Predicting Falls from Operational Data: Insights and Limitations of

Using a Non-Specialized Database

Julien R

aker

1 a

, Patrick Elfert

1 b

, Cletus Brauer

2 c

, Marco Eichelberg

1 d

Frerk M

uller-von Aschwege

1 e

and Andreas Hein

1 f

R&D Division Health, OFFIS - Institute for Information Technology, Oldenburg, Germany

Johanniter-Unfall-Hilfe e.V., Oldenburg, Germany

ﬁ

Keywords:

Fall Prediction, Machine Learning, Health Data, Elderly Care, Predictive Analytics.

Abstract:

Falls among the elderly are a signiﬁcant public health concern. This study investigates the feasibility of

predicting falls using an operational dataset from Johanniter-Unfall-Hilfe (JUH) home emergency call system,

which was not created under laboratory conditions for scientiﬁc purposes. An anonymized dataset containing

records from 160,281 participants in Germany was analyzed. Statistical analysis identiﬁed 104 out of 400

features signiﬁcantly associated with falls, though with weak correlations (Cramer’s V ranging from 0.006 to

0.071). A one-class Support Vector Machine (SVM) was employed due to the absence of explicit non-fall

cases, achieving a true positive rate of 55.10%. The lack of explicit non-fall data prevented evaluation of

speciﬁcity and overall accuracy. The study demonstrates the potential of using operational datasets for fall

prediction but highlights signiﬁcant limitations due to data quality issues, such as the lack of explicit fall

records, absence of non-fall cases, lack of temporal data, and missing values. Recommendations are made to

improve data collection practices to enhance the utility of such datasets for predictive modeling.

1 INTRODUCTION

The global demographic shift towards an aging

population presents signiﬁcant challenges for

healthcare systems (Nicholas and Smith, 2006).

With advances in medicine and public health,

people are living longer, but this increased longevity

often comes with a higher prevalence of chronic

conditions and age-related impairments. One of the

most signiﬁcant risks associated with aging is the

increased likelihood of falls (Comans et al., 2013).

Falls remain one of the leading cause of injury and

morbidity in individuals aged 65 and older, often

resulting in severe consequences such as fractures,

hospitalization, and loss of independence (Comans

et al., 2013; Pfortmueller et al., 2014).

Preventing falls is thus a critical component in

improving the quality of life for older adults and

https://orcid.org/0009-0005-5153-906X

https://orcid.org/0000-0002-9834-0702

https://orcid.org/0000-0002-2924-7934

https://orcid.org/0000-0002-8590-3318

https://orcid.org/0009-0001-4960-4097

https://orcid.org/0000-0001-8846-2282

reducing the economic burden on healthcare systems

(Becker and Rapp, 2011; Walther et al., 2008).

Traditional preventive strategies include physical

therapy paired with home modiﬁcations, Checking for

incompatibilities or unwanted interactions with the

current medication or managing fear of falling (Zeeh

et al., 2017; Becker and Rapp, 2011; Jansen et al.,

2021).

In recent years, machine learning (ML) including

artiﬁcial intelligence (AI) have emerged as promising

tools for enhancing preventive healthcare efforts

(Vuppalapati et al., 2019). Through predictive

analytics, AI models can analyze large datasets to

identify risk factors and early warning signs of

health events such as falls, potentially enabling timely

interventions (Chattu, 2021).

Numerous studies have explored the use of ML

for fall prediction, utilizing various datasets and

methodologies. Most of these studies have relied on

data collected from sensors, which can be broadly

categorized into wearable and non-wearable devices.

Wearable devices often include inertial measurement

units (IMUs) that track movement and detect

anomalies indicative of a fall (El-Bendary et al.,

2013; Usmani et al., 2021). Approximately two-

774

Räker, J., Elfert, P., Brauer, C., Eichelberg, M., Müller-von Aschwege, F. and Hein, A.

Predicting Falls from Operational Data: Insights and Limitations of Using a Non-Specialized Database.

DOI: 10.5220/0013298600003911

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 774-780

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

thirds of these studies initially collected data under

controlled laboratory conditions to ensure data quality

and reliability (Usmani et al., 2021). A notable study

using non-sensor data was conducted by Marschollek

et al., who aimed to predict whether a patient would

fall during an inpatient stay in a geriatric center based

on medical records (Marschollek et al., 2012). They

analyzed the medical records of over 5,000 patients

over a period of 1.5 years. Using the C4.5 decision

tree algorithm, trained on the binary attribute of fall

or no fall, the model achieved an accuracy of 66%,

with a sensitivity of 55.4% and a speciﬁcity of 67.1%.

While their dataset included results from various

geriatric assessments, which signiﬁcantly inﬂuenced

the outcomes, such assessments are cost-intensive and

not routinely performed, limiting the generalizability

of the model. In contrast, general patient-related data

such as age, gender, pre-existing conditions (such

as chronic diseases or walking impairments), and

medication are more commonly available and have

also been shown to inﬂuence the occurrence of falls

(Jansenberger, 2011). However, there is a lack of

studies that utilize such general data for fall prediction

without relying on specialized assessments or sensor

data. This gap highlights the need to explore whether

more readily available patient data can be leveraged to

predict fall risk, potentially making predictive models

more accessible and scalable in real-world settings.

In this study, we aimed to address this gap

by exploring the potential of using an existing,

non-research-speciﬁc database from the Johanniter-

Unfall-Hilfe (JUH) home emergency call system to

predict falls. The JUH system provides emergency

support to elderly or disabled individuals, allowing

them to call for help in case of incidents, many of

which involve falls (J

org L

ussem, Thomas M

ahnert,

Hubertus v. Puttkamer, 2021). The database was

originally designed for operational purpose, not for

research, and thus lacks explicit fall records and

comprehensive temporal data. The data includes

personal information such as age, medication and

previous illnesses, which were recorded once when

the home emergency call system was activated (more

details in section 2.1). Fall-related incidents were

inferred from emergency call records, where falls

were often cited as the reason for the call, introducing

several limitations in the dataset’s applicability for

predictive modeling.

We aim to assess whether a ML model,

speciﬁcally a one-class Support Vector Machine

(SVM), can be used to predict falls based on the

available data. The one-class SVM was selected

due to the dataset’s imbalanced nature, containing

only fall-related cases and no explicit records of

individuals who have not fallen. This exploratory

study seeks to evaluate the feasibility of leveraging

a general-purpose dataset for predictive healthcare

and to identify key limitations, particularly those

inherent to this type of dataset, that future studies

must address.

2 METHODS

2.1 Data Source

The data used in this study originates from the JUH

home emergency call system, a service designed to

assist elderly or disabled individuals by providing

immediate emergency support (J

org L

ussem, Thomas

ahnert, Hubertus v. Puttkamer, 2021). The dataset

spans from January 1, 2012, to October 16, 2020, and

includes anonymized records of participants located

in Lower Saxony and Bremen, Germany.

2.1.1 Data Characteristics

The initial dataset comprised records from 160,281

participants, with an average age of 82.73 years

(±9 years) and a predominance of female participants

(71.84%). The dataset included:

• Demographic Information: Age, gender, and

living conditions.

• Medical History: Chronic conditions, prior

illnesses, and medications.

• Emergency Call Records: Details of emergency

calls, including timestamps and free-text

comments describing the nature of the incident.

• Care Level: Classiﬁed according to the German

care insurance system, indicating the degree of

care required by the individual (Nadash et al.,

2018).

This is static data that was recorded once when

the patient joined the home emergency call system,

with the exception of the degree of care, which was

updated depending on the status.

2.1.2 Data Limitations

Fall incidents were not systematically documented

but had to be inferred from free-text comments in

the emergency call logs, which were recorded by

the operator during or after the call. Consequently,

explicit records were also unavailable for individuals

who did not experience falls, as the information

on falls was derived solely from emergency calls.

This introduces a challenge, as people may fall

Predicting Falls from Operational Data: Insights and Limitations of Using a Non-Specialized Database

775

without contacting emergency services, complicating

the creation of a reliable control group.

2.2 Data Preprocessing

The following preprocessing steps were undertaken to

prepare the data for analysis:

1. Extraction of Fall Incidents:

• A keyword search on the free-text comments

was applied in the emergency call records to

identify potential fall incidents. Keywords

included different terms of the word

”

fall“

in German(e.g.

”

gefallen“,

”

gest

urzt“,

”

hingefallen“).

• To increase accuracy, the text has been

converted to lower case and umlauts have been

removed.

2. Handling Missing Values:

• Due to the signiﬁcant proportion of missing

data, with only 33.57% of participants having

complete data available, individuals with

incomplete datasets were excluded from the

analysis.

3. Feature Selection:

• The Pearson chi-square test of independence

was used to identify features signiﬁcantly

associated with fall incidents (see section 2.3).

4. Encoding Categorical Variables:

• Categorical variables were converted into

numerical format using one-hot encoding to

make them suitable for ML algorithms.

5. Feature Scaling:

• Numerical features were standardized to ensure

that all features contribute equally to the model

training process.

6. Data Splitting:

• The dataset was split into training and testing

sets using a 70/30 ratio.

2.3 Statistical Analysis

To identify features signiﬁcantly associated with fall

incidents, a Pearson chi-square test was applied to

the dataset. This test evaluates the relationship

between categorical variables (e.g., age group,

chronic conditions) and the likelihood of a fall by

comparing observed frequencies of falls within one

year with expected frequencies under the assumption

of independence. The strength of these associations

was quantiﬁed using Cramer’s V, to provide insights

into which feature were most relevant to fall

prediction.

2.4 Predictive Modeling

Since the dataset contains only one class, the selection

of suitable models is limited. We used algorithms

from the ﬁeld of outlier or anomaly detection for

this purpose. Examples of this are one-class SVMs,

Isolation Forest or Local Outlier Factor.

Given that one-class SVMs are well-suited for

problems where the goal is to identify patterns

in datasets containing only positive instances (An

et al., 2015), they were chosen for this task. The

participants were divided into two groups. Those

who made one or more calls calls due to falls within

one year, and those who made no fall-related call

within that year. To ensure all features have values

in the same range, the data was scaled using the

StandardScaler by scikit-learn. Furthermore, the

dataset was split into training and test sets, with

the training exclusively conducted on the class of

participants who experienced falls, and the testing

carried out on all data.

2.4.1 Model Evaluation

Due to the absence of non-fall records, model

evaluation was conducted by focusing on the models

ability to correctly identify actual fall instances within

the test set. The following metrics were used to

evaluate model performance:

• True Positive Rate (Sensitivity): Proportion of

actual falls correctly identiﬁed by the model.

• False Negative Rate (Speciﬁcity): Proportion of

actual falls incorrectly classiﬁed as non-falls.

3 RESULTS

3.1 Descriptive Statistics

For the chi-square test, a total of 146,263 participants

were included. The participants had an average age

of 83.81 years (±8.14 years), with ages ranging from

53 to 110. The majority of participants were female

(71.75%), while 27.32% were male. 15,262 (10.43%)

of those participants are marked as fallen. For 89.57%

of participants, no information is available as to

whether they have fallen or not.

For the predictive model, only the participants

who are marked as fallen and did not have any missing

values were included into analysis. The dataset was

HEALTHINF 2025 - 18th International Conference on Health Informatics

776

thus reduced to a total of 6038 participants, with

an average age of 84.65 (±7.39 years), ranged from

53 to 105. Here too, women were more frequently

represented (63.23%) than men (36.40%).

3.2 Statistical Associations

The Pearson chi-square test was applied to identify

features signiﬁcantly associated with fall incidents.

Among the 400 features analyzed, 104 demonstrated

a statistically signiﬁcant association with falls

(p < 0.05).

The strength of these associations was quantiﬁed

using Cramer’s V, with values ranging from 0.006 to

0.071, indicating a very weak correlation. The 20

features with the highest correlation can be seen in

Table 1.

3.3 Model Performance

A one-class SVM was employed to predict fall

incidents within a year based on the signiﬁcant

features identiﬁed in the statistical analysis. The

OneClassSVM class from scikit-learn was used for

this task. A grid search was performed to identify

the best parameters. It was found that the default

settings with the sigmoid kernel performed best. The

following performance metrics were obtained:

• The model achieved a sensitivity of 55.10%,

meaning it correctly identiﬁed 55.10% of

individuals who experienced falls.

• The false negative rate was 44.90%, indicating

the proportion of fall incidents that the model

failed to predict.

• The model classiﬁed overall 45.39% of the test

participants as at risk of falling.

The confusion matrix (shown in Table 2)

summarizes the model’s performance.

4 DISCUSSION AND

CONCLUSION

This study investigated the feasibility of using a non-

research-speciﬁc, operational dataset from the JUH

home emergency call system to predict falls among

the elderly using a one-class SVM. The ﬁndings

highlight both the potential and the limitations of

leveraging such datasets for predictive healthcare

modeling.

4.1 Interpretation of Results

The one-class SVM achieved a true positive rate

(sensitivity) of 55.10%, correctly identifying over

half of the fall incidents in the test set. However,

a signiﬁcant proportion of actual falls (44.90%)

were not detected by the model. The statistical

analysis revealed that 104 out of 400 features

were signiﬁcantly associated with falls (p < 0.05),

although the strength of these associations was weak

(Cramer’s V ranging from 0.006 to 0.071). However,

it must be considered here that in the Chi-Square

test the group of participants who had fallen was

compared with the unclear group (not known whether

they had fallen or not). The strength of the correlation

can therefore only be viewed with caution.

Notably, several of the 20 features with the highest

Cramer’s V values correspond to known risk factors

for falls, such as walking disabilities, obesity and

parkinson (see Table 1). Only the back pain factor

showed no known connection to falls. Despite

the strong limitation of the data and thus the low

strength of the correlation, it was possible to identify

characteristics that are known to be associated with

falls.

4.2 Limitations of Using Operational

Data

Several inherent limitations of the JUH dataset

impacted the study’s outcomes:

• Fall incidents were not explicitly recorded

but inferred from free-text comments in the

emergency call logs. This method may have led

to underreporting, variations in terminology or

spelling errors.

• The dataset did not include explicit records

of individuals who did not experience falls,

preventing the creation of a control group.

This limitation restricted the modeling approach

to one-class classiﬁcation and impeded the

evaluation of the models speciﬁcity and overall

accuracy.

• When changes are made to the data, the

previous records are simply overwritten, with

no trace left of the alterations. This lack of

historical data prevents tracking whether and

how patient information has been updated over

time. Consequently, it becomes challenging to

incorporate temporal trends and monitor changes

in patients’ conditions. Temporal data, however,

are essential for identifying patterns that may

Predicting Falls from Operational Data: Insights and Limitations of Using a Non-Specialized Database

777

Table 1: The 20 characteristics with the highest Cramer’s V. For characteristics with a speciﬁed source, the correlation has

already been established in previous studies. For characteristics without a source, no correlation is known to date.

Feature Cramers‘V References

Walking disability 0.070856 (Bergland et al., 2003)

Degree of care 0.048084 (Palm, 2024)

Obesity 0.046093 (Fjeldstad et al., 2008)

Parkinson 0.042153 (Kerr et al., 2010)

Living condition 0.041385 (Walther et al., 2008)

Antihypertensive medication 0.040181 (Klein et al., 2013)

Hearing, seeing and walking impairment 0.039203 (Bergland et al., 2003)

Hearing aid 0.036770 (Bergland et al., 2003)

Back pain 0.031729 none

Hearing impairment 0.031387 (Bergland et al., 2003)

Arthrosis 0.030739 (Rodrigues et al., 2014)

Alcohol abuse 0.024808 (Lima et al., 2009)

Diabetes mellitus (insulin-dependent) 0.024274 (Vinik et al., 2017)

Unsteady blood pressure 0.024247 (Klein et al., 2013)

Diabetes mellitus 0.022882 (Vinik et al., 2017)

Gender 0.021822 (Stevens and Sogolow, 2005)

Hearing impaired 0.021373 (Bergland et al., 2003)

Insulin 0.020749 (Vinik et al., 2017)

Circulatory disorders 0.020547 (Jansen et al., 2016)

Apoplexy 0.020485 (Su et al., 2021)

Table 2: Confusion matrix of the one-class classiﬁcation

model.

Predicted

Fallen Unknown

Actual Fallen 0.5510 0.4490

Actual Unknown 0.4485 0.5514

signal an increased risk of falls and for enhancing

the accuracy of predictive models.

• A signiﬁcant proportion of records had missing

values in key features such as medical history

and medication. It was unclear whether missing

entries indicated the absence of conditions or a

lack of data entry. This uncertainty led to the

exclusion of most participants from the analysis

and potentially introduced bias.

These limitations underscore the challenges of

using operational data not originally designed for

research purpose. Data structure, completeness, and

quality signiﬁcantly inﬂuence the effectiveness of ML

models in healthcare applications.

Still, it makes sense to use operational data

because it is real data. Recording data under

laboratory conditions is time-consuming, costly and

does not necessarily reﬂect the data in reality. Using

existing data can therefore be effective.

4.3 Recommendations for Improved

Data Collection Practices

To enhance the utility of operational datasets for

predictive modeling, the following recommendations

are proposed:

• The introduction of a structured ﬁeld in the

emergency call database table in which the

operator can choose between the top n call

reasons. This change would improve the accuracy

and completeness of incident data and facilitate

more precise analyses. Additionally, large

language models (LLMs) could be employed

to classify free-text ﬁelds more effectively,

providing further insights in categorizing incident

reasons.

• Collect information on individuals who have

not experienced falls, possibly through periodic

questionnaires or assessments. Participants

could report falls or near-falls annually, along

with updates on health status and medication

changes. This data would enable the creation of

a control group and allow the use of traditional

binary classiﬁcation models. We are aware

that this approach is difﬁcult to implement and

entails signiﬁcant effort, requiring considerable

resources and long-term participant commitment.

• Implement mechanisms to retain historical

HEALTHINF 2025 - 18th International Conference on Health Informatics

778

records with timestamps when updates are made

to patients information. Maintaining a temporal

record would allow for the analysis of trends

and the inclusion of time-dependent features in

predictive models, potentially enhancing their

accuracy.

• Differentiate between the absence of a condition

and missing data by including explicit indicators.

For example, use a speciﬁc code to denote

”

known illness“ versus

”

data not provided“. This

clariﬁcation would improve data integrity and

allow for more accurate analyses.

Implementing these recommendations might

enhance the quality and research utility of

operational datasets, facilitating more effective

predictive modeling and contributing to improved

healthcare outcomes. However, we acknowledge that

implementation can be a challenge, as this involves

highly sensitive data (e.g. medication, previous

illnesses) that requires careful handling of the data.

A balanced solution must therefore be found that

meets both the need for comprehensive data and

the requirement to minimize data in order to protect

privacy.

4.4 Contribution to the Field and

Future Work

This study contributes to the ﬁeld of predictive

healthcare modeling by illustrating both the potential

and challenges of using non-specialized, operational

datasets for fall prediction among older adults. The

ﬁndings highlight the critical importance of data

quality, structure, and completeness in developing

effective ML models.

Future Work: In the future, collaboration

with data providers, such as JUH, to enhance

data collection practices can improve the quality

of operational datasets. Additionally, exploring

advanced modeling techniques, such as semi-

supervised learning, may improve the predictive

performance. Another promising area is the

integration of additional data sources, such as

electronic health records or sensor data from wearable

devices, to provide a more comprehensive view of

patient health and further enhance model accuracy.

Finally, conducting longitudinal studies that preserve

temporal data will enable a deeper analysis of how

changes over time correlate with fall risk, providing

valuable insights into long-term patient outcomes.

4.5 Conclusion

In conclusion, while operational datasets like the

JUH home emergency call records hold promise

for predictive healthcare applications, signiﬁcant

challenges remain due to data limitations. The

study demonstrates that, despite certain predictive

capability exists, the effectiveness of ML models is

heavily dependent on the quality and structure of the

underlying data.

Enhancing data collection and management

practices is essential to unlock the full potential of

such datasets. By implementing the recommended

improvements, organizations could transform

operational data into more valuable resources for

research, ultimately contributing to better healthcare

outcomes for the older population.

ACKNOWLEDGMENTS

This work was carried out as part of the project

”

LivingSmart - Wohnquartiere neu gedacht –

Service-gesteuert: lebensnah, integrativ, intelligent,

innovativ“ (Elfert et al., 2023) funded by the

German Federal Ministry of Education and Research

(reference: 02K17A052). We would like to thank the

JUH, especially Alexandra Kolozis, for providing the

database used in this research.

REFERENCES

An, W., Liang, M., and Liu, H. (2015). An improved

one-class support vector machine classiﬁer for outlier

detection. Proceedings of the institution of mechanical

engineers, part c: Journal of mechanical engineering

science, 229(3):580–588.

Becker, C. and Rapp, K. (2011). St

urze in der geriatrie. Der

Internist, 52(8):939–945.

Bergland, A., Jarnlo, G.-B., and Laake, K. (2003).

Predictors of falls in the elderly by location. Aging

clinical and experimental research, 15:43–50.

Chattu, V. K. (2021). A review of artiﬁcial intelligence,

big data, and blockchain technology applications in

medicine and global health. Big Data and Cognitive

Computing, 5(3):41.

Comans, T. A., Currin, M. L., Quinn, J., Tippett, V., Rogers,

A., and Haines, T. P. (2013). Problems with a great

idea: referral by prehospital emergency services to

a community-based falls-prevention service. Injury

prevention, 19(2):134–138.

El-Bendary, N., Tan, Q., Pivot, F. C., and Lam, A. (2013).

Fall detection and prevention for the elderly: A review

of trends and challenges. International Journal on

Smart Sensing and Intelligent Systems, 6(3).

Predicting Falls from Operational Data: Insights and Limitations of Using a Non-Specialized Database

779

Elfert, P., Eichelberg, M., and Hein, A. (2023).

LivingSmart: Intelligente Auswerte- und

Informationsmodule zur Notfallpr

adiktion und -

avention f

ur hochaltrige Personen, pages 21–36.

Springer Vieweg, hmd edition.

Fjeldstad, C., Fjeldstad, A. S., Acree, L. S., Nickel, K. J.,

and Gardner, A. W. (2008). The inﬂuence of obesity

on falls and quality of life. Dynamic Medicine, 7(1):1–

Jansen, C.-P., Gross, M., Kramer-Gmeiner, F.,

Blessing, U., Becker, C., and Schwenk, M.

(2021). Empfehlungspapier f

ur das k

orperliche

gruppentraining zur sturzpr

avention bei

alteren,

zu hause lebenden menschen. Zeitschrift f

Gerontologie und Geriatrie, pages 1–11.

Jansen, S., Bhangu, J., de Rooij, S., Daams, J., Kenny,

R. A., and van der Velde, N. (2016). The association

of cardiovascular disorders and falls: a systematic

review. Journal of the American Medical Directors

Association, 17(3):193–199.

Jansenberger, H. (2011). Sturzpr

avention in Therapie und

Training. Georg Thieme Verlag.

org L

ussem, Thomas M

ahnert, Hubertus v. Puttkamer

(2021). Aus liebe zum leben: Jahresbericht 2020 der

johanniter-unfall-hilfe e.v. https://assets.johanniter.

de/JUH/Bundesgeschaeftsstelle/Dokumente/

Jahresberichte/2020 Jahresbericht Johanniter.pdf.

(besucht am 26.07.2022).

Kerr, G. K., Worringham, C. J., Cole, M. H., Lacherez,

P. F., Wood, J. M., and Silburn, P. (2010). Predictors

of future falls in parkinson disease. Neurology,

75(2):116–124.

Klein, D., Nagel, G., Kleiner, A., Ulmer, H., Rehberger, B.,

Concin, H., and Rapp, K. (2013). Blood pressure and

falls in community-dwelling people aged 60 years and

older in the vhm&pp cohort. BMC geriatrics, 13:1–7.

Lima, M. C. P., Sim

ao, M. O., Oliveira, J. B. d.,

Cavariani, M. B., Tucci, A. M., and Kerr-Correa, F.

(2009). Alcohol use and falls among the elderly in

metropolitan s

ao paulo, brazil. Cadernos de saude

publica, 25:2603–2611.

Marschollek, M., G

overcin, M., Rust, S., Gietzelt,

M., Schulze, M., Wolf, K.-H., and Steinhagen-

Thiessen, E. (2012). Mining geriatric assessment

data for in-patient fall prediction models and high-risk

subgroups. BMC medical informatics and decision

making, 12(1):1–6.

Nadash, P., Doty, P., and von Schwanenﬂ

ugel, M.

(2018). The german long-term care insurance

program: evolution and recent developments. The

Gerontologist, 58(3):588–597.

Nicholas, P. K. and Smith, M. F. (2006). Demographic

challenges and health in germany. Population

Research and Policy Review, 25:479–487.

Palm, R. (2024). Inkontinenz und st

urze erschweren die

pﬂege zu hause. MMW-Fortschritte der Medizin,

166(3):40–41.

Pfortmueller, C. A., Kunz, M., Lindner, G., Zisakis, A.,

Puig, S., and Exadaktylos, A. K. (2014). Fall-related

emergency department admission: fall environment

and settings and related injury patterns in 6357

patients with special emphasis on the elderly. The

Scientiﬁc World Journal, 2014.

Rodrigues, I. G., Fraga, G. P., and Barros, M. B. d. A.

(2014). Falls among the elderly: risk factors in

a population-based study. Revista Brasileira de

Epidemiologia, 17(3):705–718.

Stevens, J. A. and Sogolow, E. D. (2005). Gender

differences for non-fatal unintentional fall related

injuries among older adults. Injury prevention,

11(2):115–119.

Su, Q., Gao, Y., Zhang, J., Tang, J., Song, M., Song, J.,

Mao, Y., and Pi, H. (2021). Prevalence of fear of

falling and its association with physical function and

fall history among senior citizens living in rural areas

of china. Frontiers in public health, 9.

Usmani, S., Saboor, A., Haris, M., Khan, M. A., and Park,

H. (2021). Latest research trends in fall detection

and prevention using machine learning: a systematic

review. Sensors, 21(15):5134.

Vinik, A. I., Camacho, P., Reddy, S., Valencia, W. M.,

Trence, D., Matsumoto, A. M., and Morley, J. E.

(2017). Aging, diabetes, and falls. Endocrine

practice, 23(9):1120–1142.

Vuppalapati, J. S., Kedari, S., Ilapakurti, A., Vuppalapati,

C., Vuppalapati, R., and Kedari, S. (2019). Machine

learning infused preventive healthcare for high-risk

outpatient elderly. In Intelligent Systems and

Applications: Proceedings of the 2018 Intelligent

Systems Conference (IntelliSys) Volume 2, pages 496–

511. Springer.

Walther, L., Nikolaus, T., Schaaf, H., and H

ormann, K.

(2008). Schwindel und st

urze im alter. HNO,

56(8):833–842.

Zeeh, J., Reinhardt, Y., and Heppner, H.-J. (2017). St

urze

im alter. MMW-Fortschritte der Medizin, 159(13):52–

58.

HEALTHINF 2025 - 18th International Conference on Health Informatics

780