Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection

Anna Sjörs Dahlman

1,2 a

, Stefan Candefjord

1 b

, Xuezhi Zeng

1 c

Bengt Arne Sjöqvist

1 d

and

Kaj Lindecrantz

1,3 e

Chalmers University of Technology, Department of Electrical Engineering, Gothenburg, Sweden

VTI - Swedish National Road and Transport Research Institute, Linköping, Sweden

KTH - Royal Institute of Technology, Department of Biomedical Engineering and Health Systems, Stockholm, Sweden

Keywords: Driver Monitoring, Health Monitoring, AI-Driven Decision Support, Traffic Safety, Driver Support System.

Abstract: A relatively large proportion of fatalities on our roads are due to sudden illness in drivers, with the majority

of these cases attributed to cardiovascular disease. Being able to detect and manage these sudden events could

save many lives. This paper consolidates results from a literature review and three small-scale studies that

investigated and developed the possibilities of detecting sudden driver illness by measuring physiological

signals from cardiac activity with unobtrusive sensors including single-lead Electrocardiogram (ECG),

consumer-grade pulse sensors, and research grade radar technology. In general, the experiments have shown

that there is potential for the evaluated technologies to help detect and quantify cardiac illness events, but

significant development is needed to implement the technologies in real-world driving. It is challenging to

succeed in detecting driver states with high accuracy based on measurements of cardiac activity alone due to

both individual variations in heart activity patterns and an environment that complicates measurements, and

additional data from other sensors is probably needed. Physiological monitoring of drivers is challenging due

to vehicle vibrations, the driver's movements and thick clothing. There is a need for further research and

development of unobtrusive measurement technologies to detect driver states.

1 INTRODUCTION

A non-negligible number of deaths that occur during

driving are caused by sudden illness of the driver

(Halinen & Jaussi, 1994; Sjögren et al., 1996; Tervo

et al., 2013). Therefore, striving towards the vision of

zero traffic fatalities, the health of the driver and any

sudden loss of ability to control the vehicle is an

important factor to address. In most cases, the

underlaying cause is related to cardiovascular disease

(CVD), but other sudden pathologies such as syncope

and diabetic reactions may also be of importance.

Investigations of real crash data have shown that

medical conditions are a contributing factor in crash

causation. In-depth at-scene investigation of 298 road

crashes in the Adelaide metropolitan area in which at

least one person was transported to hospital or fatally

injured as a result of injuries sustained in the crash

https://orcid.org/0000-0003-2530-4126

https://orcid.org/0000-0001-7942-2190

https://orcid.org/0000-0002-6606-0386

https://orcid.org/0000-0002-6564-737X

https://orcid.org/0000-0003-4853-7731

showed that a medical condition was the main causal

factor in 13% of the crashes investigated (Lindsay &

Baldock, 2008). Close to a third of these were

cardiac-related (Lindsay & Baldock, 2008).

Moreover, studies have reported that medical

conditions can be the direct cause for between 1.5 and

24.7% of all road fatalities (Halinen & Jaussi, 1994;

Sjögren et al., 1996; Tervo et al., 2013). Some of

these deaths are due to the disease itself and in other

cases the medical condition causes a crash where the

crash impact is the cause of death. Tervo et al. (2008)

reported that in fatal crashes caused by medical

conditions, 43% of the deaths were caused by the

crash impact and 57% were caused by the disease. An

in-depth study of road fatalities in older drivers in

Sweden showed that 29% of all road fatalities in

drivers 50 years or older were attributable to acute

disease and in 9% of acute disease-triggered crashes,

Dahlman, A. S., Candefjord, S., Zeng, X., Sjöqvist, B. A. and Lindecrantz, K.

Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection.

DOI: 10.5220/0013400400003911

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 1117-1126

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

1117

another road user was injured (Skyving et al., 2023).

Almost all disease related deaths (97.5%) in this study

were caused by a cardiovascular event. In most

studies, cardiovascular related conditions are the

dominating medical conditions related to road

fatalities (Halinen & Jaussi, 1994; Lindsay &

Baldock, 2008; Tervo et al., 2008). Other medical

conditions include aortic rupture, cerebral circulatory

conditions and epilepsy. Studies have shown that less

than half of the drivers can stop the car on their own

in this situation (Büttner et al., 1999; Tervo et al.,

2008).

By continuously measuring the driver's vital signs

while driving, it would be possible to take necessary

safety measures in case a sudden deterioration is

detected. Survival from a cardiovascular event is in

some cases possible with fast, appropriate medical

care. Being able to monitor the onset of sudden

cardiovascular disease (CVD) in drivers could thus

save lives by providing valuable information about

the condition of the driver to first responders, helping

them to prioritize and allocate resources. Driver

assistance systems or automatic stop manoeuvres

triggered by detection of driver incapacitation could

also limit the effects of a disease episode by

preventing collisions with other road users. Various

types of driver monitoring systems already exist in

new vehicles, and organizations like the European

New Car Assessment Programme (Euro NCAP)

emphasize that the next generation of driver

monitoring systems should be able to detect and

manage sudden sickness in drivers (Fredriksson et al.,

2021). However, only a limited number of studies

have examined the detection of critical medical

conditions in vehicles to date.

Clinically, cardiovascular conditions such as

myocardial infarction and cardiac arrest are detected

by 12-lead Electrocardiogram (ECG) (Lee & Kim,

2023). From a twelve-lead ECG, where the electrodes

are placed at standardized positions, very

comprehensive diagnoses of cardiac infarction as

well as diagnoses of other cardiopathies can be

deduced. The morphology of the ECG is then an

important feature of the signal. Ischemia is typically

most clearly reflected in changes in ST-segment and

the T-wave. Depending on the localization of the

infarction in the cardiac muscle, the signs are more or

less visible in different leads, thus the need for several

leads for a good detection of local

ischemia/infarction. Cardiac arrythmias are seen as

irregular R-R-intervals and/or as lost synchronization

between P-wave and R-wave, and they are important

clues to cardiac problems. Arrythmia detection

depends less on many leads but put other restrictions

on the ECG. Assuming an absence of cardiac

arrythmia, heart rate (HR), sometimes referred to as

pulse rate or just pulse, can be derived easily from a

good quality single ECG signal from an arbitrary

lead, but there are also several other ways of

recording HR and short term, beat-to-beat variations

in HR, often referred to as Heart Rate Variability

(HRV). As HR and HRV are influenced by a

multitude of factors they alone are poor indicators of

cardiac distress, but they can provide complementary

information to other signs.

Measuring 12-lead ECG is not feasible in vehicles

during everyday driving. Driver monitoring is

therefore dependent on measurements of vital signs

using technology that does not disturb or affect the

driver, for example via electrodes in the steering

wheel, sensors in the seatbelt or driver seat, or

elsewhere in the vehicle (Arakawa, 2021). There are

different types of sensors that are suitable for use in

vehicles, it can be via electrodes that measure single

lead ECG or HR (pulse) when in contact with the

skin, via light (photoplethysmography), via radar

sensors, cameras, or other technologies. Another

challenge is to be able to handle shorter interruptions

in the measurements and disturbances such as noise

or artifacts due to non-optimal measurement

conditions that may occur during driving.

The aim of this paper is to explore the possibility

of detecting signs of sudden cardiovascular disease in

drivers by measuring HR or ECG via sensors in the

vehicle. This paper consolidates results from a

literature review and three small-scale exploratory

studies performed at the Department of Electrical

Engineering at Chalmers University of Technology to

provide a comprehensive view of the feasibility of in-

vehicle cardiovascular event detection. The work

focused on various aspects of cardiovascular disease

(CVD) detection technologies, including how to

1) measure HR and/or ECG unobtrusively in vehicles

2) handle noise and disturbances in physiological

measurements taken in vehicle environments and

3) detect specific CVDs in single-lead ECG

measurements. The rationale being that HR

measurements were regarded as more likely to be

implemented for continuous monitoring in vehicles.

However, one-lead ECG, for instance via steering

wheel electrodes, might be possible to implement for

assessment of suspected CVD when an abnormal HR

has been detected. Results from these studies

combined with the review of what has already been

done and is currently being done in the field forms the

basis for a general assessment of feasibility and

direction for further research and development.

DEMS 2025 - Special Session on Design and Evaluation of Monitoring Systems

1118

2 SMALL-SCALE STUDIES

2.1 In-Vehicle HR Measurement Using

Commercial Wearables

This study tackled the problem of measuring vital

data reliably in vehicles (Andersson et al., 2024).

Vital data was collected in a vehicle during driving

using different types of measuring equipment

including ECG with gel electrodes (which was used

as the gold standard), ECG via chest strap, and HR

measurement with photoplethysmography (PPG).

The sensitivity to noise and disturbances was

compared between different technologies. The

measurements taken in the car were carried out under

different conditions; stationary, driving on asphalt,

and driving on gravel roads. On the gravel road,

measurements were carried out driving through

bumps and during hard braking. Reference

measurements were carried out in stationary vehicles

with and without the engine running.

The measurements were carried out with

commercially available equipment; a Movesense

chest strap with electrodes and an Inertial

measurement unit (IMU) integrated in the strap

(Movesense Oy, Vantaa, Finland), a wrist-worn Polar

Vantage M (Polar Electro Oy, Kempere, Finland),

which measures HR using PPG, and a Vitaport II

single-lead ECG which measures a lead-II ECG using

gel electrodes placed on the chest (Temec

Technologies, Heerlen, Netherlands).

The signals were visually inspected to identify

when noise, artifacts and other disturbances occurred.

Signal processing was then performed to reduce noise

and extract relevant information from the signals.

Basic filtering was done using bandpass, notch and

Savitzky-Golay filters. This was tested on signals

from all sensors to investigate how well different

types of noise can be handled. The Pan-Tompkins

algorithm was applied to calculate HR by detecting

the peaks of the QRS complex in ECG signals (Pan &

Tompkins, 1985). The method uses bandpass filtering

and derivation, followed by squaring and integrating

the signal.

Baseline drift and noise from the car's vibrations

were evident in the signals. Baseline drift was

reduced by signal processing with bandpass filters

with successful noise reduction in all cases. High

frequency noise and vibrations could be reduced by

Savitzky-Golay filtering with a reduced amplitude in

the signals after filtering. The R-peaks of the ECG

were easy to identify in the filtered signals, showing

that the noise is not problematic to reduce, and that

pulse identification can be easily done. Motion

artifacts were also present in all measurements. They

were difficult to filter out and problematic because

they can distort the signals and make it difficult to

correctly identify the HR and arrhythmias.

The chest strap and wrist worn PPG sensor gave

the same average HR but there were differences in

individual HR values. The wrist worn PPG sensor

showed greater variability and was consistently

deficient in measuring the HR correctly in shorter

time intervals. Thus, wrist worn PPG is less reliable

for detection of short-term abnormal HR.

2.2 Unobtrusive HR Measurement

Using Radar Technology

The aim was to investigate the potential for detecting

signs of sudden illness in the form of arrhythmia

using radar technology by measuring HR (Björkman

et al., 2022). This included determining the

appropriate position of the radar, developing the

signal processing for both high and low HRs to be

detected, and investigating whether it was possible to

measure HR when the subject was wearing a jacket

and performed movements that mimic driving.

The radar sensor was the AWR1642BOOST

manufactured by Texas Instruments. The

AWR1642BOOST is a Frequency Modulated

Continuous Wave (FMCW) radar that operates at 77

GHz and supports a bandwidth of 4 GHz. Three

different radar sensor positions were examined: in

front of the chest, next to the left side of the chest and

behind the left side of the back (Figure 1). For each

position, measurements were performed at normal

HR and at high HR (> 100 beats per minute) created

by physical exertion. Measurements with normal HR

and high HR were conducted under stationary and

moving conditions (steering wheel movements). In

addition, two different jackets were used, a thinner

and a thicker jacket.

To study how accurately the radar could detect the

heart rhythm, a lead-II ECG measured with gel

electrodes on the chest was used as a reference

(PLUX Biosignals, Lisbon, Portugal). Signal

processing was then used to extract information about

the HR from the radar signal.

HR detection had the highest accuracy when the

radar sensor was placed behind the left side of the

back. In a car, this corresponds to the radar being

implemented in the driver’s backrest. Regular HR in

the range of 55-140BPM could be detected with an

overall accuracy of 91% when the radar was placed

behind the back and the subjects sat still without a

jacket.

Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection

1119

Figure 1: Experimental setup with the radar sensor placed

behind the participant.

The result also showed that it is possible to use the

FMCW radar to measure the HR even though the

subject is wearing a jacket. However, the results

indicate that different materials can affect the

accuracy. The results from measurements with

steering wheel movements showed that the radar

follows the HR from the ECG poorly when automatic

filtering is used. With manual filtering it was possible

to achieve a higher accuracy of 82%.

2.3 CVD Detection Based on

Single-Lead ECG

The study by Widengård et al. (2024) focused on

developing algorithms to detect CVD using databases

of physiological data (ECG) with signals from patients

affected by cardiovascular disease. With recent

advances in machine learning, many studies have

shown that computerized algorithms can perform

ECG-based detection of CVD with high accuracy

(Ansari et al., 2023; Feng et al., 2019; Gibson et al.,

2022; Kora, 2017; Liu et al., 2018; Martin et al., 2021).

In this study, existing algorithms were further

developed to work with single-lead ECG, which is

more realistic to implement in vehicles.

The ECG data used in this work were taken from

the database "Physikalisch-Technische Bundesanstalt

(PTB) Diagnostic ECG Database", abbreviated PTB

(Bousseljot et al., 1995), as well as "PTB-XL, a large

publicly available electrocardiography dataset",

abbreviated PTB-XL (Wagner et al., 2020) both via

PhysioNet (Goldberger et al., 2000). These open

datasets contain ECG data from healthy subjects as

well as subjects with various cardiovascular diseases.

The recordings were divided into training, validation

and test data. To avoid data leakage, the division was

made so data from the same patient did not end up in

both training and test data.

Two different types of deep learning models were

tested: Convolutional Neural Network (CNN), and

Recurrent Neural Network (RNN). Several versions

of the algorithms were developed with the starting

point in the structure described in the report by Liu et

al. (2018). To validate the algorithms, four different

measures were used; accuracy, sensitivity,

specificity, and F1 measures. In addition, the

generalizability of the best performing algorithm was

evaluated by training it on the PTB-XL dataset and

then evaluating with test data from another dataset,

the PTB dataset.

CNN Version 2 showed the best results of the

algorithms developed (Table 1). The algorithm also

had good performance when trained with one dataset

and evaluating with another. This suggests that the

algorithm is robust against different types of datasets,

and not overfitted or subject to data leakage.

The results for CNN version 2 are relatively

similar to the results of previous studies that used

neural networks for heart attack detection (Feng et al.,

2019; Gibson et al., 2022; Liu et al., 2018). All

algorithms were tested with single-lead data from

ECG lead I, but the algorithms were also tested with

data from ECG lead II, with no significant difference

in performance. The position of the myocardial

infarction determines the propagation of the changes

in the ECG signal, which allows localization of the

myocardial infarction (Morris & Brady, 2002). This

also means that some heart attacks do not cause

changes in all ECG leads. By only using one lead

there is a risk of missing out on information that

would have been visible in another lead.

3 DISCUSSION AND

STATE-OF-THE-ART

Driver monitoring systems for the detection of driver

Table 1: Results from the algorithm evaluation.

Algorithm

Accuracy

Sensitivity

Specificity

CNN Version 1 (training and evaluation using PTB)

0,839

0,80

0,86

0,845

CNN Version 2 (training and evaluation using PTB-XL)

0,92

0,95

0,90

0,93

CNN Version 2 (training using PTB-XL, evaluation using PTB)

0,83

0,97

0,76

0,81

RNN

0,49

0,50

0,42

0,668

DEMS 2025 - Special Session on Design and Evaluation of Monitoring Systems

1120

incapacitation require the ability to unobtrusively

monitor the driver's condition. Systems for the

detection of sudden cardiovascular disease in drivers

are based on monitoring vital signs linked to the

activity of the heart, mainly ECG and pulse. Driver

health monitoring systems are thus dependent on the

ability to measure these vital parameters reliably in

vehicles. Any available technique will be a

compromise of many aspects such as usability,

sensitivity, specificity, latency of detection, etc.

Being able to detect a specific condition at a specific

time is one of the biggest challenges with driver

monitoring in general. From the vehicle

manufacturers' perspective, it is not always the

detection of a specific condition that is the most

important, but the ability to handle an incapacitated

driver, regardless of the cause. New vehicles are

already equipped with many sensors, such as

cameras, radar for passenger detection, etc. It can

therefore be a challenge to bring in additional sensors,

but also an opportunity to use sensors that already

exist.

The results from the small-scale studies show that

there is potential for continuous monitoring of driver

vital signs for CVD detection in the future, but

substantial further development of in-vehicle HR and

ECG monitoring technologies is needed. Andersson

et al. (2024) showed that commercially available

wearable sensors can be used to monitor mean HR in

real-world driving. However, wrist worn PPG was not

able to detect sudden changes in HR and may thus be

less suitable for arrhythmia detection. Furthermore, in

all measurement technologies, noise and disturbances

that do not occur periodically and that arise from

sudden events were difficult to handle. Björkman et

al. (2022) showed that it is possible to measure HR in

a wide frequency range with radar technology in a

laboratory setting, which suggests that there is good

potential to detect arrhythmia. However, there are still

several challenges with signal processing before radar

can be implemented in a vehicle. Among other things,

it is necessary to develop the method further so that it

becomes more robust and can measure continuously.

Widengård et al. (2024) showed that if a single-lead

ECG of good signal quality is available, it is possible

to detect CVD (myocardial infarction) with high

accuracy in ECG data from lead I with a machine

learning algorithm. Furthermore, a CNN was

considered the most appropriate machine learning

algorithm for this purpose.

In a controlled environment and without

restrictions on how to apply sensors, there are

excellent methods for rapid detection of

cardiovascular disease states. ECG recordings offer

the possibility to detect CVD based on the

morphology of the signal, whereas HR sensors limit

the possibilities for CVD detection to detection of

abnormalities in the heart rhythm. Recent studies

show that it is possible to detect myocardial infarction

with single-lead ECG (Gibson et al., 2022; Hannun et

al., 2019). Abnormalities in various vital parameters

are common hours before a heart attack and clear

deteriorations can be found before cardiac arrest

(Andersen et al., 2016; Churpek et al., 2012; Kang et

al., 2016; Oh et al., 2016). With access to HR only or

lower quality data, there are still opportunities to

detect cardiovascular problems, but likely with lower

sensitivity and specificity. Additional difficulties in

the vehicular environment are the fact that in real-

world driving, the measurement must be carried out

in an unobtrusive manner. The collected signals are

often plagued with various types of disturbance and

noise that make further analysis very difficult. For

example, motion artifacts are of great importance for

the quality of the signals (Kawasaki & Kajiwara,

2023) and the traffic environment can influence

signal quality (Leicht et al., 2022). Wearable sensors

often provide more robust measurements of HR than

remote sensing in a vehicle environment (Arakawa,

2021; Leicht et al., 2022), although with the

disadvantage that it requires the driver to put on the

equipment.

Leonhardt et al. (2018) published an extensive

review of different methods for unobtrusive vital sign

monitoring in an automotive environment. The most

explored monitoring technologies were ECG

(conductive, capacitive or hybrid),

ballistocardiography, optical methods (PPG, PPG

imaging, far infrared imaging, other camera-based

methods), magnetic induction, and radar-based

methods. Some of the techniques have been more

extensively studied than others and the review

provides a thorough comparison between methods. A

general conclusion was that all methods were fragile

and sensitive to disturbances. Relatively few of the

methods have been tested in environments close to

realistic conditions. Most early tests were in

laboratory settings and without the influence from

movement.

Previous studies have shown that it is possible to

detect HR with remote sensing, including radar

technology (Qiao et al., 2022; Schires et al., 2018;

Tang et al., 2017). However, few studies have used

radar technology to investigate the possibility of

detecting CVD. Radar-based systems can be used to

detect both breathing rate and HR, but they tend to be

very sensitive to both body and vehicle motion. Tang

et al. (2017) proposed a method using two radar

Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection

1121

sensors, one in front of the person and one at the back.

Utilizing the fact that movements of the entire body

tend to give an out-of-phase signal whereas

movements of heart and lung give an in-phase signal.

In a review article by Arakawa (2021), different

approaches to measuring drivers' HR were analyzed.

A steering wheel sensor with electrode measurement

and steering wheel sensor with LED sensor managed

to measure HR accurately when both hands were on

the wheel. Steering wheel electrodes have the

potential to be used for single-lead ECG measurement

but with the disadvantage that the hands must always

be kept on the steering wheel. Another method was a

car seat with a 24 GHz doppler sensor and

accelerometer. This car seat was able to measure HR

accurately within 5-10 beats per minute depending on

the type of driving. Arakawa (2021) found that

wearables such as rings and watches could measure

HR with good reliability with the disadvantage that

the driver have to wear them. Finally, it was reported

that several studies used video cameras to measure

HR by measuring color changes in people's faces.

However, this method has not been tested in vehicles.

Arakawa (2021) also states that pulse measurements

alone today are not enough to assess the driver's state

of health.

Ultimately, the goal of the sensors is to provide

signals that can be processed into information

pertinent to the detection of driver health status.

Signal processing including preprocessing,

parametrization/information enhancement, and

finally information extraction/disease detection is

key. Preprocessing is primarily necessary to increase

the signal-to-noise ratio (SNR). Low SNR is clearly

one key obstacle in all the reviewed methods. Few of

the sensors described in the literature have been tested

in a car during driving; and those that have been

subjected to realistic conditions have demonstrated

unsatisfactory performance. Parametrisation is based

on the notion that some specific aspect or aspects of a

signal carries relevant information. For instance, HR,

HRV or ECG morphology as potential indicators of

cardiac problems. The intended parametrization may

put requirements and limitations on the

preprocessing. If the sought parameter is HR only a

narrow band filtering may be a good component in

the preprocessing, but such filtering may extinguish

all possibilities to capture ECG morphology.

Maximizing the SNR in the sense signal power

divided with noise power is counterproductive as the

maximum power of the signal does not necessarily

coincide with power of the morphology alterations.

All good signal processing is therefore based on a

clear idea of the information bearing part of the signal

and knowledge of the interfering signal components.

A powerful method to reduce noise from the

signal is adaptive filtering which is based on the idea

of placing a second transducer, a reference

transducer, in close proximity to the original, signal

transducer (Eilebrecht et al., 2012). The purpose of

the reference sensor is to get a signal that mainly

registers noise, and no signal. Ideally, the difference

between the signals from the two sensors would give

a clean signal. In reality, an appropriate filter will be

inserted after the reference transducer; whose task is

to adjust the amplitude and phase of the reference

signal to mimic the noise from the signal transducer.

The filter is adaptively adjusted to changes in

transmission characteristics during recording

(Widrow et al., 1975).

Sensor fusion is a potential method that can be

applied to deal with poor signal quality. Future

research may benefit from including several vital

parameters to obtain a more complete understanding

of the physiological processes. Combining more vital

parameters can create a stronger basis for identifying

CVD. For instance, if abnormal vital signs are

detected the driver could be asked to put both hands

on the steering wheel to enable a single-lead ECG

measurement for more detailed health assessment.

Another possibility is to connect mobile devices to

the vehicle, in this way measurements from smart

watches or other wearables can be used in the

assessment. The focus of this report is on

cardiovascular events, but it is worth noting that both

epileptic seizures and strokes can affect HR and HRV

(Chen et al., 2014; Nei, 2009; van Elmpt et al., 2006;

Zangróniz et al., 2017; Zijlmans et al., 2002). This

opens for the possibility that also some cases of

severe cerebral attacks could be detected via heart

signals.

The way forward is therefore to use physiological

signals that contain as much information as possible

and that can realistically be acquired routinely during

driving. The diagnostic qualities of any single one of

these signals will likely be too low, and in periods the

signal-to-noise will be very poor. However,

combining several signals from different sources may

compensate for the shortcomings of individual

technologies. Suggested next steps are to test and

evaluate sensor tools allowing detection of

morphological changes in single lead ECG, HR

recording, analysis of HRV, and recording of

respiration rate. In parallel, the types of interference

that affect the different sensors should be studied to

look for means of acquiring a good reference signal,

allowing adaptive noise cancelling. When the sensor

DEMS 2025 - Special Session on Design and Evaluation of Monitoring Systems

1122

data acquisition is optimized for each individual

sensor, sensor fusion algorithms can be developed

that extract the physiologically relevant parameters.

Lastly, the decision system that can provide the final

decision whether the driver is in full health or actions

are required needs to be developed.

A major challenge for the development and

evaluation of sensors and algorithms is that sudden

illness events while driving are rare and they cannot

be manipulated experimentally for validation of the

systems as is done, for example, when testing fatigue

warning systems. Early development must rely on

data measured under other conditions or simulated

data. An important limitation is that the studies by

Andersson et al. (2024) and Björkman et al. (2022)

use only measurements from healthy individuals,

which limits the possibility to generalize to real

disease situations. Future studies should include data

from people diagnosed with arrhythmia to assess

whether these conditions can be detected in the

signals. Another challenge is to minimize the number

of false positives as well as false negatives. Vehicle

support systems such as automatic stop maneuvers or

automatic alarms should not be activated unless it is

a real emergency. This underscores the need to

carefully consider the pros and cons of different

detection methods to ensure proper interpretation and

management of CVD events.

The traffic safety and societal benefit of being

able to detect sudden CVD onset in drivers is mainly

linked to two scenarios. The first is to integrate driver

monitoring with support systems into the vehicle so

that the vehicle can take control and perform a safe

stop. This would reduce the risk of collisions with

other road users and the risk of single-vehicle

accidents. The second aspect is about being able to

inform emergency responders about the driver’s

condition to allocate appropriate resources to the

scene. A recent project called TEAPaN (Traffic Event

Assessment, Prioritizing and Notification) explored

how such traffic incident information could be used

for prioritization of emergency resources

(Söderholm, 2023). The project designed and

explored an IT-infrastructure and incident alert

handling solution directly interfacing with

prehospital care resources. Detecting a medical

condition before the driver is completely

incapacitated can provide the opportunity to handle

the medical emergency at an early stage. Early

detection would also increase the possibility of

survival if the right care can be given more quickly.

Both actions, stopping the vehicle and calling for

emergency care, can have far-reaching consequences.

It is not possible to achieve 100% sensitivity and

100% specificity for this type of detection. False

CVD detection can potentially lead to the vehicle

making unnecessary and dangerous maneuvers

and/or sending false alarms to dispatch centers. False

alarms are costly and can mean that resources that

could have saved lives elsewhere are blocked.

In the study by Widengård et al. (2024), the

recordings in the PTB and PTB-XL databases were

often taken several days after the onset of CVD,

which is important to take into account when

analyzing and drawing conclusions about the

performance of the algorithm. Moreover, the datasets

were recorded in a clinical environment, i.e. not in

vehicle environment which is the intended

environment for the practical implementation of the

algorithm. For further development of a machine

learning algorithm with the aim of detecting acute

CVD, a large amount of new data is required.

Optimally, this data would contain real-time infarcts

and be collected in a vehicle environment where

characteristic noise and motion artifacts occur.

However, such data collection is very difficult to

implement.

Privacy and regulatory issues are other

important areas to consider when handling health

data. Health monitoring would involve the storage of

sensitive personal data. Therefore, all data will

probably need to be processed locally in the vehicle

to reduce the risk of exposing (sensitive) personal

data. The ability to turn off or deny monitoring can

also be a solution to privacy issues. Diagnostic tools

are regarded as medical devices and the regulations

surrounding medical devices are extensive. If a

detection system in a vehicle makes a medical

decision, the device needs to be tested and approved

as a medical device. Whether a detection system that

warns of abnormal vital parameters in a driver is a

medical device has not been investigated here.

4 CONCLUSIONS

The results showed that there are significant

challenges in building a functional system to detect

sudden cardiovascular disease in drivers, and

substantial development work remains. Provided

there is a high-quality ECG signal, the prospects for

detecting heart attacks are promising, and a high-

quality pulse signal enables the detection of

arrhythmias in drivers. Further development work is

required to ensure sufficiently high sensitivity and

specificity in real cases of illness. However, the

biggest obstacle is that reliable unobtrusive

technologies for measuring ECG morphology and HR

Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection

1123

in the vehicle environment are not yet available.

Arrhythmias can be detected using wearables such as

chest straps whereas sensors not worn by the driver

but integrated into the vehicle are sensitive to noise

and interference. To enable the detection of sudden

illnesses in vehicles, more vital parameters need to be

examined, and multiple measurement systems need to

be integrated to provide sufficient and reliable data.

The results indicate that there is a need for further

research and development of unobtrusive

measurement methods to detect driver states.

ACKNOWLEDGEMENTS

This work was funded by grants from the strategic

vehicle research and innovation (FFI) program at

Sweden’s Innovation Agency (VINNOVA), grant

number 2020-05157 and the Swedish Road

Administration (Skyltfonden), grant number

TRV2023/28021. The authors would like to

acknowledge the work by bachelor students Julia

Björkman, Zakaria Hersi, Abdinaser Muse, Krister

Mattsson, Anton Widengård, David Ruin, Emmy

Alvius, Lukas Pettersson, Lukas Wallén, Molly

Lundqvist, Joel Andersson, Petter Enlund, Ebba

Fredlund, Emma Hedberg, Love Stoopendahl, and

Stina Ström who performed the small-scale studies.

REFERENCES

Andersen, L. W., Kim, W. Y., Chase, M., Berg, K.-M.,

Mortensen, S. J., Moskowitz, A., Novack, V., Cocchi,

M. N., & Donnino, M. W. (2016). The prevalence and

significance of abnormal vital signs prior to in-hospital

cardiac arrest. Resuscitation, 98, 112-117.

https://doi.org/https://doi.org/10.1016/j.resuscitation.2

015.08.016

Andersson, J., Enlund, P., Fredlund, E., Emma Hedberg,

Stoopendahl, L., & Ström, S. (2024). EKG- och

pulsmätning i fordon. Analys av brus och störningar i

fordon för att detektera kardiovaskulära sjukdomar med

EKG- och pulsmätning. Chalmers Tekniska Högskola].

Göteborg.

Ansari, Y., Mourad, O., Qaraqe, K., & Serpedin, E. (2023).

Deep learning for ECG Arrhythmia detection and

classification: an overview of progress for period 2017–

2023 [Systematic Review]. Frontiers in Physiology, 14.

https://doi.org/10.3389/fphys.2023.1246746

Arakawa, T. (2021). A Review of Heartbeat Detection

Systems for Automotive Applications. Sensors, 21(18),

6112. https://www.mdpi.com/1424-8220/21/18/6112

Björkman, J., Hersi, Z., Muse, A., & Mattsson, K. (2022).

Detektering av kännetecken på arytmi med en

millimetervågsradar Chalmers Tekniska Högskola].

Göteborg. http://hdl.handle.net/20.500.12380/306079

Bousseljot, R., Kreiseler, D., & Schnabel, A. (1995).

Nutzung der EKG-Signaldatenbank CARDIODAT der

PTB über das Internet. Biomedical Engineering /

Biomedizinische Technik, 40(s1), 317-318.

https://doi.org/doi:10.1515/bmte.1995.40.s1.317

Büttner, A., Heimpel, M., & Eisenmenger, W. (1999).

Sudden natural death ‘at the wheel’: a retrospective

study over a 15-year time period (1982–1996). Forensic

Science International, 103(2), 101-112.

https://doi.org/10.1016/s0379-0738(99)00063-8

Chen, W., Guo, C.-L., Zhang, P.-S., Liu, C., Qiao, H.,

Zhang, J.-G., & Meng, F.-G. (2014). Heart rate changes

in partial seizures: analysis of influencing factors

among refractory patients. BMC Neurology, 14(1),

135. https://doi.org/10.1186/1471-2377-14-135

Churpek, M. M., Yuen, T. C., Park, S. Y., Meltzer, D. O.,

Hall, J. B., & Edelson, D. P. (2012). Derivation of a

cardiac arrest prediction model using ward vital signs*.

Critical Care Medicine, 40(7).

https://journals.lww.com/ccmjournal/fulltext/2012/070

00/derivation_of_a_cardiac_arrest_prediction_model.1

3.aspx

Eilebrecht, B., Wartzek, T., Willkomm, J., Schommartz, A.,

Walter, M., & Leonhardt, S. (2012). Motion Artifact

Removal from Capacitive ECG Measurements by

Means of Adaptive Filtering. In Á. Jobbágy, 5th

European Conference of the International Federation

for Medical and Biological Engineering Berlin,

Heidelberg.

Feng, K., Pi, X., Liu, H., & Sun, K. (2019). Myocardial

Infarction Classification Based on Convolutional

Neural Network and Recurrent Neural Network.

Applied Sciences, 9(9), 1879. https://www.mdpi.com/

2076-3417/9/9/1879

Fredriksson, R., Lenné, M. G., van Montfort, S., & Grover,

C. (2021). European NCAP Program Developments to

Address Driver Distraction, Drowsiness and Sudden

Sickness [Perspective]. Frontiers in Neuroergonomics,

2(33). https://doi.org/10.3389/fnrgo.2021.786674

Gibson, C. M., Mehta, S., Ceschim, M. R. S., Frauenfelder,

A., Vieira, D., Botelho, R., Fernandez, F., Villagran, C.,

Niklitschek, S., Matheus, C. I., Pinto, G., Vallenilla, I.,

Lopez, C., Acosta, M. I., Munguia, A., Fitzgerald, C.,

Mazzini, J., Pisana, L., & Quintero, S. (2022).

Evolution of single-lead ECG for STEMI detection

using a deep learning approach. International Journal of

Cardiology, 346, 47-52. https://doi.org/https://doi.org/

10.1016/j.ijcard.2021.11.039

Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff, J.

M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,

G. B., Peng, C. K., & Stanley, H. E. (2000).

PhysioBank, PhysioToolkit, and PhysioNet:

components of a new research resource for complex

physiologic signals. Circulation, 101(23), E215-220.

https://doi.org/10.1161/01.cir.101.23.e215

Halinen, M. O., & Jaussi, A. (1994). Fatal road accidents

caused by sudden death of the driver in Finland and

Vaud, Switzerland. European Heart Journal, 15(7), 888-

DEMS 2025 - Special Session on Design and Evaluation of Monitoring Systems

1124

894. https://doi.org/10.1093/oxfordjournals.

eurheartj.a060606

Hannun, A. Y., Rajpurkar, P., Haghpanahi, M., Tison, G.

H., Bourn, C., Turakhia, M. P., & Ng, A. Y. (2019).

Cardiologist-level arrhythmia detection and

classification in ambulatory electrocardiograms using a

deep neural network. Nature Medicine, 25(1), 65-69.

https://doi.org/10.1038/s41591-018-0268-3

Kang, M. A., Churpek, M. M., Zadravecz, F. J., Adhikari,

R., Twu, N. M., & Edelson, D. P. (2016). Real-Time

Risk Prediction on the Wards: A Feasibility Study.

Critical Care Medicine, 44(8), 1468-1473.

https://doi.org/10.1097/ccm.0000000000001716

Kawasaki, R., & Kajiwara, A. (2023, 22-25 Jan. 2023).

Driver's vital-signs monitoring with a single 60GHz

sensor. 2023 IEEE Radio and Wireless Symposium

(RWS),

Kora, P. (2017). ECG based Myocardial Infarction

detection using Hybrid Firefly Algorithm. Computer

Methods and Programs in Biomedicine, 152, 141-148.

https://doi.org/https://doi.org/10.1016/j.cmpb.2017.09.

015

Lee, C. H., & Kim, S. H. (2023). ECG Measurement

System for Vehicle Implementation and Heart Disease

Classification Using Machine Learning. IEEE Access,

11, 17968-17982. https://doi.org/10.1109/

ACCESS.2023.3245565

Leicht, L., Walter, M., Mathissen, M., Antink, C. H.,

Teichmann, D., & Leonhardt, S. (2022). Unobtrusive

Measurement of Physiological Features Under

Simulated and Real Driving Conditions. IEEE

Transactions on Intelligent Transportation Systems,

23(5), 4767-4777. https://doi.org/10.1109/

TITS.2022.3143004

Leonhardt, S., Leicht, L., & Teichmann, D. (2018).

Unobtrusive vital sign monitoring in automotive

environments—A review. Sensors, 18(9), 3080.

Lindsay, V., & Baldock, M. (2008, November). Medical

conditions as a contributing factor in crash causation.

Australasian Road Safety Research, Policing and

Education Conference,

Liu, N., Wang, L., Chang, Q., Xing, Y., & Zhou, X. (2018).

A Simple and Effective Method for Detecting

Myocardial Infarction Based on Deep Convolutional

Neural Network. Journal of Medical Imaging and

Health Informatics, 8(7), 1508-1512.

https://doi.org/10.1166/jmihi.2018.2463

Martin, H., Izquierdo, W., Cabrerizo, M., Cabrera, A., &

Adjouadi, M. (2021). Near real-time single-beat

myocardial infarction detection from single-lead

electrocardiogram using Long Short-Term Memory

Neural Network. Biomedical Signal Processing and

Control, 68, 102683. https://doi.org/https://doi.org/

10.1016/j.bspc.2021.102683

Morris, F., & Brady, W. J. (2002). ABC of clinical

electrocardiography: Acute myocardial infarction—

Part I. BMJ, 324(7341), 831-834.

https://doi.org/10.1136/bmj.324.7341.831

Nei, M. (2009). Cardiac effects of seizures. Epilepsy Curr,

9(4), 91-95. https://doi.org/10.1111/j.1535-

7511.2009.01303.x

Oh, H., Lee, K., & Seo, W. (2016). Temporal patterns of

change in vital signs and Cardiac Arrest Risk Triage

scores over the 48 hours preceding fatal in‐hospital

cardiac arrest. Journal of advanced nursing, 72(5),

1122-1133.

Pan, J., & Tompkins, W. J. (1985). A Real-Time QRS

Detection Algorithm. IEEE Transactions on

Biomedical Engineering, BME-32(3), 230-236.

https://doi.org/10.1109/TBME.1985.325532

Qiao, J. H., Qi, F. G., Liang, F. L., Ma, J., Lv, H., Yu, X.,

Xue, H. J., An, Q., Yan, K. D., Shi, D., Qiao, Y. H.,

Wang, J. Q., & Zhang, Y. (2022). Contactless

multiscale measurement of cardiac motion using

biomedical radar sensor. Front Cardiovasc Med, 9,

1057195. https://doi.org/10.3389/fcvm.2022.1057195

Schires, E., Georgiou, P., & Lande, T. S. (2018). Vital Sign

Monitoring Through the Back Using an UWB Impulse

Radar With Body Coupled Antennas. IEEE

Transactions on Biomedical Circuits and Systems,

12(2), 292-302. https://doi.org/10.1109/

TBCAS.2018.2799322

Sjögren, H., Eriksson, A., & Öström, M. (1996). Role of

disease in initiating the crashes of fatally injured

drivers. Accident Analysis & Prevention, 28(3), 307-

314. https://doi.org/https://doi.org/10.1016/0001-

4575(96)00062-0

Skyving, M., Möller, J., & Laflamme, L. (2023). What

triggers road traffic fatalities among older adult drivers?

An investigation based on the Swedish register for in-

depth studies of fatal crashes. Accident Analysis &

Prevention, 190, 107149. https://doi.org/https://doi.org/

10.1016/j.aap.2023.107149

Söderholm, H. M. (2023). TEAPaN: Traffic Event

Assessment, Prioritization and Notification.

https://www.vinnova.se/en/p/teapan-1-sossum---

traffic-event-assessment-prioritizing-and-notification-

steg1-simulations-and-demonstrators/

Tang, M. C., Wang, F. K., & Horng, T. S. (2017). Single

Self-Injection-Locked Radar With Two Antennas for

Monitoring Vital Signs With Large Body Movement

Cancellation. IEEE Transactions on Microwave Theory

and Techniques, 65(12), 5324-5333.

https://doi.org/10.1109/TMTT.2017.2768363

Tervo, T., Räty, E., Sulander, P., Holopainen, J. M.,

Jaakkola, T., & Parkkari, K. (2013). Sudden Death at

the Wheel Due to a Disease Attack. Traffic Injury

Prevention, 14(2), 138-144. https://doi.org/10.1080/

15389588.2012.695827

Tervo, T. M. T., Neira, W., Kivioja, A., Sulander, P.,

Parkkari, K., & Holopainen, J. M. (2008).

Observational Failures/Distraction and Disease

Attack/Incapacity as Cause(s) of Fatal Road Crashes in

Finland. Traffic Injury Prevention, 9(3), 211-216.

https://doi.org/10.1080/15389580802040303

van Elmpt, W. J., Nijsen, T. M., Griep, P. A., & Arends, J.

B. (2006). A model of heart rate changes to detect

Feasibility of Driver Monitoring for Sudden Cardiac Illness Detection

1125

seizures in severe epilepsy. Seizure, 15(6), 366-375.

https://doi.org/10.1016/j.seizure.2006.03.005

Wagner, P., Strodthoff, N., Bousseljot, R.-D., Kreiseler, D.,

Lunze, F. I., Samek, W., & Schaeffter, T. (2020). PTB-

XL, a large publicly available electrocardiography

dataset. Scientific Data, 7(1), 154.

https://doi.org/10.1038/s41597-020-0495-6

Widengård, A., Ruin, D., Alvius, E., Pettersson, L., Wallén,

L., & Lundqvist, M. (2024). Detektion av akut hjärt-

kärlsjukdom i fordon. Utveckling av

maskininlärningsalgoritm för detektion av akut hjärt-

kärlsjukdom hos förare med enkelavlednings-EKG.

(Publication Number EENX16-2024-37) Chalmers

Tekniska Högskola]. Göteborg.

Widrow, B., Glover, J. R., McCool, J. M., Kaunitz, J.,

Williams, C. S., Hearn, R. H., Zeidler, J. R., Dong, J.

E., & Goodlin, R. C. (1975). Adaptive noise cancelling:

Principles and applications. Proceedings of the IEEE,

63(12), 1692-1716. https://doi.org/10.1109/PROC.

1975.10036

Zangróniz, R., Martínez-Rodrigo, A., Pastor, J. M., López,

M. T., & Fernández-Caballero, A. (2017).

Electrodermal Activity Sensor for Classification of

Calm/Distress Condition. Sensors (Basel, Switzerland),

17(10). https://doi.org/10.3390/s17102324

Zijlmans, M., Flanagan, D., & Gotman, J. (2002). Heart rate

changes and ECG abnormalities during epileptic

seizures: prevalence and definition of an objective

clinical sign. Epilepsia, 43(8), 847-854.

https://doi.org/10.1046/j.1528-1157.2002.37801.x

DEMS 2025 - Special Session on Design and Evaluation of Monitoring Systems

1126