Datasetpublic

xyz1901901/patientcare

Patient Care Activity Recognition Dataset Dataset Summary A multi-class video action recognition dataset covering 23 clinically relevant patient care activities recorded in hospital-like environments. The dataset includes both real-life recordings and synthetically generated videos. The dataset is intended to support research in automated patient monitoring and clinical activity recognition using computer vision. Supported Tasks Video… See the full description on the dataset page: https://huggingface.co/datasets/xyz1901901/patientcare.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes365downloads

Dataset Card

Patient Care Activity Recognition Dataset

Dataset Description

—Homepage: https://huggingface.co/datasets/Avinash190/patientcare
—Repository: https://huggingface.co/datasets/Avinash190/patientcare
—Task: Video-based action recognition for patient care monitoring in hospital settings
—Languages: English (label annotations only)

Dataset Summary

A multi-class video action recognition dataset covering 23 clinically relevant patient care activities recorded in hospital-like environments. The dataset includes both real-life recordings and synthetically generated videos.

The dataset is intended to support research in automated patient monitoring and clinical activity recognition using computer vision.

Supported Tasks

—Video Classification / Action Recognition: Classify a video clip into one of 23 patient care activity classes. ---

Data Instances

Each instance consists of:

—An .mp4 video clip of a single patient care activity.
—An integer class label (0–22).

Data Splits

Config	Split	Videos	Description
real	train	180	Real-life training clips
real	validation	200	Real-life validation set
real	test	997	Main real-world evaluation set (~1000 clips across all classes)
realtest200	test	200	Real-life validation set (standalone config)
synthetic	train	9,267	Synthetic training clips (clean + noisy)
synthetic	test	1,064	Synthetic test clips

The synthetic config contains two sub-variants: clean/and noisy/

Class Labels

ID	Class Name	Description
01	patientlyingon_bed	Patient resting flat or on their side on a hospital bed.
02	standingupfrom_bed	Patient transitioning from lying or sitting to a standing position.
03	fallfromstanding	Patient losing balance and falling from a standing position.
04	patientsittingupfrombed	Patient moving from a flat lying position to sitting upright in bed.
05	persistent_coughing	Patient exhibiting repeated or prolonged coughing episodes.
06	eatingfoodwith_tray	Patient eating food from a hospital meal tray.
07	taking_medicine	Patient taking oral medication, typically with water.
08	walkingwithwalker	Patient ambulating using a walker or other mobility aid.
09	removingoxygenequipment	Patient removing oxygen mask or nasal cannula from their face.
10	chokingonfood	Patient showing signs of airway obstruction or choking while eating.
11	patientinvoluntarymovements	Patient exhibiting uncontrolled, seizure-like full-body movements.
12	tamperingivlines_catheter	Patient touching, pulling, or tampering with IV lines or catheter.
13	walkingwithnormal_gait	Patient walking independently with a steady, normal gait.
14	wheelchair_seated	Patient seated stationary or being moved in a wheelchair.
15	nausea_vomiting	Patient showing signs of nausea, retching, or vomiting.
16	walkingabnormalgait	Patient walking with an irregular, limping, or unsteady gait.
17	patientchestdistress	Patient clutching chest or showing signs of chest pain or distress.
18	scratching_body	Patient scratching skin on arms, legs, or other body areas.
19	usingphonetablet	Patient interacting with a smartphone or tablet device.
20	bedsideselfdirected_exercise	Patient performing self-initiated physical exercises beside the bed.
21	headache_distress	Patient holding or clutching head, indicating headache or head pain.
22	gesturinghandwaving	Patient making deliberate hand gestures or waving motions.
23	coveringfacewith_hands	Patient covering face with one or both hands.

Dataset Creation

Source Data

Initial Data Collection and Normalization

Synthetic videos (train + test) are the original contribution of this work and were generated using a Wan2.1 video diffusion model.

Real-life videos used for evaluation were curated and re-labelled from five publicly available datasets:

Source Dataset	Description
Toyota Elderhomes	Daily activity recognition in eldercare smart-home environments
NTU RGB+D 120	Large-scale RGB+D action recognition dataset including medical/daily activities
MedVideoCap-55k	Medical procedure video captioning dataset with clinical activity clips
UR Fall Detection Dataset	Benchmark dataset for RGB and depth-based fall detection
WU-SAHZU-EMU Dataset	Clinical patient activity dataset from hospital ward monitoring

Clips were trimmed, re-annotated according to the 23-class taxonomy defined in this dataset, and split into train/test sets.

Who are the source language producers?

Video content is non-linguistic. Original source datasets were produced by their respective authors. Re-annotation for this dataset was performed by domain experts in clinical care and computer vision.

Annotations

Annotation Process

Synthetic videos are labelled by construction. Real-life clips sourced from external datasets were re-annotated by expert annotators according to the 23-class taxonomy.

Who are the annotators?

Expert annotators with knowledge of clinical patient care activities.

Personal and Sensitive Information

No real patient data or personally identifiable information (PII) is included. Real-life video clips are sourced from publicly released research datasets whose subjects consented to research use. Synthetic videos contain no real individuals.

Considerations for Using the Data

Social Impact

This dataset aims to accelerate research in automated patient monitoring, which can improve hospital safety (e.g., fall detection, tamper/IV-line alerts) and reduce caregiver burden.

Discussion of Biases

—Synthetic data may not fully capture the appearance diversity of real patients.
—Real-life evaluation clips are drawn from heterogeneous source datasets with varying recording conditions, camera angles, and subject demographics.
—Some high-risk classes (e.g., falls, choking) are underrepresented relative to routine activities across all source datasets.
—Source datasets have their own demographic biases (e.g., NTU RGB+D skews younger; Toyota Elderhomes skews older).

Additional Information

Licensing Information

The synthetic subset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

The real-life subset is derived from the following source datasets, each with their own license. Users must comply with the terms of each source:

Dataset	License
Toyota Elderhomes	Research/Non-commercial use
NTU RGB+D 120	Research use only
MedVideoCap-55k	See repository
UR Fall Detection Dataset	Public research use
WU-SAHZU-EMU Dataset	See repository

bibtex

@inproceedings{Das2019ToyotaSmarthome,
  title     = {Toyota Smarthome: Real-World Activities of Daily Living},
  author    = {Das, Srijan and others},
  booktitle = {ICCV},
  year      = {2019}
}

@article{Liu2020NTURGBD120,
  title   = {NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding},
  author  = {Liu, Jun and others},
  journal = {IEEE TPAMI},
  year    = {2020}
}

@article{Kepski2014URFall,
  title   = {Human Fall Detection on Embedded Platform Using Depth Maps and Wireless Accelerometer},
  author  = {Kepski, Michal and Kwolek, Bogdan},
  journal = {Computer Methods and Programs in Biomedicine},
  year    = {2014}
}

Contributions

Contributions and issue reports are welcome via the Hugging Face dataset repository.