{RMSE} = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} \left( y_i - \hat{y}_i \right)^2 }$$

$$\text{MAPE} = \frac{1}{n} \sum_{i=1}^{n} \left| \frac{y_i - \hat{y}_i}{y_i} \right| \times 100$$

The baselines were Kuang et al. (hierarchical contrastive, orientation-feature GCN) and Deb et al. (spatio-temporal GCN for rehabilitation). On the NRC dataset, the reported results were: RMSE 0.291 for RAST-G@, compared with 1.961 for Kuang et al. and 1.030 for Deb et al.; MAPE 0.259 for RAST-G@, compared with 2.836 and 3.805; and MAD 0.321 for RAST-G@, compared with 0.739 and 0.892. The paper states that RAST-G@ outperforms baselines by a large margin, and further reports ablation findings confirming the importance of frame normalization, temporal attention, and using both ND and stroke data for learning movement dynamics.

The dataset was also used in a longitudinal setting: Data was collected for up to 4 months with periodic actions; the model tracks user trends over time, supporting longitudinal analysis. This makes the benchmark relevant not only for single-trial assessment but also for recovery monitoring.

6. Significance, scope, and acronym ambiguity

The paper characterizes NRC as the first open-sourced, rigorously annotated dataset containing both daily-life and clinical ROM activities collected from both stroke and healthy participants in a domiciliary (home) context (Lim et al., 27 Sep 2025). It is presented as supporting home-based, patient-centered rehabilitation monitoring, AI-driven feedback systems for patients and clinicians, and long-term tracking of patient recovery trends. The accompanying repository is given as https://github.com/LimSuH/NRC-rehab.

A recurring source of confusion is the acronym itself. In arXiv literature, NRC is also used for distinct concepts and resources. It denotes non-negative representation based classification in pattern recognition (Yin et al., 2019); the Nested Relational Calculus, a high-level query language and basis for language-integrated query (Cheney et al., 2021); the NRC VAD Lexicon v2, which has human ratings of valence, arousal, and dominance for more than 55,000 English words and phrases (Mohammad, 30 Mar 2025); the NRC Affect Intensity Lexicon, a manually created affect intensity lexicon with close to 6,000 English words (Mohammad, 2017); and Neural Regression Collapse, a layer-wise structure studied in deep regression models (Rangamani et al., 25 Mar 2026). In the rehabilitation literature, however, the phrase NRC dataset refers specifically to the domiciliary upper-limb stroke rehabilitation dataset introduced with the RAST-G@ system.

The dataset’s reported limitations are largely implicit in its formulation. The benchmarked representation is skeleton-only despite multimodal acquisition, and the test set contains stroke data only. This suggests that current reported use emphasizes clinically relevant score prediction under a specific deployment scenario rather than exhaustive multimodal fusion. Within that scope, the NRC dataset occupies a specific niche: a supervised, clinically scored, home-context benchmark for upper-limb rehabilitation assessment based on skeleton sequences.