Routine Pattern Learning and Anomaly Detection Applied to Lone

Workers Through Topic Modeling

Ana Cravid

ao Pereira

, Mar

ılia Barandas

and Hugo Gamboa

1,2

Fraunhofer Portugal AICOS, Rua Alfredo Allen 455/461, 4200-135 Porto, Portugal

Laborat

orio de Instrumentac¸

ao, Engenharia Biom

edica e F

ısica da Radiac¸

ao (LIBPhys-UNL), Departamento de F

ısica,

Faculdade de Ci

encias e Tecnologia da Universidade Nova de Lisboa, Monte da Caparica, 2829-516 Caparica, Portugal

Keywords:

Pattern Learning, Anomaly Detection, Topic Modeling, Incremental Learning, Lone Workers.

Abstract:

Learning routines and identifying anomalous behaviors play a critical role in worker safety. Identifying devia-

tions from normal patterns helps prevent accidents, ensuring enhanced safety in complex environments. Topic

modeling is frequently used to discover hidden semantics patterns and is well-suited to the complexity of rou-

tines in human behavior. However, its utility in complex time-series analysis and as a baseline for anomaly

detection has not been widely explored. This work proposes a novel solution to accurately model complex rou-

tines using topic modeling, enabling the identiﬁcation of anomalies through a statistical approach. A dataset

of human movement recordings was collected over up to seven consecutive months, capturing the routines of

three lone workers, with each accumulating between 414 and 955 hours of recording time. This dataset served

as the basis for a comprehensive analysis of the results, showing strong alignment between visually observed

patterns in routines and the outcomes of the proposed method. Additionally, detecting anomalies across mod-

els with varying training days conﬁrms that online learning enhances the accuracy and adaptability of routine

modeling. Topic modeling allows for in-depth learning of routines, capturing latent patterns undetectable to

humans. This capability prevents abnormal events, leading to a proactive approach to predictive risk assess-

ment.

1 INTRODUCTION

Working alone is a common practice across vari-

ous sectors, spanning from transport and logistics to

healthcare and the manufacturing industry. With the

intensiﬁcation of remote working worldwide, it has

also become frequent in other ﬁelds. According to the

UK Ministry of Defence (2008), a lone worker is de-

ﬁned as someone who cannot be immediately assisted

by another person during their working day, whether

for a short or extended period. The market for lone

worker safety solutions is growing as more organi-

zations recognize the need for comprehensive man-

agement policies and safety systems to ensure em-

ployee well-being. Current strategies primarily focus

on detecting falls or prolonged inactivity. However,

lone workers may also face other hazards, such as as-

sault or hostage situations, which are inadequately ad-

dressed by existing safety reactive measures.

To address these challenges, key methodologies

for a more proactive risk prediction and categoriza-

tion for lone workers in real-time include routine pat-

tern learning, anomaly detection tailored to individual

worker routines, and online learning to incorporate

new patterns and respond to previously unseen sce-

narios.

Contributions. This study proposes a framework that

learns and adapts to each worker’s unique routine pat-

terns, allowing it to identify deviations that may indi-

cate critical anomalies. The proposed solution was

evaluated in a real-world setting, involving three lone

workers from different sectors, ranging from ofﬁce

work to the manufacturing industry. These workers

recorded their daily routines over 45 days across ﬁve

to seven consecutive months. Results demonstrated

the framework’s ability to learn routines and detect

anomalies in those routines. These contributions re-

sulted from a research path that addressed the follow-

ing research questions:

• RQ1. How can different worker routines be

learned in an unsupervised manner?

• RQ2. Are anomaly detection methods effective in

identifying deviations from an individual worker

routine?

• RQ3. How do different online learning tech-

niques impact routine pattern learning?

Pereira, A. C., Barandas, M. and Gamboa, H.

Routine Pattern Learning and Anomaly Detection Applied to Lone Workers Through Topic Modeling.

DOI: 10.5220/0013236600003911

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 1, pages 935-942

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

935

2 RELATED WORK

Over recent years, various strategies have been em-

ployed for pattern learning on human movement, uti-

lizing activity and spatial routines on daily or weekly

cycles. In the context of pattern learning, topic mod-

eling emerges as a relevant technique. Huynh et al.

(2008) contributed to the early application of topic

modeling for recognizing daily routines using time-

series activity data. Their approach begins by cluster-

ing features relevant to routines, followed by applying

Latent Dirichlet Allocation (LDA) to uncover hidden

topics within these routines. Similarly, a related study

from Tang (2021) explored the use of LDA to identify

location patterns in user trip behaviors using Global

Positioning System (GPS) coordinates. However, this

study did not consider the temporal factor of patterns.

Steinhauer et al. (2019) adopted a similar approach

in their study of telecommunication networks, inter-

preting run-time variable readings as words in LDA to

obtain representative topics. Furthermore, Seiter et al.

(2014) employed LDA to discover activity routines of

stroke rehabilitation patients, subsequently applying

KNN to map the discovered topics to speciﬁc rou-

tines. Sun et al. (2021) used a two-dimensional prob-

ability distribution, a variation of LDA, to discover

routine patterns in individual travel behavior. This

work does not consider the order of occurrences, es-

sential for routine comprehension.

In the context of anomaly detection in human rou-

tines, Sun et al. (2021) also explored the detection

of anomalies by using perplexity as a scoring func-

tion to assess whether new behaviors were predictable

(thus normal) or anomalous. Steinhauer et al. (2019)

also aimed to detect anomalies following topic mod-

eling by establishing a ”normal model” and relying

on an expert’s visual comparison to conﬁrm the pres-

ence and nature of anomalies. In their study on de-

tecting patterns in anomalous clusters, Abraham and

Nair (2018) applied LDA to deﬁne the normal topics

within the data, subsequently using a statistical test

to detect anomalies. Similarly, Thornton et al. (2020)

used LDA to identify normal patterns in avionics net-

works and employed a combination of variational in-

ference and statistical analysis to detect anomalies.

Based on studies on routine pattern learning, topic

modeling has shown promising results for in-depth

analysis of human movement routines. This work will

therefore explore its potential in complex time series

analysis, focusing on a critical use case: lone worker

safety. Human routines are complex, and a few days

or weeks may be insufﬁcient to capture them accu-

rately. To the best of our knowledge, no study has

addressed long-term routine learning in human move-

ment using both activity patterns and GPS data and

only a few have focused on anomaly detection within

individual routines, which is the aim of this study.

3 METHODS

The primary aim of this work is to develop a gener-

alized approach to predictive risk assessment based

on deviations from a user’s routine within the time-

series domain. To achieve this, a lone worker use

case was selected, and a general approach for rou-

tine learning, anomaly detection, and model updating

was developed. Figure 1 provides an overview of the

proposed approach. For the routine learning stage, a

topic modeling technique was applied, consisting of a

clustering step, document creation, LDA model train-

ing, and topic activation. For the anomaly detection

stage, a statistical approach was developed. Since the

solution is designed for real-life implementation, rou-

tine evolution was also considered. Therefore, model

updates were explored, using both retraining and In-

cremental Learning (IL) approaches.

3.1 Datasets and Preprocessing

For this study’s dataset, three lone workers volun-

teered to collect data on their daily work routines dur-

ing the ﬁrst semester of 2024. The typical daily rou-

tine of LW1 includes a commuting period via sub-

way or car and a work period in an ofﬁce. LW2 com-

mutes by car and has work periods that include being

in an ofﬁce and walking outdoors within a speciﬁc re-

gion. Lastly, LW3 is a manufacturing worker who has

a long commuting period by car and spends part of

the workday walking through facilities.

The recording protocol involved installing a log-

ger Android application on the workers’ smartphones,

which was activated by the user at the beginning and

end of each workday. During the collection period,

the lone workers did not record data every weekday

due to holidays, application issues, battery drain, and

forgetfulness in activating the application. As a re-

sult, this led to a total of 45, 46, and 81 days of usable

data for routine learning for LW1, LW2, and LW3,

respectively. Additionally, the volunteers were asked

to annotate any days on which activities and/or loca-

tions occurred that were not part of their routine. This

request was made to prevent the contamination of rou-

tine learning with anomalous patterns. Therefore, the

test days were selected based on these annotations,

corresponding to a percentage between 23% and 34%

of all available days. Table 1 summarizes the dataset

details for the data collected from each lone worker.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

936

Clustering

ROUTINE LEARNING

Statistical

approach

ANOMALY DETECTION

Retrain

MODEL UPDATE

Document creation

LDA model

Topics

Incremental

Learning

Figure 1: Overview of the main stages of the proposed approach.

Table 1: Summary of data for lone workers. Mean day du-

ration and total duration are presented in hours

LW 1 LW 2 LW 3

Number of train days 30 35 62

Number of test days 15 11 19

Mean day duration 9,22 11,03 11,79

Total duration 414,69 507,43 955,37

The application recorded GPS information (lati-

tude, longitude, and accuracy), human activity data

(i.e., a set of pre-deﬁned, automatically recognized

activities with associated conﬁdence levels), and tem-

poral records. The activity data was obtained with the

Activity Recognition API from Google (2021), which

detects user activities through mobile device sensors.

The detected activities are in vehicle, on bicycle, on

foot, running, still, tilting, unknown, and walking.

For preprocessing the human activity data, low-

conﬁdence activity labels were ﬁltered out using a

threshold of 80%. For GPS information, only loca-

tions with an accuracy of up to 30 meters were con-

sidered. Finally, to align the time intervals for both

activities and locations, we applied linear interpola-

tion with a one-minute step, resulting in a data struc-

ture where each minute contains location data and a

conﬁdence value for each activity label. The time of

day, weekday, and day of the month were added as

features in the ﬁnal data structure.

3.2 Routine Learning

This section is divided into two subsections: the ﬁrst

provides a brief overview of topic modeling, and the

second presents our proposed approach using topic

modeling for routine learning.

3.2.1 Topic Modeling

Topic modeling is a methodology mainly used in text

processing that allows the discovery of hidden pat-

terns and structures, i.e., topics, in collections of data.

In the case of a text document, topic modeling intents

to ﬁnd common patterns across the document and to

understand how its words are associated to make up

the existing topics. As long as the elements that com-

pose the document can be processed into sequences

of characters, latent patterns can be discovered. Thus,

this methodology can also be applied to routine iden-

tiﬁcation. The most common approach for topic mod-

eling is LDA, a generative probabilistic model devel-

oped by Blei et al. (2003). A graphical model repre-

sentation of LDA is presented in Figure 2, illustrating

how LDA transforms a collection of documents into

a probabilistic distribution of topics. Topic model-

ing assumes documents are formed by combining the

words from the existing topics in the correct propor-

tions - both the proportions of the words within the

topics and of the topics within the document. This

document is the input to train the model.

3.2.2 Proposed Approach

Before applying the LDA model, the input data struc-

ture is normalized, followed by the K-means cluster-

ing algorithm. The output clusters represent crucial

combinations of data elements that form the routine.

The elbow method was used to determine the opti-

mal number of clusters, having been chosen 7 clusters

for LW1 and LW3 and 8 clusters for LW2. Addition-

ally, one drawback of K-means is that it does not han-

dle points far from the centroids as outliers, assigning

them to one of the known clusters. To address this is-

sue, we adapted K-means by creating an ”outlier clus-

ter” for anomalous events. During the training phase,

no points are assigned to this outlier cluster. However,

in the testing phase, if the distance between the data

point and the cluster centroids exceeds three times the

Routine Pattern Learning and Anomaly Detection Applied to Lone Workers Through Topic Modeling

937

Topic 1 Topic 2 Topic 3

Doc 1 Doc 2 Doc 3 Doc 4

Document

Collection of

documents

Topic distribution along

documents of a corpus

Figure 2: Graphical model representation of LDA. The outer plate, M, represents documents and the inner plate, N, represents

the words and topics within a document. Each word, w, is assigned to a topic, whose association is represented by z. Parameter

α deﬁnes the topic distribution within the corpus, θ deﬁnes the topic distribution within a document and β deﬁnes the word

distribution within topics. The diagram is presented in the work of Blei et al. (2003).

standard deviation of the training data, is assigned to

the outlier cluster.

Before creating the document, distances must be

transformed into probabilities so that shorter dis-

tances to a centroid correspond to higher association

probabilities with its cluster. This conversion was

done according to the following equation presented

in the work of Huynh et al. (2008):

−d

/σ

∑

j=1

−d

/σ

, (1)

where ω

represents the weight, i.e., probability of a

cluster, σ denotes the standard deviation of the dis-

tances of all instances, d

is the distance of the in-

stance to the centroid of the cluster to which it be-

longs, d

is the distance of the instance to the centroid

of cluster j and K is the number of clusters.

Since an event spans several seconds rather than

a single point in time, a sliding window of ten min-

utes was deﬁned. This window captures information

over intervals by summing the probabilities associ-

ated with each instance within that time frame and

normalizing the result. The window size was selected

considering the duration of the events targeted. For

each sliding window, it was produced a list whose el-

ements were the clusters, with their quantities propor-

tional to the probabilities obtained. The collection of

these lists forms the document used to train the model.

The LDA model trained outputs an activation

function that indicates the composition of the top-

ics, i.e., the contribution of each cluster to each topic,

which is illustrated by the following example:

Topic

= Cluster

× a

1,1

+ ...+ Cluster

× a

1,K

...

Topic

= Cluster

× a

N,1

+ ... + Cluster

× a

N,K

(2)

where N represents the number of topics, K the num-

ber of clusters and a

N,K

corresponds to the activation

value of Cluster

in Topic

. This function is used to

decode the latent structures in the test data, which un-

dergoes the same processing as the training data, by

multiplying the cluster probabilities by the activation

function factors.

Despite the creation of an outlier cluster, the num-

ber of topics remains unchanged, as topics are inﬂu-

enced only by the clusters identiﬁed during training.

Therefore, this approach does not generate a topic that

detects anomalies. Instead, as highlighted by Equa-

tion 2, it minimizes the values of all other topics,

which might be an indicator of an anomaly.

3.3 Anomaly Detection

To the best of our knowledge, very few studies have

applied anomaly detection after pattern learning in

time series and no standard method exists. Therefore,

we implemented a statistical approach over the pat-

tern learning results to identify potential deviations.

This approach was applied to compare the behavior

of the topics in the training data with the activation of

the topics in the test data. For this purpose, the mean

value and corresponding standard deviation were cal-

culated for each topic and each timestamp across all

training days.

Let z be a set composed of N topics, z =

,...,z

}, which form a user’s routine. Let z

x,i

denote the topic z

at time instant i, and the mean and

standard deviation across all training days for z

x,i

represented by µ

x,i

and σ

x,i

, respectively. Given a test

instance where ˜z

x,i

represents the activation of z

x,i

, an

anomaly is detected through the following rule:

∑

x=1





˜z

x,i

− µ

x,i



> σ

x,i



(3)

where {.} is an indicator function that takes the

value 1 (anomaly) if the expression inside is true, and

0 (normal) otherwise. An anomaly is detected if more

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

938

Topic 0 Topic 1 Topic 2 Topic 3

(a) Normal day. (b) Anomalous day.

Figure 3: Top: Annotation of activities performed along the complete workday of a normal and anomalous day. The annotated

activities are not exhaustively descriptive. For instance, Commuting includes walking to and from the train station, subway

travel, and car transportation. Bottom: Topic activations from the model selected, applied to the same test day.

than half of the topics at time instant i deviate from

their expected range, deﬁned as µ

x,i

± σ

x,i

A post-processing technique was implemented to

require a minimum streak of samples ﬂagged as

anomalies to conﬁrm the detection, and a minimum

streak of normal samples to revert the status to nor-

mal. This safeguards the algorithm against inaccura-

cies and insigniﬁcant variations.

3.4 Model Update

Anomalies might be detected based on a pre-

established number of days. However, routine orig-

inally identiﬁed might have evolved. Therefore, it is

essential to maintain an updated routine baseline.

For the purpose of model updating, two ap-

proaches were studied: (a) Retraining the models

from scratch and (b) Applying Incremental Learning

(IL). While IL is computationally lighter but may re-

sult in catastrophic forgetting, model retraining con-

siders all data equally. In the case of the LDA model,

the implementation used supports incremental up-

dates and, therefore, both approaches were tested.

However, the approach of topic modeling includes

a preliminary step: the clustering algorithm before

the LDA model. In the IL approach, the clusters are

not updated with the new data, as it would demand

producing a new document for the LDA model and,

therefore, require retraining it instead of incremen-

tally updating it. Hence, it was assumed that the fea-

tures (i.e. clusters) are well-deﬁned based on the ini-

tial training data.

4 RESULTS AND DISCUSSION

Regarding routine learning, it is difﬁcult, if not im-

possible, to access all factors that compose a user’s

routine. Without a ground-truth routine, quantita-

tive evaluation is compromised, requiring relative

comparisons with other methods and the visualiza-

tion of semi-annotated data. Therefore, to evaluate

the model, we visualized topic activations across test

days, comparing them with the semi-annotated data.

The evaluation of anomaly detection was divided

into two phases. Firstly, it was hypothesized that

the number of detected anomalies decreases with the

number of training days, reﬂecting a stabilization

of the learned routine. Secondly, it was evaluated

whether the expected anomalies were being detected.

To achieve this, LW1 was asked to annotate perceived

anomalies in their routine, allowing for the applica-

tion of quantitative measures on the annotated data.

However, it is important to note that users’ percep-

tion of an anomaly is often limited compared to the

model’s considerations. As a result, these annotations

may be biased. For this reason, the evaluation focuses

exclusively on the limited set of annotated anomalies,

calculating the true positive rate for assessment pur-

poses.

Finally, for online learning evaluation, a compar-

ison was made between retraining the model from

scratch and applying incremental updates over dif-

ferent training days, using the number of detected

anomalies as a proxy for performance.

In summary, the performance evaluation com-

prised the following steps to address the previously

Routine Pattern Learning and Anomaly Detection Applied to Lone Workers Through Topic Modeling

939

Normal day Anomalous day

Standard Deviation

(a) Topic 0. (b) Topic 1.

Figure 4: Comparison of activated topics on a normal and an anomalous day, overlaid with the normality range obtained from

training days.

identiﬁed research questions:

1. Visual analysis of different test days compared to

the learned routine (RQ1).

2. Evaluation of the number of detected anomalies

as a function of the number of training days used

for routine learning (RQ2).

3. Comparison of the percentage of anomalies cor-

rectly detected (true positive rate) across the dif-

ferent tested models (RQ2).

4. Comparison of model updating strategies using

the number of detected anomalies as a proxy for

performance evaluation (RQ3).

4.1 Routine Learning: Visual Analysis

LW1 semi-annotated one typical and one anomalous

day from their routine to assess how well the model

learned the routine. Figure 3 illustrates the activa-

tion of topics throughout the workday, along with

semi-annotated labels for both days. On the normal

day (Figure 3a), it is observed that the topics acti-

vated by the model closely correspond to the annota-

tions shown on top. For example, the periods labeled

Home, Commuting and Working display a consistent

topic conﬁguration within their respective pairs. This

alignment suggests that the model effectively extracts

meaningful information from the data, organizing it

into logical categories. On the anomalous day (Fig-

ure 3b), the topics do not align with the annotated

activities. While both Home periods display a con-

sistent topic conﬁguration, all other periods exhibit

topics indicative of anomalous behavior. Throughout

most of the day, changes in annotated activities are

indistinguishable based on the activated topics, sug-

gesting that these activities are not recognized due to

their absence in the training data.

Another visualization uses the proposed approach

for anomaly detection. It is assumed that topics acti-

vated during the testing days should fall within each

topic’s mean ± standard deviation (referred to as the

normality range), to be considered normal. Thus,

each activated topic on both the normal and anoma-

lous days is shown in Figure 4, along with the cor-

responding normality range obtained during training.

For most of the normal day, Topics 1, 2 and 3 fall

within the normality range while Topic 0 does not. It

is important to note that even on training days, not all

topic activations consistently fall within the normality

range, emphasizing the importance of the combina-

tion of topics in routine formation. In contrast, on the

anomalous day, all topics except Topic 3 fall outside

the normality range for most of the day, conﬁrming

its anomalous nature.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

940

Model Retrain

Model IL5 Model IL10

(a) LW1 (b) LW2 (c) LW3

Figure 5: Number of anomalies detected with the increased number of training days for the three lone workers and three tested

approaches: Model Retrain, IL5, and IL10.

4.2 Anomaly Detection and Model

Updates

To evaluate anomaly detection performance and ad-

dress research questions RQ2 and RQ3, the topic

modeling technique was tested in the context of model

updating using the following approaches:

1. Model Retrain. Topic model retrained each time

new data is added to the training set.

2. Model IL5. Topic model trained with IL, starting

from a base model of ﬁve days and updating each

new day individually.

3. Model IL10. Topic model trained with IL, start-

ing from a base model of 10 days and updating

each new day individually.

As the number of training days grows, it is ex-

pected that the number of detected anomalies de-

creases, indicating that the routine is being bet-

ter learned and that instances previously considered

anomalous are now recognized by the model. This hy-

pothesis is conﬁrmed in Figure 5. For all lone work-

ers’ routines, it is observed a decreasing trend in de-

tected anomalies across all models evaluated, with a

more pronounced drop during the ﬁrst 10 days. These

results suggest that ten consecutive days are sufﬁcient

for the model to learn a user’s routine. However, a

gradual decrease in detected anomalies is still evident

for LW1 and LW3, indicating that new data continues

to provide learnable information. In the case of LW2,

this decreasing pattern appears to plateau earlier, with

the Model Retrain approach even showing an increase

in detected anomalies toward the ﬁnal days. It is ob-

served a more stable pattern in IL models, whereas

Model Retrain shows more ﬂuctuation, with each new

day having a greater impact. Comparing the three ap-

proaches, no clear conclusions can be drawn from this

analysis, as the differences in anomalies detected be-

tween models do not appear to be signiﬁcant across

users. More information about routine data would be

needed to determine which model performs best.

Regarding the correctly detected anomalies evalu-

ation, a rough annotation of anomalies was performed

for the test set of LW1 to serve as ground truth. It was

established that for an anomaly to be correctly de-

tected, the model must identify at least one instance

within the annotated anomaly. In Figure 6, the true

positive rate of annotated anomalies was evaluated

across the different models. This analysis reveals that

Model Retrain tends to outperformed both IL5 and

IL10 models with more training days, but there is

not a model that consistently produces a higher true

positive rate. Additionally, there does not appear to

be a signiﬁcant difference between the performances

of Models IL5 and IL10, likely because the routine

learning stabilizes by day 10, as indicated in Figure 5.

5 CONCLUSIONS

In the context of lone workers, detecting anomalies

in their routines can help prevent risky situations and

promote preventive actions.

To discover the patterns that form the routines,

a topic modeling approach was implemented. This

method reveals hidden patterns, thereby identifying

relationships that, although coherent with annota-

tions, are not readily discernible to humans due to the

complexity of routines. Through visual inspection,

it was found that the LDA model identiﬁed routines,

and the derived topics accurately represented the hu-

man activities performed. Moreover, it was concluded

that the routine does not correspond to a single topic,

but rather to the combination of all topics activated.

Routine Pattern Learning and Anomaly Detection Applied to Lone Workers Through Topic Modeling

941

11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

100

Number of training days

True Positive Rate (%)

Retrain IL5 IL10

Figure 6: Comparison of correctly detected anomalies across different days for all models on data from Lone Worker 1.

The patterns identiﬁed were subsequently used as

the ground truth for deﬁning the routine and detect-

ing any deviations from it. The anomaly detection

results were satisfactory, demonstrating effective per-

formance in detecting anomalies that had been anno-

tated. However, this evaluation is not straightforward,

as the annotated anomalies do not fully correspond to

the ground truth, due to their complex nature.

Finally, the routine’s continuous evolution was ad-

dressed using IL techniques. A comparison of the re-

sults between models updated with IL and those fully

retrained shows that the latter detected a higher num-

ber of anomalies when provided a higher number of

training days. However, further studies are necessary

to draw more deﬁnitive conclusions.

While the results presented are promising, there

are some limitations and future work to consider.

Firstly, parameters such as the number of topics need

further testing and optimization. Additionally, the sta-

tistical approach to anomaly detection does not con-

sider potential relationships between activated topics,

although the results indicate a strong inter-topic rela-

tionship. Thus, alternative anomaly detection meth-

ods should be explored, and more annotated data

should be collected to enhance the robustness and

generalizability of the methodologies. This would

also facilitate a more thorough comparison between

retraining and incremental learning.

ACKNOWLEDGMENTS

This work was supported by European funds through

the Recovery and Resilience Plan, project ”Center

for Responsible AI”, project number C645008882-

00000055.

REFERENCES

Abraham, B. L. and Nair, A. P. (2018). Anomalous topic

discovery based on topic modeling from document

cluster. International Research Journal of Engineer-

ing and Technology (IRJET), 5(2):966–972.

Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). La-

tent dirichlet allocation. J. Mach. Learn. Res.,

3(null):993–1022.

Google (2021). Google Activity Recognition API.

https://developers.google.com/location-context/

activity-recognition. [Accessed: November 4, 2024].

Huynh, T., Fritz, M., and Schiele, B. (2008). Discovery of

activity patterns using topic models. pages 10–19.

Seiter, J., Derungs, A., Schuster-Amft, C., Amft, O., and

oster, G. (2014). Activity routine discovery in stroke

rehabilitation patients without data annotation. Pro-

ceedings - REHAB 2014, pages 270–273.

Steinhauer, H., Helldin, T., Mathiason, G., and Karlsson,

A. (2019). Topic modeling for anomaly detection in

telecommunication networks. Journal of Ambient In-

telligence and Humanized Computing, 14.

Sun, L., Chen, X., He, Z., and Miranda-Moreno, L. (2021).

Routine pattern discovery and anomaly detection in

individual travel behavior. Networks and Spatial Eco-

nomics, 23.

Tang, T. (2021). Characterizing shared mobility operator

and user behavior using big data analytics and ma-

chine learning. Master’s thesis, University of Virginia,

School of Engineering and Applied Science.

Thornton, A., Meiners, B., and Poole, D. (2020). La-

tent dirichlet allocation (lda) for anomaly detection in

avionics networks. In 2020 AIAA/IEEE 39th Digital

Avionics Systems Conference (DASC), pages 1–5.

UK Ministry of Defence (2008). JSP 375 Volume 1: Man-

agement of Health and Safety in Defence: Arrange-

ments. Last updated December 7, 2023.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

942