Recognising Care-Related Activities with a Convolutional Neural

Network Using Inertial Measurement Units

Fenja T. Hesselmann

, Jannik Fleßner

, Alexander Pauls

, Vanessa Cobus

Frauke Koppelin

and Frank Wallhoff

Jade University of Applied Sciences, Oldenburg, Germany

Keywords:

Human Activity Recognition, CNN, Nurse Care, Accelerometer.

Abstract:

Sensor-based human activity recognition is a growing ﬁeld of research. In addition to recognising everyday

movements, situation-dependent activities can also be detected. This paper therefore aims to detect care-

speciﬁc movements. For this purpose, 13 different nursing activities were recorded with Inertial Measurement

Units (IMUs) worn on the body. In this paper, we present an approach on how the sensor data can be used

for recognition. Convolutional neural networks were used for classiﬁcation. The focus of this work is on two

different fusion approaches of the data to check which approach achieves better results. In the ﬁrst approach,

all data is fused at the beginning, while in the second one, a separate pipeline is designed for each sensor and

fused later. The results show that a later fusion technique provides a better F1 score of 90.2 % compared to a

model that considers all signals from the beginning (F1 score: 82.5 %).

1 INTRODUCTION

Activity recognition refers to the ability of a system

to recognise different activities on the basis of previ-

ously learned data from various sensors. The sensors

used can be accelerometers, gyroscopes or cameras,

for example. Based on the increasing availability of

such sensors, interest in research into Human Activity

Recognition (HAR) is also growing. The applications

are diverse: in sports medicine (Nadeem et al., 2021),

in fall detection (Stampﬂer et al., 2023; Ferrari et al.,

2020) or the recognition of activities of daily living

(Stampﬂer et al., 2023; Wan et al., 2020). Activities

can be recognised on the basis of static and dynamic

gestures using sensor data. These can be simple activ-

ities such as walking or lying down, or more complex

action sequences such as cooking (Ramanujam et al.,

2021).

Various activities are also carried out in the care

sector, e.g. giving medication or food or treating

wounds. These activities are repeated regularly and

must be documented to ensure quality and for liabil-

ity reasons. However, studies have shown that the

https://orcid.org/0000-0003-0395-3854

https://orcid.org/0000-0002-4367-1651

https://orcid.org/0000-0001-9175-5756

https://orcid.org/0000-0002-4690-0800

https://orcid.org/0000-0002-1827-9279

https://orcid.org/0000-0002-7791-3225

documentation effort amounts to approx. 20-30 % of

the regular working time (Joukes et al., 2018; Murad

et al., 2024). In addition, due to a lack of time, the

documentation is sometimes completed collectively

at the end of the shift, which can lead to errors, as

important aspects may have been forgotten to be en-

tered at the beginning of the shift (Moy et al., 2021).

However, surveys among carers have shown that they

would like to use the technology to make their daily

work easier (Seibert et al., 2020). HAR is one way

that can be used to relieve carers of the burden of doc-

umentation and reduce the time required.

Recognising nursing activities has been identiﬁed

as an important ﬁeld of research. The ﬁrst Nurse

Care Activity Recognition Challenge was launched in

2019. The aim was to classify various care activities

recorded with multiple sensors such as camera-based

systems (Lago et al., 2019). However, it should be

noted that dealing with patients always involves sen-

sitive data and for this reason aspects of data protec-

tion and ethics, as well as the avoidance of monitoring

the nursing staff must be taken into account.

The aim of the Data-Driven Health research

project is therefore to develop and evaluate auto-

mated nursing documentation based on technology-

supported activity analyses to relieve and support

clinical nursing practice. This paper presents the

ﬁrst results of activity recognition of nursing activ-

ities based on data measured with Inertial Measure-

ment Units (IMUs). The focus should be on the fusion

Hesselmann, F. T., Fleßner, J., Pauls, A., Cobus, V., Koppelin, F. and Wallhoff, F.

Recognising Care-Related Activities with a Convolutional Neural Network Using Inertial Measurement Units.

DOI: 10.5220/0013257400003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 727-733

ISBN: 978-989-758-731-3; ISSN: 2184-4305

727

of the used data.

In the following, we ﬁrst present related work that

has dealt with activity recognition in nursing. This is

followed by a description of our methods, including

the sensors used and the development of a machine

learning model. The results are then presented and

discussed. The paper concludes with a summary and

next steps.

2 RELATED WORK

Other researchers have also recognised the relevance

of relieving the burden on nursing staff by automat-

ing the documentation. Kaczmarek et al. attached

three IMUs to the upper body, speciﬁcally to the up-

per arms and to a belt. The IMUs recorded accelerom-

eter, magnetometer and gyroscope data. Seven sub-

jects performed six different activities, with all ac-

tivities having the context of mobilising or changing

the patient’s position. To recognise the classes, both

an unsegmented approach and recognition with seg-

mented data were performed. Two neural networks,

an LSTM and a CNN, were evaluated. The accuracy

was 55 %, but could be improved with further locali-

sation information, segmented data and daily routine

information (Kaczmarek et al., 2023).

In addition to accelerometer data, Konak et al.

also used camera data. The ﬁve accelerometer sen-

sors were attached to the wrists, ankles and pelvis.

Ten subjects performed 13 different activities, includ-

ing washing the body in bed, pushing a wheelchair

or preparing medication. Three deep learning mod-

els were trained and compared to recognise the dif-

ferent classes. It was also tested whether the use of

poses from the camera data has added value in terms

of recognition. The authors were able to achieve an

accuracy of up to 86 % with a Residual Neural Net-

work. It was also found that the optimal sensor po-

sitions are the wrists and pelvis (Konak et al., 2023).

However, this statement is difﬁcult to implement in

practical care, as objects on the wrists are usually pro-

hibited for reasons of hygiene and risk of injury. The

use of cameras is also difﬁcult in practice for data pro-

tection reasons. In addition, the selection of activi-

ties to be recorded is limited, so that activities such as

wound care are not taken into account.

Lago et al. took a different approach. They

recorded various activities using a marker-based mo-

tion capture system. For this purpose, 29 markers

were attached to the subjects’ bodies and their po-

sition was recorded using 16 infrared cameras. In

addition, the accelerometer data from a smartphone

in the subject’s breast pocket was recorded. This

data set was made available as part of several chal-

lenges. While six activities were to be recognised in

the ﬁrst Nurse Care Activity Recognition Challenge,

in which all modalities were available (Lago et al.,

2019), 28 activities had to be classiﬁed using only

the accelerometer data in the fourth challenge (Inoue

et al., 2022). Although the challenges make an impor-

tant contribution, a critical reﬂection must be made on

the choice of sensor modalities. The use of a motion

tracking system with several cameras is complex and

not practical. The wearing or use of smartphones is

also often not permitted in everyday clinical practice.

The work presented shows that there are already

various approaches for recognising care-related activ-

ities. However, the selected sensors and sensor po-

sitions must be critically considered in the design.

In collaboration with care professionals, Bruns et al.

have established functional and non-functional re-

quirements that a system for recognising care-related

activities should fulﬁl (Bruns et al., 2024). The work

by Wallhoff and Hesselmann (Wallhoff and Hessel-

mann, 2025) presented a concept of how an AI-based

assistance system could be structured to support doc-

umentation. In addition to the identiﬁcation of activi-

ties, the concept provides a review by the nursing staff

in order to improve recognition results.

3 METHODS

This chapter presents the methods used in this paper

to recognise care activities. Starting with the sen-

sors used and the procedure for data acquisition, fol-

lowed by the subsequent data processing. The chapter

concludes with a description of the classiﬁcation and

evaluation methods.

3.1 Data Recording

The data for this study was recorded using 10 Movella

DOT (previously Xsens DOT) Inertial Measurement

Units (IMUs) (Movella Inc., 2024). The sensors were

located on the outer sides of the wrists, upper arms,

ankles and thighs, as well as one sensor on the upper

back and one on the lower back. Although the po-

sition on the wrists is unfavourable in real life, we

chose this position for later comparison with other

work. A total of 23 output signals were recorded for

each sensor, each with a sampling rate of 60 Hz: four-

dimensional quaternion values, four-dimensional an-

gular velocities determined by the derivative of the

quaternion values, three-dimensional accelerometer

values, three-dimensional velocity values determined

by the derivative of the accelerometer values, three-

HEALTHINF 2025 - 18th International Conference on Health Informatics

728

dimensional magnetic ﬁeld values, three-dimensional

Euler angles and three-dimensional gyroscope values.

In addition, the study was recorded with a Mi-

crosoft Azure Kinect (Microsoft, 2024). The result-

ing video was used to label the IMU data and assign

it to the correct activities. The Azure Kinect was set

up approx. 130 cm away from the patient’s bed, at a

height of approx. 214 cm, so that the entire bed was

recorded from the side. The Kinect recorded at a sam-

pling frequency of 30 Hz.

During the study, the participants were asked to

carry out 14 different activities. These activities were

selected on the basis of an online survey in which

subjects were asked about the most frequently docu-

mented care activities. Figure 1 shows the distribution

of the individual activities in seconds. The activities

were performed on a simulation mannequin as a pa-

tient.

The data set was analysed in the university’s com-

petence and simulation laboratory for applied nursing

sciences at the Jade University of Applied Sciences.

Four trained nurses took part in the study, average

age: 44 years, average professional experience: 20

years. One session, in which the 13 activities were

carried out one after the other, lasted 23 minutes on

average. The participants performed two rounds each,

with a break between the rounds. In order to make it

as realistic as possible, no minimum implementation

time was speciﬁed.

The experiment complied with the Declaration of

Helsinki and was approved by the Ethics Commit-

tee of the University of Oldenburg under the approval

number Drs.EK/2024/027.

Figure 1: Distribution of the different activities in seconds.

3.2 Data Preprocessing

The activities were labelled manually and synchro-

nised with the acceleration signal. The Kinect video

recordings were used as a reference for annotating the

IMU data. Distinctive movements in both data sets

(e.g. performing a jumping jack) were used to syn-

chronise the time of the modalities. By annotating

the start and stop times in the video data, the corre-

sponding IMU data could be assigned to the respec-

tive activity performed. In addition to the activities,

e.g. the measurement of blood pressure, the individ-

ual actions, e.g. the placement of the blood pressure

cuff or its inﬂation, were also annotated. However,

the actions are not taken into account for this study.

The video recording of one participant stopped

during the recording of the second session, which

means that only part of this session is available and

the data could not be fully annotated.

The data set was divided into small, overlapping

windows, with a window length of 1 s (60 frames) and

an overlap of 50 %. A smaller window size is associ-

ated with better classiﬁcation performance. The pro-

cedure also facilitates the training process of a Con-

volutional Neural Network (CNN) (Bevilacqua et al.,

2019). The overlapping of the windows leads to a

higher number of training and test data. The recorded

356013 frames containing activities were thus divided

into 11868 windows. Each window was assigned to

an activity or class, which was determined by the re-

spective class majority. If there was only unlabelled

data in a window, e.g. because materials were re-

trieved, this window was ignored. Because the ac-

tivities have different execution times and the partici-

pants performed the activities at different speeds, the

data set is not balanced, see ﬁgure 1.

The data is then normalised so that all signals

are in a similar range. To do this, we used Z-

normalisation, which is already used in various works

unzner et al., 2017). The value x can be normalised

using formula 1, where µ and σ are the mean and stan-

dard deviation of the individual signals. For normal-

isation, the data is ﬁrst divided into training and test

data. µ and σ are determined for the training part and

then applied to the test part. This ensures that the val-

ues of the training data set cannot inﬂuence the nor-

malisation process.

′

= (x − µ)/σ (1)

3.3 Classiﬁcation

To classify the care activities, a Convolutional Neural

Network (CNN) is used in this work. We opted for

this because a CNN has delivered good results in other

work compared to other neural networks (Ramanujam

et al., 2021).

In addition to the input layer, a CNN consists of

one or more convolutional and pooling layers, fully

connected or dense layers and an output layer. The

Recognising Care-Related Activities with a Convolutional Neural Network Using Inertial Measurement Units

729

convolutional layer can extract features from the input

data using kernels, while the pooling layer reduces the

resolution of the features by discarding unnecessary

information. Finally, all nodes in the fully connected

layers are connected to all nodes in the output layer

(Bhatt et al., 2021). The output, e.g. the classiﬁca-

tion result, is displayed in the output layer. CNNs are

often used in image recognition (Lecun and Bengio,

1995; Bhatt et al., 2021), but are also used in other

applications (Ersavas et al., 2024; Bhatt et al., 2021).

As several different signals were recorded with the

IMUs (including quaternion and accelerometer data),

two models were trained. The models were devel-

oped with Tensorﬂow (Abadi et al., 2015). The CNN

contains two consecutive convolutional layers, each

with 64 ﬁlters and a kernel size of 3. The convo-

lutional layers use a ReLu activation function. This

is followed by a dropout layer (50 %), a fully con-

nected layer with 100 neurons and a softmax layer.

The model was trained with an Adam optimiser, a

batch size of 128 and 100 epochs. The architecture

and parameters were chosen based on (Gholamiango-

nabadi et al., 2020).

The two models differ in their fusion approach i.e.

their data input. In Model

All

, all signals from all sen-

sors were merged at the beginning and transferred to

the input layer. As a result, the data input had a size

of N x 230 x 60, where N is the number of segments.

The 230 describe the 23 signals multiplied by ten sen-

sors and the 60 denote the frames of a window. The

schematic structure is shown in ﬁgure 2 a).

In Model

Signal

, the data is divided up according

to the signals. A separate pipeline is developed for

each signal, which is fused before the fully connected

layer. This allows signal-speciﬁc models to be de-

veloped. Depending on the dimension of the signal

(three- or four-dimensional), the data input of the in-

dividual pipelines was N x 30 x 60 or N x 40 x 60.

The structure of the CNN is shown in ﬁgure 2 b).

3.4 Evaluation

The two models Model

All

and Model

Signal

are each

tested with a stratiﬁed 10-fold cross validation. The

classiﬁcation accuracy and the weighted F1 score

were calculated as evaluation metrics. The weighted

F1 score considers the correct classiﬁcation of each

class in equal parts by using the precision, recall and

proportion of the class in the data set. As the data set

is unbalanced, the accuracy is a less suitable measure

than the F1 score. Nevertheless, it was calculated for

comparison with other work. The training and valida-

tion loss was also considered to ensure that the model

does not overﬁt. This can often occur with smaller

data sets.

Figure 2: Schematic structure of the two CNN architectures

and fusion approaches. a) shows the early fusion of all sen-

sor data in Model

All

, b) shows the separate pipelines and

the late fusion in Model

Signal

4 RESULTS

Figure 3 shows the average training and validation

loss curve as well as the curve of the F1 score of

Model

All

over the ten folds. The loss respectively the

F1 score is plotted on the y-axis and the epochs on the

x-axis. While the loss of the training data decreases

constantly, the loss of the validation data starts to in-

crease slightly from epoch 43 onwards. It can also be

seen that the loss of the validation data is higher than

the loss of the training data. The F1 diagram (see Fig.

3 below) shows that the training F1 score level off at

around 91 %, while the test F1 score stabilises at just

over 80 %.

Figure 4, on the other hand, shows the training

and validation loss of Model

Signal

. Here, too, it can be

seen that the loss curve of the validation data is higher

than the loss curve of the training data, although the

two curves almost equalise. However, the curves de-

crease up to a point of stability and do not rise again

afterwards. The performance of the model no longer

improves around the 55th training epoch. The F1 di-

agram (see Fig. 1 below) shows that the graphs of the

training and validation data approximate each other.

While the training data achieves an F1 score of 96 %,

the F1 scores of the validation data stabilise at around

91 %.

Table 1 shows the accuracy and F1 score for both

models. With an accuracy of 90.3 % and an F1

HEALTHINF 2025 - 18th International Conference on Health Informatics

730

Figure 3: Loss and F1 curve of the training and validation

data of Model

All

Figure 4: Loss and F1 curve of the training and validation

data of Model

Signal

Table 1: Performance comparison of Model

All

and

Model

Signal

Model Model

All

Model

Signal

Accuracy 82.6 % ± 1.3 % 90.3 % ± 1.8 %

F1 score 82.5 % ± 1.2 % 90.2 % ± 1.7 %

score of 90.2 %, Model

Signal

has a better result than

Model

All

, which has an accuracy of 82.6 % and an F1

score of 82.5 %.

5 DISCUSSION

The main objective of this work is to evaluate the im-

pact of model selection and data fusion on the ability

of the model to classify care activities. The results

have shown that later fusion of the data provides bet-

ter accuracies and F1 scores than fusion right at the

beginning. This result can be explained by the fact

that later fusion models are based on the idea of deter-

mining individual convolutional ﬁlters for each sen-

sor modality. This allows each pipeline of the model

to extract the most meaningful features. In contrast,

when the data is fused early, the model appears to dis-

card important features, resulting in lower calculated

scores.

As we have attempted to recognise different nurs-

ing activities in this paper than the work presented in

the Related Work chapter, the calculated scores can-

not be fully compared. Nevertheless, the accuracy

and F1 score for both CNN models are in a similar

range to other studies that have performed HAR with

a CNN and evaluated with cross-validation (Konak

et al., 2023; Tsokov et al., 2021; Moya Rueda et al.,

2018).

However, the loss curve should also be consid-

ered. The decreasing training curves of the loss val-

ues show that the models are gradually reﬁning their

predictions. The validation curve, on the other hand,

shows the performance of the model with unseen data.

In the case of Model

All

, the continuing fall in the

training curve, the slight rise in the validation curve

and the difference between the curves indicate a slight

overﬁt of the model. The model makes accurate pre-

dictions for the training data, but not for new data.

This could be due to the small but also unbalanced

data set. With Model

Signal

, on the other hand, the

convergence of the two curves shows that the model

learns meaningful features and overﬁtting to the train-

ing data is avoided. As a result, this model is prefer-

able to Model

All

Related work has shown that the use of k-fold

cross-validation provides better accuracy than leave-

one-subject-out cross-validation (LOSOCV) (Gho-

lamiangonabadi et al., 2020; Konak et al., 2023). This

is due to the fact that each subject has individual dif-

ferences in the performance of activities and with k-

fold cross validation, data from each subject is avail-

able in the training set. To ensure that the model can

also recognise the activities of new, unknown persons,

a LOSOCV should also be carried out. However, this

was not possible with the limited number of subjects

available.

6 CONCLUSION

In this paper, two CNNs were developed to recognise

13 care activities. Various signals from ten body-worn

IMUs were used for this purpose. The paper describes

the pre-processing procedure, the classiﬁers used and

the results obtained. The results show an accuracy

of 82.6 % and an F1 score of 82.5 % for the CNN

with all signals and an accuracy of 90.3 % and an F1

score of 90.2 % for the signal-dependent CNN. It has

been shown that with multisensor data, it is useful to

Recognising Care-Related Activities with a Convolutional Neural Network Using Inertial Measurement Units

731

calculate a separate pipeline for each signal and then

fuse them. This procedure provides better accuracies

and F1-scores than fusing the data at the beginning.

The next step is to analyse more data from other

care professionals and investigate which sensor posi-

tion is most relevant. As soon as several participants

have taken part, a leave one subject out cross valida-

tion can be carried out instead of a 10-fold cross val-

idation in order to check how accurate the model is

with an unknown participant. Furthermore, the data

from the Kinect camera will be integrated into the ac-

tivity recognition to check whether this can improve

recognition and the F1 score.

ACKNOWLEDGEMENTS

This study was supported by the Lower Saxony Min-

istry for Science and Culture with funds from the gov-

ernmental funding initiative zukunft.niedersachsen

of the Volkswagen Foundation, project ”Data-driven

health (DEAL)”.

REFERENCES

Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z.,

Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin,

M., Ghemawat, S., Goodfellow, I., Harp, A., Irving,

G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kud-

lur, M., Levenberg, J., Man

e, D., Monga, R., Moore,

S., Murray, D., Olah, C., Schuster, M., Shlens, J.,

Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Van-

houcke, V., Vasudevan, V., Vi

egas, F., Vinyals, O.,

Warden, P., Wattenberg, M., Wicke, M., Yu, Y., and

Zheng, X. (2015). TensorFlow: Large-scale machine

learning on heterogeneous systems. Software avail-

able from tensorﬂow.org.

Bevilacqua, A., MacDonald, K., Rangarej, A., Widjaya, V.,

Caulﬁeld, B., and Kechadi, T. (2019). Human Activ-

ity Recognition with Convolutional Neural Networks.

In Brefeld, U., Curry, E., Daly, E., MacNamee, B.,

Marascu, A., Pinelli, F., Berlingerio, M., and Hurley,

N., editors, Machine Learning and Knowledge Dis-

covery in Databases, pages 541–552, Cham. Springer

International Publishing.

Bhatt, D., Patel, C., Talsania, H., Patel, J., Vaghela, R.,

Pandya, S., Modi, K., and Ghayvat, H. (2021). CNN

Variants for Computer Vision: History, Architecture,

Application, Challenges and Future Scope. Electron-

ics, 10(20):2470.

Bruns, F. T., Pauls, A., Koppelin, F., and Wallhoff, F.

(2024). Activity Recognition of Nursing Tasks in a

Hospital: Requirements and Challenges. In Salvi, D.,

Van Gorp, P., and Shah, S. A., editors, Pervasive Com-

puting Technologies for Healthcare, pages 235–243,

Cham. Springer Nature Switzerland.

Ersavas, T., Smith, M. A., and Mattick, J. S. (2024). Novel

applications of Convolutional Neural Networks in the

age of Transformers. Scientiﬁc Reports, 14(1).

Ferrari, A., Micucci, D., Mobilio, M., and Napoletano, P.

(2020). On the Personalization of Classiﬁcation Mod-

els for Human Activity Recognition. IEEE Access,

8:32066–32079.

Gholamiangonabadi, D., Kiselov, N., and Grolinger, K.

(2020). Deep Neural Networks for Human Activ-

ity Recognition With Wearable Sensors: Leave-One-

Subject-Out Cross-Validation for Model Selection.

IEEE Access, 8:133982–133994.

Inoue, S., Hamdhana, D., Garcia, C., Kaneko, H., Nahid,

N., Hossain, T., Alia, S. S., and Lago, P. (2022). 4th

Nurse Care Activity Recognition Challenge Datasets.

Joukes, E., Abu-Hanna, A., Cornet, R., and de Keizer,

N. F. (2018). Time Spent on Dedicated Patient Care

and Documentation Tasks Before and After the Intro-

duction of a Structured and Standardized Electronic

Health Record. Appl Clin Inform, 09(01):046–053.

Kaczmarek, S., Fiedler, M., Bongers, A., Wibbeling, S.,

and Grzeszick, R. (2023). Dataset and Methods for

Recognizing Care Activities. In Proceedings of the

7th International Workshop on Sensor-Based Activity

Recognition and Artiﬁcial Intelligence, iWOAR ’22,

New York, NY, USA. Association for Computing Ma-

chinery.

Konak, O., Wischmann, A., van De Water, R., and Arnrich,

B. (2023). A Real-time Human Pose Estimation Ap-

proach for Optimal Sensor Placement in Sensor-based

Human Activity Recognition. In Proceedings of the

8th International Workshop on Sensor-Based Activity

Recognition and Artiﬁcial Intelligence, iWOAR ’23,

New York, NY, USA. Association for Computing Ma-

chinery.

Lago, P., Alia, S. S., Takeda, S., Mairittha, T., Mairit-

tha, N., Faiz, F., Nishimura, Y., Adachi, K., Okita,

T., Charpillet, F., and Inoue, S. (2019). Nurse care

activity recognition challenge: summary and results.

In Adjunct Proceedings of the 2019 ACM Interna-

tional Joint Conference on Pervasive and Ubiquitous

Computing and Proceedings of the 2019 ACM Inter-

national Symposium on Wearable Computers, Ubi-

Comp/ISWC ’19 Adjunct, page 746–751, New York,

NY, USA. Association for Computing Machinery.

Lecun, Y. and Bengio, Y. (1995). Convolutional Networks

for Images, Speech, and Time-Series.

Microsoft (2024). Azure kinect dk hardware speciﬁ-

cations. https://learn.microsoft.com/en-us/previous-

versions/azure/kinect-dk/hardware-speciﬁcation, Last

accessed on 2024-11-07.

Movella Inc. (2024). DOT. https://www.movella.com/

products/wearables/movella-dot, Last accessed on

2024-11-07.

Moy, A. J., Schwartz, J. M., Chen, R., Sadri, S., Lucas, E.,

Cato, K. D., and Rossetti, S. C. (2021). Measurement

of clinical documentation burden among physicians

and nurses using electronic health records: a scoping

review. Journal of the American Medical Informatics

Association, 28(5):998–1008.

HEALTHINF 2025 - 18th International Conference on Health Informatics

732

Moya Rueda, F., Grzeszick, R., Fink, G. A., Feldhorst,

S., and Ten Hompel, M. (2018). Convolutional Neu-

ral Networks for Human Activity Recognition Using

Body-Worn Sensors. Informatics, 5(2):26.

unzner, S., Schmidt, P., Reiss, A., Hanselmann, M.,

Stiefelhagen, R., and D

urichen, R. (2017). CNN-

based Sensor Fusion Techniques for Multimodal Hu-

man Activity Recognition. In Proceedings of the 2017

ACM International Symposium on Wearable Comput-

ers, UbiComp ’17, page 158–165. ACM.

Murad, M. H., Vaa Stelling, B. E., West, C. P., Hasan,

B., Simha, S., Saadi, S., Firwana, M., Viola, K. E.,

Prokop, L. J., Nayfeh, T., and Wang, Z. (2024). Mea-

suring Documentation Burden in Healthcare. Journal

of General Internal Medicine, 39(14):2837–2848.

Nadeem, A., Jalal, A., and Kim, K. (2021). Automatic

human posture estimation for sport activity recog-

nition with robust body parts detection and entropy

markov model. Multimedia Tools and Applications,

80(14):21465–21498.

Ramanujam, E., Perumal, T., and Padmavathi, S.

(2021). Human Activity Recognition With Smart-

phone and Wearable Sensors Using Deep Learning

Techniques: A Review. IEEE Sensors Journal,

21(12):13029–13040.

Seibert, K., Domhoff, D., Huter, K., Krick, T., Rothgang,

H., and Wolf-Ostermann, K. (2020). Application of

digital technologies in nursing practice: Results of a

mixed methods study on nurses’ experiences, needs

and perspectives. Zeitschrift f

ur Evidenz, Fortbildung

und Qualit

at im Gesundheitswesen, 158–159:94–106.

Stampﬂer, T., Elgendi, M., Fletcher, R. R., and Menon, C.

(2023). The use of deep learning for smartphone-

based human activity recognition. Frontiers in Public

Health, 11.

Tsokov, S., Lazarova, M., and Aleksieva-Petrova, A.

(2021). Accelerometer-based human activity recogni-

tion using 1d convolutional neural network. IOP Con-

ference Series: Materials Science and Engineering,

1031(1):012062.

Wallhoff, F. and Hesselmann, F. T. (2025). Nursing Ac-

tivity Recognition for Automated Care Documenta-

tion in Clinical Settings. In Bramer, M. and Stahl,

F., editors, Artiﬁcial Intelligence XLI, pages 244–250,

Cham. Springer Nature Switzerland.

Wan, S., Qi, L., Xu, X., Tong, C., and Gu, Z. (2020).

Deep Learning Models for Real-time Human Activ-

ity Recognition with Smartphones. Mobile Networks

and Applications, 25(2):743–755.

Recognising Care-Related Activities with a Convolutional Neural Network Using Inertial Measurement Units

733