Heart failure (HF) is a complex clinical syndrome
characterized by the inability of the heart to maintain
adequate circulation to meet metabolic demands. Despite
advances in therapy, HF remains associated with high
morbidity, mortality, and healthcare costs, largely due to
delayed diagnosis and limited accessibility of definitive
imaging methods. Early identification, particularly in
primary care settings, remains a major unmet clinical need.
To address this challenge, we revisited the concept of
polycardiography using modern sensing technologies.
Polycardiography enables synchronous acquisition of
electrical and mechanical cardiovascular activity,
providing access to electromechanical coupling parameters.
We developed a prototype multimodal system capable of
simultaneous recording of electrocardiogram (ECG),
phonocardiogram (PCG), seismocardiogram (SCG), and
photoplethysmogram (PPG) signals. From these signals,
established HF-related biomarkers can be extracted,
including systolic time intervals (STIs) such as left
ventricular ejection time (LVET) and pre-ejection period
(PEP). Prior to clinical deployment, the SensSmartTech
validation study was conducted on healthy volunteers to
verify device performance and characterize the dependence
of key cardiovascular features on heart rate. This effort
resulted in the SensSmartTech database, publicly available
via PhysioNet, comprising multimodal cardiovascular
recordings across a wide heart rate range. The database
provides insight into physiological variability and
supports algorithm development and benchmarking.The central
focus of this work is the SensSmart clinical study,
conducted at the University Clinical Centre of Serbia. The
study has been completed and comprehensive data analysis is
underway. Its primary objective is to evaluate the
non-inferiority of HF detection based on multimodal
polycardiography relative to echocardiography. Preliminary
results obtained using AI-based classification demonstrate
promising performance in discriminating HF patients from
controls, while also indicating the complementary
diagnostic value of combining multiple sensing modalities.
At the conference, we will present interim outcomes of the
clinical analysis, including classification performance and
an assessment of the relative contribution of individual
multimodal features to HF detection.

Electrocardiographic (ECG) signals are often corrupted by
noise such as baseline wander, electromyographic noise, and
power-line interference, which can degrade clinical
interpretation. Therefore, effective filtering methods are
essential to extract clinically relevant information from
noise-contaminated signals, and wavelet-based methods are
among the most commonly used approaches for denoising.
However, the denoising performance of traditional
threshold-based wavelet methods heavily depends on the
choice of the mother wavelet. In this study, we compare
several common wavelet types (db3–db5, sym4–sym5, coif3,
Haar, bior1.5 and rbio2.4) used within a deep learning
framework based on autoencoder. The wavelet denoising
performance is evaluated under different noise conditions
using the SNR improvement and Person coefficient as
metrics. The comparison across different wavelet types
shows that all evaluated wavelets achieve similar
performance, with differences within 5% for both metrics.
This indicates that the variation between wavelet types is
relatively small and not statistically significant. These
results suggest that learning the denoising function
directly from data using deep learning can reduce the
dependence on the specific choice of wavelet.

Reliable automated seizure detection from EEG recordings is
essential for continuous, out-of-clinic monitoring and
patient care. Despite the availability of standardized
guidelines for evaluation, many developed approaches do not
follow the recommended practices, making direct comparison
difficult. Reproducibility presents another major concern
due to the lack of complete documentation of preprocessing
and training pipelines. Furthermore, subject-independent
models frequently exhibit significant performance drops
when applied to unseen subjects, implying the lack of
robustness and generalizability in existing approaches. All
these shortcomings hinder the development of seizure
detection systems suitable for real-world applications. In
this work, we analyze reported results and evaluate
multiple seizure detection techniques under standardized,
subject-independent settings. We introduce a
subject-independent lightweight temporal convolutional
model trained on continuous multichannel EEG recordings
from the CHB-MIT dataset. The proposed approach yields
78.00% sensitivity and a false alarm rate of 1.59 per hour
for subject-wise 10-fold cross-validation analysis. Our
results further emphasize the impact of evaluation
methodology on reported performance as well as the need for
transparent reporting to support reproducibility and the
development of reliable seizure detection systems for
real-world deployment.

Electrodermal activity (EDA) is widely used for assessing
autonomic nervous system activity and stress. However, when
using dry textile electrodes, signal stabilization remains
a critical challenge, particularly in wearable
applications. This work investigates the stabilization
dynamics of EDA signals acquired from the upper arm using
dry electrodes and proposes a practical protocol for
determining signal validity. Three pilot studies were
conducted to analyze stabilization time, define validation
criteria, and develop a decision logic for determining
signal readiness. The results show that stabilization time
varies across subjects and conditions, with an average of
26.1 ± 15.2 minutes. Phasic responses were observed even
before tonic stabilization, suggesting their potential role
in early signal validation; however, reliable signal
validation requires the combined evaluation of both
components. Based on these findings, a 5.5-minute
validation protocol was defined and successfully tested on
nine subjects. The proposed approach provides a practical
framework for ensuring reliable EDA acquisition in wearable
systems.

This study investigates stress dynamics during reading in
children with reading difficulties using a multimodal
approach combining eye-tracking (ET), ultra-short-term
pulse rate variability (PRV), and electrodermal activity
(EDA). The goal is to evaluate differences in ET, PRV, and
EDA profiles between struggling and control readers, while
also exploring the potential modulating effects of gender.
Forty-eight elementary school students of the second grade
(24 struggling readers, 24 control readers,
gender-balanced) read two stories. Story 1 was divided into
13 consecutive slides, and Story 2 was on one slide, while
physiological and oculomotor activity was recorded. Linear
mixed-effects models (Story 1) and standard linear models
(Story 2) with false discovery rate (FDR) corrections
showed that the three eye-tracking features (mean fixation
duration, fixation frequency, and saccade frequency) were
significantly different between groups with different
reading skills. EDA metrics indicated a group effect for
the number of EDA peaks; the struggling readers exhibited a
higher peak frequency than the control group. Boys showed
higher tonic variability and larger mean EDA peak
amplitudes than girls in Story 1, while the Interquartile
Range (IQR) of saccade duration and Multifractal Detrended
Fluctuation Analysis (MFDFA) derived short-term scaling
increment were also higher in boys. No significant Group ×
Gender interaction effects were observed across any domain
or story. The results suggest that struggling readers
demonstrate dysregulated visual attention deployment and
heightened sympathetic autonomic arousal during reading,
with ET and EDA features being the most sensitive
biomarkers.

Myoelectric control is a widely used approach in modern
prosthetic hands, where movements are driven by
electromyographic (EMG) signals. However, achieving
reliable control remains a challenge due to the
inconsistency between device capabilities and control
quality. This study evaluated various time domain feature
sets with Linear Discriminant Analysis (LDA) and k-Nearest
Neighbors (KNN) classification algorithms for pattern
recognition. Datasets included EMG recordings from three
participants performing four wrist motions. Results show
that Waveform length is the most discriminative individual
feature (F1-score up to 0.99), while Slope Sign Changes
shows high variability (accuracy 58.28–98.15%). With two
combined features, KNN had 87.36–99.54% accuracy compared
to LDA’s 86.25–96.89%. Using all features, accuracies
reached up to 98.75% for LDA and 99.09% for KNN. It is also
demonstrated that using only two features leads to
satisfactory performance. However, performance strongly
relies on individual anatomical characteristics, indicating
the need for user-specific calibration.

The standing long jump (SLJ) is a complex task requiring
explosive power and relying on effective force absorption
through the kinematic chain during landing. Movement of the
lower extremities in SLJ can be described in terms of the
rotational positions of segments and joints. This study
aims to investigate how information provided by joint and
segment sagittal plane angles relate to jump stability.
Twenty participants (aged: 18-30) performed multiple SLJ,
which were recorded using inertial measurement units (IMU)
affixed to the participants' lower extremities, arms, and
trunk. The angle signals derived from IMU data were
utilized to classify SLJ stability with a 1D convolutional
neural network (1DCNN). Leave-one-subject-out
cross-validation showed that segment signals achieved 87.5%
accuracy and an F1 score of 86.4%, outperforming
joint-based model which achieved only 77.67% accuracy with
76.17% F1 score. Grad-CAM analysis was used do determine
relevant SLJ sections, and highlighted the landing phase as
most important for stability. The study concludes that
sagittal-plane angles, especially during landing, provide
valuable information for SLJ stability detection.

Automatsko prepoznavanje emocija na osnovu pokreta tela ima
potencijal da unapredi interakciju čovek–računar i
čovek–robot, uz širok spektar primena. Ipak, istraživanja u
ovoj oblasti i dalje su u relativno ranim fazama. Ovaj rad
predstavlja preliminarni sistematski pregled automatskog
prepoznavanja emocija na osnovu pokreta tela. Analizirani
su modaliteti i metode prikupljanja podataka, obuhvaćena
emocionalna stanja, primenjene tehnike prepoznavanja,
ostvareni učinak, kao i kontekst i karakteristike učesnika
studija. Rezultati pokazuju da se većina istraživanja
zasniva na upotrebi skupih sistema za snimanje pokreta i
analizi različitih tipova pokreta, uključujući gestove, hod
i plesne sekvence. Studije su obuhvatile prepoznavanje do
osam emocionalnih stanja, primenom metoda mašinskog i
dubokog učenja. Većina radova upoređivala je više pristupa
radi identifikacije metoda sa najboljim performansama, pri
čemu su pojedine studije ostvarile tačnost veću od 95%.
Skupovi podataka uključivali su od 2 do 59 učesnika
različitih profila. Iako postoje brojni izazovi, dalji
razvoj oblasti zavisiće pre svega od unapređenja
tehnologija za prikupljanje podataka, tehnika prepoznavanja
i njihove validacije u realnim uslovima.