Hand Gesture Recognition Using MediaPipe Landmarks and

Deep Learning Networks

Manuel Gil-Martín

, Marco Raoul Marini

, Iván Martín-Fernández

, Sergio Esteban-Romero

and

Luigi Cinque

Grupo de Tecnología del Habla y Aprendizaje Automático (THAU Group), Information Processing and

Telecommunications Center, E.T.S.I. de Telecomunicación, Universidad Politécnica de Madrid (UPM),

Av. Complutense 30, 28040, Madrid, Spain

VisionLab, Department of Computer Science, Sapienza University of Rome, Via Salaria 113, Rome 00198, Italy

Keywords: Hand Gesture Recognition, Human-Computer Interaction, MediaPipe Landmarks, Deep Learning.

Abstract: Advanced Human Computer Interaction techniques are commonly used in multiple application areas, from

entertainment to rehabilitation. In this context, this paper proposes a framework to recognize hand gestures

using a limited number of landmarks from the video images. This hand gesture recognition system

comprises an image processing module that extracts and processes the coordinates of 21 hand points called

landmarks, and a deep neural network module that models and classifies the hand gestures. These landmarks

are extracted automatically through MediaPipe software. The experiments were carried out over the IPN

Hand dataset in an independent-user scenario using a Subject-Wise Cross Validation. They cover the use of

different landmark-based formats, normalizations, lengths of the gesture representations, and number of

landmarks used as inputs. The system obtains significantly better accuracy when using the raw coordinates

of the 21 landmarks through 125 timesteps and a light Recurrent Neural Network architecture (80.56 ±

1.19 %) or the hand anthropometric measures (82.20 ± 1.15 %) compared to using the speed of the hand

landmarks through the gesture (72.93 ± 1.34 %). The proposed framework studied the effect of different

landmark-based normalizations over the raw coordinates, obtaining an accuracy of 83.67 ± 1.12 % when

using as reference the wrist landmark from each frame, and an accuracy of 84.66 ± 1.09 % when using as

reference the wrist landmark from the first video frame of the current gesture. In addition, the proposed

solution provided high recognition performance even when only using the coordinates from 6 (82.15 ±

1.16 %) or 4 (81.46 ± 1.17 %) specific hand landmarks using as reference the wrist landmark from the first

video frame of the current gesture.

1 INTRODUCTION

Hand gesture recognition consists in detecting the

movement that people perform using their hands.

This technology could be useful to develop human

computer interaction systems and could improve the

user experience across a wide variety of domains.

For example, it could be seen as the basis for sign

language understanding and hand gesture control

applications. For instance, a person could ask to take

a picture using the front camera of a smartphone by

opening and closing the hand palm. In these

applications, it is crucial to accurately recognize the

hand gesture to perform specific actions with smart

devices, a computer, or an automatic transmission

machine.

Multiple previous works have been focused on

human activity recognition to optimize the physical

activity classification using wearables or cameras

(Gil-Martin, San-Segundo, Fernandez-Martinez, &

Ferreiros-Lopez, 2020, 2021; Gil-Martín, San-

Segundo, Fernández-Martínez, & de Córdoba, 2020;

Zhang et al., 2017). However, there exists a lower

number of works focused on detecting hand poses or

gestures. Most of these works used images as inputs

of their systems and follow a hand localization step

as the first stage. Afterwards, they extracted

handcrafted features or descriptors (Trindade, Lobo,

& Barreto, 2012) from the hand and fed them to an

inference algorithm that classifies the different hand

poses or gestures. For example, a previous work

(Mantecon, del-Blanco, Jaureguizar, & Garcia, 2019)

Gil-Martín, M., Marini, M. R., Martín-Fernández, I., Esteban-Romero, S. and Cinque, L.

Hand Gesture Recognition Using MediaPipe Landmarks and Deep Learning Networks.

DOI: 10.5220/0013053500003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 24-30

ISBN: 978-989-758-737-5; ISSN: 2184-433X

segmented the image into the hand in different

regions and obtained the Histogram of Oriented

Gradients (HOG) and a Local Binary Pattern (LBP)

from each region. Afterwards, they combined k-

means and Support Vector Machines (SVM) to

classify the hand poses, obtaining an F1-score near

96% using data from 25 subjects and 16 different

hand poses. Similarly a previous work (Bao,

Maqueda, del-Blanco, & Garcia, 2017) fed a deep

Convolutional Neural Network (CNN) to directly

classify hand poses in images without any previous

segmentation. They classified the hand pose with

average accuracy of 97.1% in the images with simple

backgrounds and 85.3% in the images with complex

backgrounds. They used a dataset with data from 40

subjects and seven different hand poses. Another

work (Gil-Martín, San-Segundo, & de Córdoba,

2023) used a normalization over the hand landmarks

to detect the same poses, achieving robust

performance even when the images had complex

backgrounds. This approach has the advantage of

sending less information to the recognition module

compared to traditional computer vision approaches

(landmark coordinates vs. a full image). Another

previous work (Benitez-Garcia, Olivares-Mercado,

Sanchez-Perez, Yanai, & Ieee Comp, 2021) used the

raw images to classify hand gestures via training a

ResNeXt-101 model, achieving a 86.32% of

accuracy when evaluating 13 subjects.

This paper is focused on exploring the impact of

different landmark-based input formats over the

gesture recognition task, rather than optimizing the

deep learning architecture for obtaining the

maximum performance. This work uses a state-of-

the-art deep learning architecture to understand

which input formats—specifically raw hand

landmark coordinates, speed of coordinate

movement, and anthropometric measures—yield the

most informative representations for gesture

recognition. The primary contributions of this

research are as follows:

 Analyze different landmark input formats: raw

coordinates, speed of coordinates movement,

and anthropometric measures.

 Investigate the effects of various

normalization techniques on raw landmark

coordinates.

 Minimize the number of landmarks used in the

recognition process while keeping a high

recognition performance.

2 MATERIAL AND METHODS

This section describes the dataset used, the

landmarks information extraction, and the proposed

model architecture.

2.1 Dataset Description

We have used the public dataset called IPN Hand to

evaluate our system.

The IPN Hand dataset (Benitez-Garcia et al.,

2021) includes hand gestures performed by 50

subjects for interaction with touchless screens. It

contains 4,218 gesture instances and 800,000

frames. It includes 13 gestures performed with one

hand: pointing with one or two fingers, clicking with

one or two fingers, throwing up/down/left/right,

opening twice, double-clicking with one or two

fingers, zooming in, and zooming out. During data

collection, each subject did different gestures with

three random breaks in a single video. The subjects

used their own PC or laptop to collect the RGB

videos, which were recorded in the resolution of

640x480 with the frame rate of 30 fps.

This dataset is a great choice for hand gesture

recognition due to its extensive and varied content

regarding instances, gestures and subjects, ensuring

a diverse representation of hand motions. This

diversity enhances the dataset's applicability for real-

world human-machine interaction applications, such

as touchless screens and virtual reality interfaces,

where accurate gesture recognition is crucial. Its

realistic data collection, with subjects using their

own devices to record gestures in varied

environments, ensures that models trained on the

dataset can generalize well to practical usage

scenarios.

2.2 Landmarks-Based Representations

The original images from the dataset were processed

by the MediaPipe library to extract the x and y

coordinates from specific points of the hand called

landmarks. MediaPipe (Lugaresi et al., 2019;

Quinonez, Lizarraga, & Aguayo, 2022) is a powerful

library with the capacity to track pose and hands

from input frames or video streams. This framework

can extract 21 landmarks from the hand (including

wrist and four points along the five fingers). To

standardize the input data and ensure consistent

processing, we applied zero padding at the

beginning of each gesture sequence, thereby aligning

all examples to the same length, ranging from 25 to

250 timesteps. This padding ensures that a uniform

Hand Gesture Recognition Using MediaPipe Landmarks and Deep Learning Networks

length, although it becomes dominant in shorter

gestures.

In our gesture recognition pipeline, we leverage

three distinct types of landmark information formats

to encapsulate various aspects of hand movements

and spatial relationships. Firstly, we utilize the raw

coordinates of landmarks extracted from the hand,

providing direct spatial information about the hand's

configuration in each frame of the gesture sequence.

Secondly, we incorporate the speed derived from the

computed derivatives of these coordinates, capturing

the temporal dynamics and velocity of hand

movements throughout the gesture. Lastly, we

integrate anthropometric measures (Pheasant &

Haslegrave, 2006) obtained by computing the

Euclidean distances between pairs of landmarks,

enabling us to encode spatial relationships in the

hand gestures and structural characteristics inherent

in the hand physiognomy.

For the first approach of landmark information

(raw landmark coordinates), we applied two

normalization techniques at the landmark level to

consider hand translation. The first normalization

method involved using the wrist landmark from each

frame of the gesture sequence as a reference point.

This process entailed subtracting the coordinates of

the wrist landmark from those of all other landmarks

within the same frame, aligning them relative to the

position of the wrist. The second normalization

technique employed the wrist landmark from the

first frame of the gesture sequence as a constant

reference point throughout. Regardless of

subsequent frames, the coordinates of all landmarks

were adjusted relative to the wrist landmark from the

initial frame, ensuring uniformity and consistency in

spatial representations across the entire gesture

sequence. The formulas related to these

normalizations are described in Equation (1) and

Equation (2), respectively, where 0 landmark

corresponds to the wrist, i ranges from 1 to 20 (for

the other landmarks) and t refers to a specific frame

in the gesture sequence. These normalization

approaches could contribute to reduce variability

and enhance the interpretability of landmark

coordinates for the modeling and recognition

architecture. We also evaluated the possibility of

performing a scaling normalization, but no

improvement was obtained. A possible reason for

this is that the distance between the subjects and

their laptops was relatively consistent across the

dataset.

𝑥



,

=𝑥

,

−𝑥

,

𝑦



,

=𝑦

,

−𝑦

,

(1)

𝑥



,

=𝑥

,

−𝑥

,

𝑦′





=𝑦





−𝑦



(2)

2.3 Model Architecture

The deep learning architecture used in this work

learns the evolution pattern from landmark-based

inputs using a Long Short-Term Memory (LSTM)

layer of 100 neurons and classifies the examples

using a Dense layer of 13 neurons, corresponding to

the number of classes. The input shape of the

architecture was bidimensional, including the

timesteps of the gesture (from 25 to 250) and the

number of channels (for example, 42 when using x

and y coordinates from the 21 landmarks). The

architecture included a dropout layer (0.3) after the

LSTM layer to avoid overfitting during training. The

last layer used a softmax activation function to offer

the predictions of each class for every analysis

gesture. We used categorical cross-entropy as loss

metric and the root-mean-square propagation

method as optimizer (Weiss, 2017). We adjusted the

epochs and batch size of the deep learning structure:

50 and 25, respectively. Since the objective of this

work is not optimizing the deep learning

architecture, its configuration is a sample of a state-

of-the-art RNN useful for modeling pattern sequence

(Goodfellow, 2016), which is the case of gesture

recognition task.

3 RESULTS AND DISCUSSION

To evaluate the system using the whole dataset in a

subject-independent scenario, we followed a

Subject-Wise Cross-Validation (SW-CV) alternative

as data distribution. In this 10-fold CV methodology,

the given data are divided into 10 groups or folds to

train and test a system with different data subsets in

such a way that the data from one subject is only

contained in one subset. This process is repeated by

changing the training and testing folds and the

results are the average of the partial results obtained

for all repetitions. This methodology simulates a

realistic scenario where the system is evaluated with

recordings from subjects different to those used for

training.

As evaluation metrics, we used accuracy, which

is defined as the ratio between the number of

correctly classified samples and the number of total

samples. This way, for a classification problem with

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

N testing examples and C classes, accuracy is

defined in Equation (3).

Accuracy=

P







(3)

To show statistical significance values, we used

confidence intervals, which include plausible values

for a specific metric. We will assure that there exists

a significant difference between results of two

experiments when their confidence intervals do not

overlap. Equation (4) represents the computation of

confidence intervals attached to a specific metric

value and N samples for 95% confidence level.

(

95%

)

=±1.96



metric · (100 − metric )

(4)

Regarding the experiments, we firstly analyzed

the effect of different landmark-based input formats:

raw landmark coordinates, landmarks speed and

anthropometric measures. In these experiments we

used the 21 available landmarks and different

lengths for each gesture (25, 50, 75, 100, 125, 150,

175, 200, 225 and 250 timesteps). Figure 1 shows

the evolution of accuracy considering the length of

the gestures and the different landmark-based input

formats.

This figure shows how a saturation of

performance is achieved when increasing the length

of the gestures, reaching a competitive performance

for each landmark representation at 125 timesteps,

which is close to the mean of duration of the

gestures in the dataset (140 timesteps). For example,

at this gesture length, the system obtains an accuracy

of 80.56 ± 1.19 % when using the raw coordinates,

72.93 ± 1.34 % when using their speed and 82.20 ±

1.15 % when using the hand anthropometric

measures. As observed in the figure, when it comes

to modeling gestures, finding the right balance in the

amount of real data is crucial. Using too few points

fails to capture the nuances of gestures adequately,

potentially cutting off crucial information or having

too much padding. Conversely, padding sequences

unnecessarily lengthens them, posing challenges for

recurrent layers to process efficiently.

Results also suggest that the raw landmark

coordinates and the anthropometric measures offer

significantly better performance compared to the

landmark speed along the different timestep

configurations.

Raw coordinates serve as fundamental building

blocks for hand gesture recognition, offering direct

spatial information about the positions of hand

landmarks. Anthropometric measures derived from

hand landmarks provide valuable insights into the

Figure 1: Accuracy depending on the number of timesteps

used per gesture and the landmark format.

structural characteristics and spatial relationships

within gestures. This way, relative positions of hand

key points could be useful to learn features based on

the spatial arrangement of landmarks, facilitating the

discrimination of complex gestures with subtle

variations. However, landmark speed might not

provide enough contextual information for

understanding the hand gestures, because features

such as direction, acceleration, and spatial relations

between landmarks could be lost in this

representation, which may struggle to accurately

distinguish the hand gestures.

Second, we performed some experiments

applying different landmark-based normalization

over the raw coordinates. Figure 2 shows the

evolution of accuracy considering the length of the

gestures and the landmark normalization used.

This figure shows a tendency of performance

improvement when applying a landmark

normalization over the raw coordinates, which

becomes significant when using 125 or 150

timesteps. For example, at 125 timesteps length, the

accuracy of 80.56 ± 1.19 % obtained with the raw

landmark coordinates is significantly improved until

83.67 ± 1.12 % when normalizing through the wrist

landmark from the current frame, and until 84.66 ±

1.09 % when using as reference the wrist landmark

25 50 75 100 125 150 175 200 225 250

Accuracy (%)

Timesteps

Raw Coordinates Landmarks

Landmarks Speed

Anthropometric Measures

Hand Gesture Recognition Using MediaPipe Landmarks and Deep Learning Networks

from the first video frame of each gesture.

Normalizing the coordinates using the wrist

landmark as reference removes the variability in the

inputs among users. One of the reasons for this

improvement is that thanks to these types of

normalizations, the representation of examples of the

same gesture become similar independently of the

location of the hand through the images. For

example, the representation of a hand gesture

consisting in clicking with one finger could differ

when the person performs the gesture at the right or

left side of the image. However, thanks to

normalizing using the wrist landmark, both

representations become standardized since both use

the coordinate origin as reference. Table 1 shows a

summary of the results from the previous

experiments using 125-timestep gesture length.

Figure 2: Accuracy depending on the number of timesteps

used per gesture and the landmark normalization used over

the raw landmark coordinates.

Comparing the state-of-the-art works, it is fair to

highlight that the hand pose recognition usually gets

higher performance because it deals with fixed

images, which are more stable and easier to

recognize. In contrast, the hand gesture recognition

involves sequences of images of varying lengths,

requiring normalization and handling of temporal

dynamics, which complicates the recognition task.

Table 1: Accuracy for different landmark-based input

formats using 125-timestep gesture length.

Landmar

ased input Accuracy (%)

Raw Coordinates + No

Normalization

80.56 ± 1.19

Raw Coordinates + Wrist of Current

Frame Norm.

83.67 ± 1.12

Raw Coordinates + Wrist of First

Frame Norm.

84.66 ± 1.09

Landmarks Spee

72.93 ± 1.34

Anthropometric Measures 82.20 ± 1.15

For example, a previous work (Benitez-Garcia et

al., 2021) used 37 subjects from this dataset to train

a ResNeXt-101 model with 47.56 million parameters

and evaluating the remaining 13 subjects, achieving

an 86.32% accuracy. This potentially allows for

overfitting, where the model performs well on the

test data because it has been optimized for it. Using

the same setup, the proposed approach of using the

raw coordinates and the wrist of first frame

normalization obtained an 86.38 ± 2.03 % accuracy

using a significantly lighter model with only 58,614

parameters. In addition, this work offers a

competitive 84.66 ± 1.09 % accuracy evaluating a

larger and more diverse set of 50 subjects, making

the task more challenging and the model's

generalization performance more critical.

Finally, we analyzed if using a limited number of

landmarks was informative enough to provide a high

performance. Analyzing the nature of the dataset

gestures, we observed that all the 13 classes

compromised the movement of the thumb, index and

middle fingers, and the ring and little fingers were

slightly used in the gestures. This supports the idea

that we observed in a previous study (Luna-Jimenez

et al., 2023), indicating that for example the variance

of the index finger was higher (𝜎2 =0.031) than the

one of the little finger (𝜎2 =0.019). This way, we

used the wrist and all fingertips landmarks (6

landmarks) and the wrist, and fingertips from thumb,

index and middle fingers (4 landmarks) to perform

the experiments. In this case, the anthropometric

measures were computed using only these

landmarks (Euclidean distances between those pairs

of landmarks). Figure 3 shows the accuracy

depending on the landmark-based input format and

the number of landmarks per hand.

We observed that using the wrist and the

fingertips landmarks could be enough to obtain a

high recognition performance when using raw

landmark coordinates as inputs. For example, when

using the raw coordinates and the normalization

using the first video frame, the system achieved

82.15 ± 1.16 % and 81.46 ± 1.17 % when using 6

25 50 75 100 125 150 175 200 225 250

Accuracy (%)

Timesteps

No Landmark Normalization

Wrist Landmark of First Frame

Wrist Landmark of Current Frame

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

and 4 landmarks, respectively. However, there exists

a performance decrease when using the

anthropometric measures and reducing the number

of landmarks. This could be due to a loss of spatial

relationships between the different landmarks (from

210 when using 21 landmarks to 15 and 6 when

using 6 and 4 landmarks, respectively).

Figure 3: Accuracy depending on the landmark-based

input format and the number of landmarks used per

gesture.

4 CONCLUSIONS

This paper proposes a system to detect hand gestures

using a limited number of landmarks from images.

The proposed approach automatically extracts 21

MediaPipe landmarks (x and y coordinates of

specific points) from the hand and feeds a deep

neural architecture to model and recognize different

hand gestures in a subject-independent scenario.

This system analyzes the effect of using different

landmark-based representations as inputs to a light

RNN. It obtains significantly better accuracy when

using the raw coordinates of the 21 landmarks or the

hand anthropometric measures compared to using

the speed of the hand landmarks. In addition, a

performance tendency improvement is observed

when using landmark-based normalizations over the

raw coordinates. Moreover, using a limited number

of hand landmarks (the ones with higher variance

along the gestures) provides competitive gesture

recognition performance.

As future work, it would be interesting to apply

Multi-Modal Large Language Models to recognize

gestures by feeding the models with the original

images and landmarks representation. In addition, it

would be interesting to apply this framework using

other landmark detection models like OpenPose and

for other datasets with a wider variety of gestures or

microgestures (Chan et al., 2016) and/or related to

sign language recognition. Finally, a real-time

system that infers the prediction from a stream of a

camera could be an application on a real use case of

the proposed method.

ACKNOWLEDGEMENTS

This research was funded by Programa Propio 2024

from Universidad Politécnica de Madrid. Iván

Martín-Fernández research was supported by the

Universidad Politécnica de Madrid (Programa

Propio I+D+i). Sergio Esteban-Romero research was

supported by the Spanish Ministry of Education (FPI

grant PRE2022-105516).

This work was funded by Project ASTOUND

(101071191 — HORIZON-EIC-2021-

PATHFINDERCHALLENGES-01) of the European

Commission and by the Spanish Ministry of Science

and Innovation through the projects GOMINOLA

(PID2020-118112RB-C22), TRUSTBOOST

(PID2023-150584OB-C21) and BeWord (PID2021-

126061OB-C43), funded by

MCIN/AEI/10.13039/501100011033 and by the

European Union “NextGenerationEU/PRTR”.

REFERENCES

Bao, P., Maqueda, A. I., del-Blanco, C. R., & Garcia, N.

(2017). Tiny Hand Gesture Recognition without

Localization via a Deep Convolutional Network. Ieee

Transactions on Consumer Electronics, 63(3), 251-

257.

Benitez-Garcia, G., Olivares-Mercado, J., Sanchez-Perez,

G., Yanai, K., & Ieee Comp, S. O. C. (2021, Jan 10-

15). PIS Hand: A Video Dataset and Benchmark for

Real Time Continuous Hand Gesture Recognition.

Paper presented at the 25th International Conference

on Pattern Recognition (ICPR), Electr Network.

Chan, E., Seyed, T., Stuerzlinger, W., Yang, X.-D.,

Maurer, F., & Acm. (2016). User Elicitation on

Single-hand Microgestures. [Proceedings Paper]. 34th

Annual Chi Conference on Human Factors in

Computing Systems, Chi 2016, 3403-3414.

Raw

Coordinates

Raw

Coordinates +

First Frame

Norm

Raw

Coordinates +

Current Frame

Norm

Anthropometric

Measures

Accuracy (%)

Landmark-based Input Format

21 landmarks 6 landmarks 4 landmarks

Hand Gesture Recognition Using MediaPipe Landmarks and Deep Learning Networks

Gil-Martin, M., San-Segundo, R., Fernandez-Martinez, F.,

& Ferreiros-Lopez, J. (2020). Improving physical

activity recognition using a new deep learning

architecture and post-processing techniques.

Engineering Applications of Artificial Intelligence, 92.

Gil-Martin, M., San-Segundo, R., Fernandez-Martinez, F.,

& Ferreiros-Lopez, J. (2021). Time Analysis in

Human Activity Recognition. Neural Processing

Letters.

Gil-Martín, M., San-Segundo, R., & de Córdoba, R.

(2023). Hand Pose Recognition through MediaPipe

Landmarks. Paper presented at the Modeling

Decisions for Artificial Intelligence.

Gil-Martín, M., San-Segundo, R., Fernández-Martínez, F.,

& de Córdoba, R. (2020). Human activity recognition

adapted to the type of movement. Computers &

Electrical Engineering, 88, 106822.

Goodfellow, I. (2016). Deep learning. Cambridge,

Massachusetts: Cambridge, Massachusetts The MIT

Press.

Lugaresi, C., Tang, J., Nash, H., McClanahan, C.,

Uboweja, E., Hays, M., et al. (2019). MediaPipe: A

Framework for Building Perception Pipelines. ArXiv,

abs/1906.08172.

Luna-Jimenez, C., Gil-Martin, M., Kleinlein, R., San-

Segundo, R., Fernandez-Martinez, F., & Acm. (2023).

Interpreting Sign Language Recognition using

Transformers and MediaPipe Landmarks.

[Proceedings Paper]. Proceedings of the 25th

International Conference on Multimodal Interaction,

Icmi 2023, 373-377.

Mantecon, T., del-Blanco, C. R., Jaureguizar, F., &

Garcia, N. (2019). A real-time gesture recognition

system using near-infrared imagery. Plos One, 14(10).

Pheasant, S., & Haslegrave, C. M. (2006). Bodyspace:

Anthropometry, Ergonomics and the Design of Work.

Quinonez, Y., Lizarraga, C., & Aguayo, R. (2022).

Machine Learning solutions with MediaPipe.

[Proceedings Paper]. 2022 11th International

Conference on Software Process Improvement, Cimps,

212-215.

Trindade, P., Lobo, J., & Barreto, J. P. (2012, 13-15 Sept.

2012). Hand gesture recognition using color and

depth images enhanced with hand angular pose data.

Paper presented at the 2012 IEEE International

Conference on Multisensor Fusion and Integration for

Intelligent Systems (MFI).

Weiss, N. A. (2017). Introductory statistics: Pearson.

Zhang, S., Wei, Z., Nie, J., Huang, L., Wang, S., & Li, Z.

(2017). A Review on Human Activity Recognition

Using Vision-Based Method. Journal of Healthcare

Engineering, 2017.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence