Deep Learning for Frailty Classification Using Raw Inertial Sensor

Gait Data

Arslan Amjad

, Agnieszka Szczęsna

and Monika Błaszczyszyn

Department of Computer Graphics, Vision and Digital Systems, Faculty of Automatic Control, Electronics and Computer

Science, Silesian University of Technology, Akademicka 16, 44-100 Gliwice, Poland

Department of Physical Education and Sport, Faculty of Physical Education and Physiotherapy,

Opole University of Technology, Prószkowska 76, 45-758 Opole, Poland

Keywords: Deep Learning, Frailty, Gait, Healthcare, IMU Sensor.

Abstract: The Frailty is a significant health issue in older adults that increases the risk of disability, decline in

physiologic reserve and function, hospitalization, and even death. The social and economic impact of frailty

increased due to the higher healthcare costs and the medical resources. The intervention of early frailty

detection can prevent its progression and delay the disability, ultimately improving the quality of life in the

elderly population. This study aims to propose a frailty classification system based on gait data collected from

an Inertial Measurement Unit (IMU) sensor with the utilization of the Deep Learning (DL) approach. The

individual’s frailty status is classified as robust, pre-frail, or frail. A publicly available dataset of 163

participants was utilized to analyze the raw gait signals and find the most effective DL for extracting gait

patterns for frailty classification. DeepConvLSTM model has shown effective performance on raw IMU gait

data with a balanced accuracy, precision, recall, and F1-score of 91%. The results show that the proposed

methodology successfully classifies the pre-frail individuals, which demonstrate its potential to enhance

frailty detection and intervention in clinical settings. This ultimately provides an improved healthcare system

and a quality of life in elderly populations.

1 INTRODUCTION

The number of elderly individuals is increasing

dramatically as the world's population grows

(Hernigou et al., 2024). World Health Organization

data show this demographic trend: among the 8.1

billion population, people aged 60 years and older

will become 1.4 billion by 2030 and 2.1 billion by

2050 (Sun et al., 2024; United Nation, 2024; World

Health, 2024). Frailty is one of the most common and

fatal disorders in the elderly population (Hakeem et

al., 2023). Physical frailty is a multidimensional

condition that is defined as a decline in physiological

reserves. This makes older persons more vulnerable

to stresses and increases the possibility of negative

health effects (Kojima et al., 2018). Considering its

consequences link to increased illness, disability, and

death, this raises a significant public health concern

https://orcid.org/0000-0002-6711-4382

https://orcid.org/0000-0002-4354-8258

https://orcid.org/0000-0002-1723-4001

(Pasieczna et al., 2023). To reduce the burden of

frailty on the healthcare system and enhance the

quality of life for the aging population, it is essential

to address it through early identification, precise

assessment, and effective management.

In order to reduce the risk of frailty among older

adults, it is essential to develop an objective

healthcare solution. Traditional clinical frailty

assessment methods are time-consuming and need

specialized equipment and experienced healthcare

personnel (Obbia et al., 2020). To solve this issue,

wearable technology and advanced Machine

Learning (ML) algorithms have emerged as a

potential solution (Fan et al., 2023; Minici et al.,

2022). These technologies provide continuous, real-

time remote monitoring, allowing for early

identification and classification of frailty stages.

Amjad, A., Szcze¸sna, A. and Błaszczyszyn, M.

Deep Learning for Frailty Classiﬁcation Using Raw Inertial Sensor Gait Data.

DOI: 10.5220/0013127300003890

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 311-317

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

311

This study offers a smart stack of an Inertial

Measurement Unit (IMU) sensor and Deep Learning

(DL) technologies as a promising solution for frailty

classification. An IMU sensor was worn by each

participant to collect their raw IMU gait signals.

These signals were then pre-processed and converted

to the frequency domain in order to capture the

underlying patterns. Following this, a Deep Learning

(DL) algorithm was used to extract intrinsic gait

parameters for frailty classification into frail, pre-

frail, or robust stages.

This study has two main objectives: 1) analyzing

the raw IMU gait signals for frailty classification and

2) finding the most effective DL algorithm for frailty

classification using raw IMU gait data. The ultimate

goal of this research is to develop an early frailty

detection system that will detect the frailty stage

timely and prevent the frailty from progressing in

older adults. Early detection of frailty allows

individuals to seek medical advice and take

appropriate measures, which lowers the total cost of

healthcare for society. This study proposed an

intelligent frailty assessment system that will be

expanded into a real-time application, increasing its

use and impact in clinical settings.

The paper is organized as follows: Section 2

outlines the relevant literature work; Section 3

explains the research methodology, including the

dataset and the application of the DL algorithm for

frailty classification; Section 4 provides the results

with discussion; and finally, Section 5 concludes the

paper and suggests directions for future work.

2 LITERATURE REVIEW

For clinical gait analysis, the most commonly used

DL algorithms in the previous studies are:

Convolutional Neural Networks (CNN), Recurrent

Neural Networks (RNN), and Auto-Encoders (AE).

These algorithms became popular due to their ability

to analyze complex time-series data and

automatically extract features from raw IMU sensor

data, making them appropriate for applications such

as gait analysis.

The studies (García et al., 2022; Kou et al., 2024),

and (Li et al., 2024) classify the fall risks using CNN-

LSTM and CNN-BiLSTM algorithms. The study

(García et al., 2022) used a 3-D IMU device placed

on a wrist and leveraged the CNN-LSTM model to

achieve an accuracy of 93.60%. Whereas the study

(Kou et al., 2024) achieved an F1-score of 95.18%

and the study (Li et al., 2024) obtained an accuracy of

98.40%.

The study (Kamran et al., 2021) explored the

utility of 1-D CNN for automatically assessing

balance using data from a single IMU worn on the

lower back. They also compared the results with

handcrafted features. DL provided significant results

with an AUROC of 0.81. Another study (Hauth et al.,

2021) utilized three IMU sensors while performing

daily activities. The BiLSTM model outperformed

with an AUROC score of 0.87.

Another approach used in the previous studies

(Butt et al., 2020; San-Segundo et al., 2019; Sánchez-

DelaCruz et al., 2019) is the transformation of raw

IMU signals into images. This structured format of

input leverages the DL algorithms to extract more

enhanced features. The overview of previous studies

that utilized raw IMU gait signals with DL algorithms

is shown in Table 1.

Table 1: Overview of relevant studies that utilized raw IMU

gait data for frailty analysis.

Ref. Algorithms Task Outcomes

(García et

al., 2022)

CNN-LSTM

Falls risks

Accuracy =

93.60%

(Kou et al.,

2024)

CNN-LSTM

F1-score =

95.18%

(Li et al.,

2024)

CNN-

BiLSTM

Accuracy =

98.40%.

(Kamran et

al., 2021)

1-D CNN

Assess

balance

AUROC =

0.81

(Hauth et al.,

2021)

BiLSTM

AUROC =

0.87

3 METHOD

The research methodology consists of key steps,

which include: 1) the analysis of raw IMU gait data

and assigning the frailty status label to each

participant; 2) pre-processing of the raw IMU data

and data formatting using sliding window technique

and wavelet transformation; and 3) implementation of

the DL algorithm to classify the frailty into frail, pre-

frail, or robust stages. The research methodology is

illustrated in Figure 1.

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

312

Figure 1: Research methodology.

3.1 Dataset

The GSTRIDE (García-de-Villa et al., 2023) dataset

was utilized in this study. It’s a publicly available

dataset that consists of 163 (45 mean and 118 women)

older adults. Their ages range from 70 to 98 years and

an average weight of 64.2 to 77.3 kg. The list of

parameters available in the GSTRIDE database is

shown in Table 2.

Table 2: List of Parameters available in GSTRIDE

database.

Category Parameter Description

Socio

Demo-

graphic

Age

Average age of the

subjects (years)

Gender Male/Female

Living

Environment

Type of living

environment (e.g.,

Home, Assisted

Living)

Anatomical

Weight

Average weight of

the subjects (kg)

Height

Average height of

the subjects (cm)

Body Mass Index

(BMI)

Average BMI of

the subjects

(kg/m²)

Cognitive

Global

Deterioration

Scale (GDS)

Index

Average GDS

index of the

subjects (scale 1-

Functional

4-metre Gait

Speed Test

Average time

taken (seconds)

Hand Grip

Strength

Average hand grip

strength (kg)

Timed Up and

Go (TUG)

Average time

taken for TUG test

(seconds)

Short Physical

Performance

Battery (SPPB)

Average SPPB

score

Short Falls

Efficacy Scale

International

(FES-I)

Average FES-I

score

For raw IMU signal acquisition, two IMU sensors

(CSIC and Gaitup) were used, with only one sensor

worn on the foot of each participant during 15

minutes of gait (García-de-Villa et al., 2023). The

reason for using two different sensors with varying

frequencies was to assess the effect of varying

configurations of sensors on the spatio-temporal

estimation. The authors reported a minimal effect of

these varying configurations of sensors on spatio-

temporal estimation, although there was a slight

Deep Learning for Frailty Classiﬁcation Using Raw Inertial Sensor Gait Data

313

variation in the accuracy of estimation (García-

Villamil et al., 2021).

3.2 Class Labelling of Participants

The Standardized Fried's phenotype (Fried et al.,

2001) test was adopted to label the frailty status of

each participant. In this test, the Frailty Index (FI)

score is calculated using the five parameters. Each

parameter’s score is assigned a score of 0 or 1. The

final FI score is calculated by summing the score of

all parameters (ranges from 0 to 5) (García-de-Villa

et al., 2023). The class label is assigned to each

participant based on FI score. If the FI score is 0, then

the frailty class label is “Robust”. If the FI score is 1

or 2, then the frailty class label is “Pre-frail”,

otherwise the frailty label is “Frail”, as shown in

Figure 2.

Figure 2: Criteria for assigning the frailty label to each

participant.

3.3 Data Pre-Processing

After raw IMU signal data acquisition and labelling,

pre-processing became a critical step for further

analysis. In this stage, outliers were removed from

raw IMU signals, and the signals were normalized

using the “StandardScaler” function. The data was

then segmented into smaller chunks using the

“Sliding Window” technique (Jaén-Vargas et al.,

2022), which allows the extraction of spatio-temporal

features from the time-series IMU signals. The

window size set in this study was 200, with a stride of

50. Next, a wavelet transformation is applied to each

segment to capture both time and frequency domain

features. The “pywt.wavedec” function was used for

Daubechies wavelet of order 1 (“db1”). This

frequency transformation is suitable for raw IMU

signals to capture the sharp changes in the signals

(Chakraborty et al., 2020; Kuduz et al., 2023; Michau

et al., 2022).

At the end, the segmented windows were divided

into 75% training set. The remaining segments were

divided into equal sets for validation and testing. The

code is implemented in Python, version 3.5, using

Spyder as the development environment.

3.4 Deep Learning (DL) Algorithm

Architecture

The input data is ready after pre-processing steps. It

can be input to a DL algorithm for frailty

classification. The DeepConvLSTM (Ordóñez et al.,

2016) algorithm was utilized for this purpose. The

model consists of a convolutional layer with Long

Short-Term Memory (LSTM) layers to capture both

the spatial and temporal (spatio-temporal) features in

raw IMU signals, which makes it an effective

algorithm for frailty classification.

Two different DeepConvLSTM models were

created and trained on the training dataset. The best

model was selected based on high accuracy and

minimum losses on both training and validation

datasets. After finalizing the training process, the best

model’s hyperparameters were saved and tested on

the test dataset. The models were created using an

open-source Python’s library, McFly (van Kuppevelt

et al., 2020).

In this study, the architecture of the best

DeepConvLSTM model was initialized with a

“BatchNormalization” layer followed by a reshape

operation. Following this, a 2D convolutional layer

with 54 filters was applied, followed by normalized

and activated layers. After convolution, the resulting

tensor is reshaped to prepare it for recurrent

processing. Mathematically, the convolutional

process is defined as: for an input 𝒙∈ℝ



(where 𝑇 is time, 𝑊 is width, H is height, and 𝐶 is

channels), the convolutional process is depicted in

(1).

𝒙



𝐶𝑜𝑛𝑣



(

𝒙



(1)

Conv

represents the 2D convolutional operation

in the model; the overall equation of the

convolutional process with filters F is:

𝒙



[

𝑡,𝑤,ℎ,𝑐

]



𝒙

[

𝑡,𝑤 𝑖,ℎ 𝑗,𝑘

]

∙𝑭

[

𝑖,𝑗,𝑘,𝑐

]

𝑏

[

𝑐

]



















(2)

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

314

Table 3: DeepConvLSTM model performance on test dataset.

DeepConvLSTM performance (Test Set)

Precision Recall F1-score Support Accuracy

Robust 0.94 0.93 0.93 39063

0.91

Pre-frail 0.92 0.90 0.91 38269

Frail 0.67 0.85 0.75 6541

Weighted Avg. 0.91 0.91 0.91 83873

In (2), the F

and F

are filter weight and height,

respectively. Whereas the F

represents input

channels and bias is represented with b[c].

The convolutional operation results were input

into the stack of LSTM layers. Four LSTM layers

were used in the model with 29, 95, 94, and 46 units,

respectively. After that, the dropout layer was added

to prevent the model overfitting. The model

concluded with the “TimeDistributed” layer with 3

units followed by “softmax” activation for the

classification task. The final description of a model

can be represented as:

Conv(54)−BN−LSTM(29)−LSTM(95)−LSTM(94)−

LSTM(46)−D−TD(3)−S.

4 RESULTS AND DISCUSSION

The evaluation criteria in this study consist of two

phases. In the first phase, the two created

DeepConvLSTM models were evaluated in the

training process based on their best training and

validation accuracy and minimum losses,

respectively. The second phase evaluated the best

selected DeepConvLSTM model on the testing

dataset using metrics such as accuracy, precision,

recall, and F1-score (Wasikowski et al., 2010).

In the training phase, the hyperparameters of

DeepConvLSTM models were fine-tuned using 25

epochs with a batch size of 64 and a stopping patience

of 3. The hyperparameters of the best model were

reported as a learning rate of 0.0268, a regularization

rate of 0.0004, and convolutional filters and LSTM

dimensions as 54 and (29, 95, 94, 46), respectively.

The model achieved training and validation accuracy

of 95.18% and 94.14% with corresponding losses of

0.1163 and 0.1415, respectively, as shown in Figure

Figure 3: Training and validation losses of

DeepConvLSTM in training phase.

The DeepConvLSTM model’s performance on

test data is reported in Table 3. It shows that the

DeepConvLSTM achieved an accuracy of 91% on

test data. Whereas the overall frailty stage-wise

confusion matrix is depicted in Figure 4.

The results in Table 3 suggested that the model

effectively classified the pre-frail and robust

individuals but reported low precision in the case of

the frail class. This is due to the highly class-

imbalanced, as the frail class has fewer instances,

which may overlap the features with other classes.

This problem can be overcome by adding more frail

instances

utilizing data augmentation techniques to

synthetically increase the number of frail samples or

applying class-weighted loss functions and

oversampling methods like Synthetic Minority Over-

Sampling Technique (SMOTE) (Hosseini et al.,

2024) during model training. However, this study

used raw IMU sensor signals as input, keeping the

original data with its spatio-temporal properties.

This ensures the effectiveness of the DL model for

frailty classification.

Deep Learning for Frailty Classiﬁcation Using Raw Inertial Sensor Gait Data

315

Figure 4: Confusion matrix represents the performance of

DeepConvLSTM model.

Accurate classification is a major concern in

clinical settings, as it directly influences patient care

and intervention strategies. Overall, the DL model

performed effectively; better performance on frail

individuals will enhance clinical decision-making

and personalized care.

The proposed approach may have some

challenges when applied in the clinical healthcare

system. These challenges include maintaining the

privacy of data, facilitating real-time processing with

wearable IMU sensors, and smoothly integrating into

clinical workflows. Furthermore, it is crucial to

validate the system in real-world settings and achieve

generalizability across a variety of demographics.

The method's potential for the early frailty detection

task is highlighted by its adaptability to diverse

operational circumstances and scalability to multiple

sensor configurations.

5 CONCLUSION

In the world of a growing elderly population, frailty

is an important factor in the adverse health outcomes

among elders. Early and accurate detection of frailty

can significantly enhance clinical decision-making,

leading to better patient care and management.

This study proposed a sensor-based approach with

a DL algorithm to classify the frailty into robust, pre-

frail, or frail stages. The DeepConvLSTM model

demonstrated its effectiveness in frailty classification

using raw IMU sensor data, with an overall accuracy

of 91%. The performance of the DL model has shown

it's potential to develop a frailty classification system

that depicts the real-world clinical scenario.

The limitations of this research work are: 1) The

small size of the dataset limited the performance of

the DL model; and 2) A diverse dataset and the

selection of features may also affect the DL

performance. Future studies should focus on the

diverse types of sensors for the data collection. There

is also a need to develop a real-time application to

monitor the frailty status in a real-world clinical

environment.

REFERENCES

Butt, A. H., Cavallo, F., Maremmani, C., & Rovini, E.

(2020). Biomechanical parameters assessment for the

classification of Parkinson Disease using Bidirectional

Long Short-Term Memory. 2020 42nd Annual

International Conference of the IEEE Engineering in

Medicine & Biology Society (EMBC), 5761–5764. doi:

10.1109/EMBC44109.2020.9176051

Chakraborty, J., & Nandy, A. (2020). Discrete wavelet

transform based data representation in deep neural

network for gait abnormality detection. Biomedical

Signal Processing and Control, 62, 102076. doi:

10.1016/j.bspc.2020.102076

Fan, S., Ye, J., Xu, Q., Peng, R., Hu, B., Pei, Z., Yang, Z.,

& Xu, F. (2023). Digital health technology combining

wearable gait sensors and machine learning improve the

accuracy in prediction of frailty. Frontiers in Public

Health, 11. doi: 10.3389/fpubh.2023.1169083

Fried, L. P., Tangen, C. M., Walston, J., Newman, A. B.,

Hirsch, C., Gottdiener, J., Seeman, T., Tracy, R., Kop,

W. J., Burke, G., & McBurnie, M. A. (2001). Frailty in

Older Adults: Evidence for a Phenotype. The Journals

of Gerontology Series A: Biological Sciences and

Medical Sciences, 56(3), M146–M157. doi:

10.1093/gerona/56.3.M146

García-de-Villa, S., Neira, G. G.-V., Álvarez, M. N.,

Huertas-Hoyas, E., Ruiz, L. R., Del-Ama, A. J.,

Sánchez, M. C. R., & Jiménez, A. R. (2023). A database

with frailty, functional and inertial gait metrics for the

research of fall causes in older adults. Scientific Data,

10(1), 566. doi: 10.1038/s41597-023-02428-0

García-Villamil, G., Ruiz, L., Jiménez, A. R., Granja, F. S.,

& Rodríguez-Sánchez, M. C. (2021). Influence of

IMU’s Measurement Noise on the Accuracy of Stride-

Length Estimation for Gait Analysis. IPIN-WiP.

Retrieved from https://api.semanticscholar.org/

CorpusID:247314771

García, E., Villar, M., Fáñez, M., Villar, J. R., de la Cal, E.,

& Cho, S.-B. (2022). Towards effective detection of

elderly falls with CNN-LSTM neural networks.

Neurocomputing, 500, 231–240. doi:

10.1016/j.neucom.2021.06.102

Hakeem, F. F., Maharani, A., Todd, C., & O’Neill, T. W.

(2023). Development, validation and performance of

laboratory frailty indices: A scoping review. Archives

of Gerontology and Geriatrics, 111, 104995. doi:

10.1016/j.archger.2023.104995

Hauth, J., Jabri, S., Kamran, F., Feleke, E. W., Nigusie, K.,

Ojeda, L. V., Handelzalts, S., Nyquist, L., Alexander,

N. B., Huan, X., Wiens, J., & Sienko, K. H. (2021).

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

316

Automated Loss-of-Balance Event Identification in

Older Adults at Risk of Falls during Real-World

Walking Using Wearable Inertial Measurement Units.

Sensors, 21(14), 4661. doi: 10.3390/s21144661

Hernigou, P., Bumbasirevic, M., Pecina, M., & Scarlat, M.

M. (2024). Eight billion people, sixteen billion hip

joints today: are future orthopedists prepared to treat a

world of ultra-old patients and centenarians in 2050?

International Orthopaedics, 48(8), 1939–1944. doi:

10.1007/s00264-024-06245-x

Hosseini, I., Rojas, R. F., & Ghahramani, M. (2024). Fall

Risk Assessment Using Single IMU. 2024 IEEE

International Symposium on Medical Measurements

and Applications (MeMeA), 1–6. doi:

10.1109/MeMeA60663.2024.10596880

Jaén-Vargas, M., Reyes Leiva, K. M., Fernandes, F.,

Barroso Gonçalves, S., Tavares Silva, M., Lopes, D. S.,

& Serrano Olmedo, J. J. (2022). Effects of sliding

window variation in the performance of acceleration-

based human activity recognition using deep learning

models. PeerJ Computer Science, 8, e1052. doi:

10.7717/peerj-cs.1052

Kamran, F., Harrold, K., Zwier, J., Carender, W., Bao, T.,

Sienko, K. H., & Wiens, J. (2021). Automatically

evaluating balance using machine learning and data

from a single inertial measurement unit. Journal of

NeuroEngineering and Rehabilitation, 18(1), 114. doi:

10.1186/s12984-021-00894-4

Kojima, G., Iliffe, S., & Walters, K. (2018). Frailty index

as a predictor of mortality: a systematic review and

meta-analysis. Age and Ageing, 47(2), 193–200. doi:

10.1093/ageing/afx162

Kou, J., Xu, X., Ni, X., Ma, S., & Guo, L. (2024). Fall-risk

assessment of aged workers using wearable inertial

measurement units based on machine learning. Safety

Science, 176, 106551. doi: 10.1016/j.ssci.2024.106551

Kuduz, H., & Kaçar, F. (2023). A deep learning approach

for human gait recognition from time-frequency

analysis images of inertial measurement unit signal.

International Journal of Applied Methods in

Electronics and Computers. doi: 10.58190/

ijamec.2023.44

Li, C., Cai, Y., Li, Y., & Zhang, P. (2024). Fusion of Dual

Sensor Features for Fall Risk Assessment with

Improved Attention Mechanism. Traitement Du Signal,

41(1), 73–83. doi: 10.18280/ts.410106

Michau, G., Frusque, G., & Fink, O. (2022). Fully learnable

deep wavelet transform for unsupervised monitoring of

high-frequency time series. Proceedings of the National

Academy of Sciences, 119(8). doi: 10.1073/

pnas.2106598119

Minici, D., Cola, G., Giordano, A., Antoci, S., Girardi, E.,

Bari, M. Di, & Avvenuti, M. (2022). Towards

Automated Assessment of Frailty Status Using a Wrist-

Worn Device. IEEE Journal of Biomedical and Health

Informatics, 26

(3), 1013–1022. doi: 10.1109/JBHI.

2021.3100979

Obbia, P., Graham, C., Duffy, F. J. R., & Gobbens, R. J. J.

(2020). Preventing frailty in older people: An

exploration of primary care professionals’ experiences.

International Journal of Older People Nursing, 15(2).

doi: 10.1111/opn.12297

Ordóñez, F., & Roggen, D. (2016). Deep Convolutional and

LSTM Recurrent Neural Networks for Multimodal

Wearable Activity Recognition. Sensors, 16(1), 115.

doi: 10.3390/s16010115

Pasieczna, A. H., Szczepanowski, R., Sobecki, J.,

Katarzyniak, R., Uchmanowicz, I., Gobbens, R. J. J.,

Kahsin, A., & Dixit, A. (2023). Importance analysis of

psychosociological variables in frailty syndrome in

heart failure patients using machine learning approach.

Scientific Reports, 13(1), 7782. doi: 10.1038/s41598-

023-35037-3

San-Segundo, R., Navarro-Hellín, H., Torres-Sánchez, R.,

Hodgins, J., & De la Torre, F. (2019). Increasing

Robustness in the Detection of Freezing of Gait in

Parkinson’s Disease. Electronics, 8(2), 119. doi:

10.3390/electronics8020119

Sánchez-DelaCruz, E., Weber, R., Biswal, R. R., Mejía, J.,

Hernández-Chan, G., & Gómez-Pozos, H. (2019). Gait

Biomarkers Classification by Combining Assembled

Algorithms and Deep Learning: Results of a Local

Study. Computational and Mathematical Methods in

Medicine, 2019, 1–14. doi: 10.1155/2019/3515268

Sun, Q., Xia, X., & He, F. (2024). Longitudinal association

between Body mass index (BMI), BMI trajectories and

the risk of frailty among older adults: A systematic

review and meta-analysis of prospective cohort studies.

Archives of Gerontology and Geriatrics, 124, 105467.

doi: 10.1016/j.archger.2024.105467

United Nation. (2024). World Population Prospects 2024.

Retrieved from https://population.un.org/wpp/

van Kuppevelt, D., Meijer, C., Huber, F., van der Ploeg, A.,

Georgievska, S., & van Hees, V. T. (2020). Mcfly:

Automated deep learning on time series. SoftwareX, 12,

100548. doi: 10.1016/j.softx.2020.100548

Wasikowski, M., & Chen, X. (2010). Combating the Small

Sample Class Imbalance Problem Using Feature

Selection. IEEE Transactions on Knowledge and Data

Engineering, 22(10), 1388–1400. doi: 10.1109/

TKDE.2009.187

World Health, O. (2024). Ageing. Retrieved from

https://www.who.int/health-topics/ageing#tab=tab_1.

Deep Learning for Frailty Classiﬁcation Using Raw Inertial Sensor Gait Data

317