CARE contains multimodal clinical features extracted from interviews with participants across multiple medical conditions. Access is provided for research and educational purposes only. Please complete the request form.

By requesting access to CARE, you agree to use the dataset exclusively for research or educational purposes, not to attempt to identify or re-identify any participant, not to redistribute the dataset or derived identifiable information, and to comply with the CARE Data Use Agreement (DUA). Please ensure that your request includes your full name and your institutional email address so that we can review your application.

CARE: A Multimodal Corpus for Studying Speech and Non-Verbal Communication Across Multiple Medical Conditions

CARE is a multimodal English dataset comprising approximately 144 hours of clinical interviews from 622 profiles across 12 medical conditions plus a control cohort.

The dataset contains participant metadata together with pre-computed acoustic, linguistic, and visual features extracted from interview recordings. CARE is designed to support research in speech and language processing, affective computing, multimodal machine learning, and artificial intelligence for healthcare.

Installation

Programmatic access is provided through the accompanying care_dataset Python package, available from the official source repository (The package requires an environment with Python 3.11 or newer.).

git clone https://gitlab.inesc-id.pt/u035731/care_dataset.git
cd care_dataset
pip install .

The machine learning example below additionally requires

pip install scikit-learn

Note: Before accessing CARE, authenticate with your Hugging Face account using hf auth login.

Quick Example

import numpy as np

from sklearn.metrics import roc_auc_score
from sklearn.preprocessing import StandardScaler
from sklearn.svm import SVC

from care_dataset import CAREDataset


# ---------------------------------------------------------
# Load CARE
# ---------------------------------------------------------

care = CAREDataset()

dataset = care.filter(
    label_id__in=["CONTROL", "PARKINSON"],
)
dataset = dataset.select_modalities("egemaps_func") # dataset.info()

# ---------------------------------------------------------
# Aggregate turn-level features into one vector per profile
# ---------------------------------------------------------

def aggregate_profiles(dataset):

    X, y = [], []

    for profile_id in dataset.profile_ids():

        profile = dataset.filter(profile_id=profile_id)

        turns = profile.modality("egemaps_func")
        turns = np.vstack(turns) # (n_samples * n_turns, n_feats)

        X.append( turns.mean(axis=0) ) # (n_feats,)

        y.append( profile.label_ids()[0] == "PARKINSON" )

    return np.vstack(X), np.asarray(y)


# ---------------------------------------------------------
# 5-fold cross validation
# ---------------------------------------------------------

folds, report = dataset.cross_validation(
    k=5,
    keep_duplicate_profiles=False,  # remove duplicated profiles
    report=True,
)

report.info()

aucs = []

for train, test in folds:

    X_train, y_train = aggregate_profiles(train)
    X_test, y_test = aggregate_profiles(test)

    scaler = StandardScaler()

    X_train = scaler.fit_transform(X_train)
    X_test = scaler.transform(X_test)

    clf = SVC(kernel="linear")

    clf.fit(X_train, y_train)

    scores = clf.decision_function(X_test)

    aucs.append(
        roc_auc_score(y_test, scores)
    )

print(aucs) # [0.920, 0.740, 0.914, 0.813, 0.962]
print(
    f"{np.mean(aucs):.3f} ± {np.std(aucs, ddof=1):.3f}"
) # 0.870 ± 0.091

Dataset Organization

Each sample corresponds to one interview video and is uniquely identified as

GROUP__SUBGROUP__PROFILE__VIDEO

For example

LUNGCANCER__LUNGCANCER__PROFILE_446__VIDEO_000

where

GROUP is the clinical condition,
SUBGROUP is a subdivision within the group,
PROFILE uniquely identifies a profile,
VIDEO identifies one interview clip from that profile.

Sample and Profile Metadata

Each interview clip can be accessed either by index

sample = care[0]

or by its unique sample identifier, for example

sample = care["PARKINSON__PARKINSON__PROFILE_522__VIDEO_000"]

Profile-level metadata associated with that interview is then available through

sample.profile()

The complete profile metadata tables can also be obtained with

care.profiles()

care.patients()

care.controls()

Documentation

A complete description of all available modalities, feature dimensions, extraction procedures, and output formats is provided in the accompanying CARE Feature Documentation.

Feature archives

In addition to the Python API, CARE provides compressed archives containing the original extracted features.

The features/ directory contains one archive per clinical condition (e.g., CONTROL.tar.gz, PARKINSON.tar.gz). Each archive includes the extracted multimodal features for all participants belonging to that condition.

These archives are intended for users who wish to work directly with the feature files without using the CARE Python package.

Citation

If you use CARE in your research, please cite

<CARE paper under review in npj Scientific Data>

License

CARE is distributed under the CARE Data Use Agreement (DUA).

By requesting access to or downloading the dataset, users agree to comply with the terms of the DUA.

In particular, users agree to:

use the dataset exclusively for research and educational purposes unless explicitly authorized otherwise;
not attempt to identify, re-identify, or contact any participant;
not combine CARE with external information for participant re-identification;
not redistribute the dataset or any derived identifiable information without prior authorization;
appropriately cite the CARE dataset and the associated publication(s);
comply with all applicable ethical, legal, and institutional requirements governing the use of human participant data.

Users are responsible for ensuring that their use of CARE complies with both the DUA and the policies of their institution.

Contact

For questions, bug reports, or feature requests, please open an issue in the repository or contact the corresponding authors of the CARE dataset publication.

Downloads last month: 99

Total file size:

156 GB