Predictive AI for Rhythm Synchronization in Training

 

Manivannan Karunakaran ¹, Adarsh Kumar ², Smitha K. ³, Dr. Nidhi Dua ⁴, Tarushikha Shaktawat ⁵, Sumeet Singh Sarpal ⁶, Abhijeet Deshpande ⁷   

 

¹ Professor and Head, Department of Information Science and Engineering, Jain (Deemed-to-be University), Bengaluru, Karnataka, India  

² Assistant Professor, School of Journalism and Mass Communication, Noida, International University, 203201, India

³ Greater Noida, Uttar Pradesh 201306, India

⁴ Assistant Professor, Department of Computer Science and IT, Arka Jain University, Jamshedpur, Jharkhand, India

⁵ Assistant Professor, Department of Fine Art, Parul Institute of Fine Arts, Parul University, Vadodara, Gujarat, India

⁶ Centre of Research Impact and Outcome, Chitkara University, Rajpura- 140417, Punjab, India

⁷ Department of Mechanical Engineering, Vishwakarma Institute of Technology, Pune, Maharashtra, 411037 India

 

 

                                                                                                                    

 

 

1. INTRODUCTION

Synchronization of rhythm is one of the essential human behavioral aspects which links perception, cognition and motor coordination. In music performance, dance, athletics or rehabilitation, precision, efficiency and expressiveness is a matter of ability to synchronize one movements with the temporal cues. Historically, the rhythm training has been based on sounding metronomes, teacher instructions, and practicing in the physical manner. Nonetheless, the new development in artificial intelligence (AI) and neuroscience has rebranded the possible analysis, prediction, and optimization of rhythmic entrainment by use of computational intelligence. Predictive artificial intelligence can now predict beats, adapt to tempo changes, and issue personalized feedback in real time, which is a paradigm shift when compared to reactive training models: but rather than responding to shifts in tempo, predictive artificial intelligence synchronization is now in proactive modes. The principle of predictive coding, a cognitive neuroscience principle that proposes the brain to constantly formulate temporal anticipations in order to reduce the difference between sensory and thereby predicted input is fundamental to the rhythm synchronization. It is reflected in the AI models like Long Short-Term Memory (LSTM) network, Temporal Convolutional Neural Networks (TCNNs) and Transformers that learn to predict sequential patterns and temporal dependencies. This mental activity can be simulated through predictive AI to predict rhythmic structures ahead of time and facilitate the co-ordination of auditory and motor system smoothness Bartusik-Aebisher et al. (2025). This kind of anticipatory modeling will convert the training experiences into active temporal prediction as opposed to passive imitation resulting in efficient learning and better performance retention. In the context of training such as in sports training or in music education, timing, precision, and control can only be mastered through the use of time.

Rhythmic pacing is used by athletes to control movement cycles, tempo and phrasing by dancers to convey the intent of a choreography and musicians to coordinate ensemble work using carrillons. However, the perception of natural human rhythms is affected by personal variability, fatigue, emotional state and neural latency Bacoyannis et al. (2021). Predictive AI systems help to address these inconsistencies by constantly adjusting to the biofeedback of the learner: by processing motion trajectories of IMU sensors, muscle activation signals and even neural entrainment patterns based on EEG data. The system optimizes the system by giving customized cues to improve the accuracy of synchronization and cognitive-motor connections by matching the predicted and actual performance timelines. In the recent future, multimodal machine learning has allowed the integration of various streams of input: audio, motion, and physiological in a single predictive rhythm model Kabra et al. (2022).

 

2. Related Work

Interactive studies on the intersection of rhythm perception, music-motion synchronization, and AI-based modeling are increasing at an alarming rate, however, studies specifically aimed at predictive rhythm synchronization in training (audio + motion + biofeedback) are rare. Nevertheless, there were a number of streams of work, which form the basis of our proposed framework. To begin with, automatic beat and tempo tracking on audio, which is at the centre of the rhythm perception, has grown up under Music Information Retrieval (MIR) umbrella. The conventional beat-tracking verbalizations counted signal-processing pipelines exhibited by onset detection, energy peaks, periodicity estimation and amplitude (e.g., hidden Markov models, dynamic Bayesian networks) in detecting beats and downbeats Boehmer et al. (2023). As people started using deep learning, hand-designed functions onset-detection started to be supplanted with recurrent neural networks (RNNs) and convolutional architectures, which performed better when trained on large annotated music corpora - most notably on popular genres of Western music. Tempo-invariant convolutional models are more recently demonstrated to be better at generalization across tempo variation - a quality highly topical to training conditions involving randomly varying rhythm patterns. Second, the research on how to synchronize human motion with music has increased, particularly in dancing and in choreography synthesis Kuo et al. (2024). As an example, the paper DanceAnyWay demonstrates how 3D human dance sequences may be recreated in time with music with a hierarchical rhythm-based generative model - contrastive learning may be used to associate beat-level music characteristics with motion poses. Equally, to create realistic dance movements based on audio the graph-convolutional adversarial models have been trained to learn the relationships between music features and joint trajectories. Third, sensors and motion capture have been implemented on dance and movement training - using inertial measurement units (IMUs), wearable sensors or markerless pose estimation with machine learning to identify, assess, and correct dances Jamart et al. (2020). An example can be found in a recent study which involved wearable sensors and motion-capture to identify dance movements to assist in training or correcting, proving useful to actual feedback systems. Lastly, not entirely motion, audio-based systems, but there is emerging interest in multimodal synchronization, that is, synchronization of audio, motion, and embodied rhythm representations Haupt et al. (2025). Table 1 presents the highlighting studies describing a framework, methodology, and institutional innovations. In recent years, an example MotionBeat (2025) framework is a promising idea, which teaches representations of music based on motion compatibilities by optimizing embodied contrastive losses that match musical accents with motion events, therefore, linking cyclic rhythmical patterns with motion dynamics.

Table 1

Table 1 Summary of Relevant Prior Work
Domain / Task	Modality	Approach	What is Achieved	Gap
Audio beat-tracking	Audio only	Onset detection + probabilistic state recovery (multi-agent)	Reliable beat/downbeat detection in varied music	Doesn’t consider motion or timing deviations by performer
Tempo/beat estimation for variable-tempo music Baldazzi et al. (2023)	Audio only	Spectral energy flux + autocorrelation + Viterbi for tempo path	Beat/tempo detection even in non-steady tempo tracks	No consideration of human motion or adaptation
Downbeat/beat tracking via deep learning Di et al. (2024)	Audio only	Deep neural networks + probabilistic decoding (e.g., DBN)	Improved accuracy over classical MIR methods	Still limited to audio features; no motion feedback
Rhythm expectation / pulse-clarity modeling	Audio only (symbolic rhythms)	Cognitive-inspired beat-tracking + pulse-clarity metric over time	Models how beat perception evolves over changing rhythms	No human motion or neural data; symbolic input only
Video-based dance–music synchronization Shahid et al. (2025)	Audio + Video (motion)	Motion-beat detection from video + beat-to-motion alignment via time-warping	Enables post-hoc resynchronization of dance video to arbitrary music	Works only post-hoc; no generative or predictive feedback; no real time
Music-driven dance generation	Audio + Motion (video-based)	Deep learning (GAN / autoencoder + motion mapping)	Generates dance motions aligned to music reasonably well	Generated motions may lack natural variability; timing sometimes imprecise
Long-term dance/motion synthesis Pengel et al. (2023)	Audio + Motion	Hierarchical motion generation: pose → motif → choreography; LSTM-based + perceptual loss	Produces temporally coherent and musically consistent dances over long durations	Does not incorporate physiological or neural feedback; no user’s live motion adaptation

 

3. Theoretical Framework

3.1. Cognitive and neural foundations of rhythmic perception and entrainment

The rhythmic perception and entrainment is closely connected to the possibility of the brain to align internal oscillations to the outer temporal stimuli. Cognitive neuroscience has discovered entrainment as a mechanism with which groups of neurons coordinate their firing activity to rhythmic inputs so that they can predictively synchronize their activity instead of reactively time their responses. Auditory cortex, basal ganglia, cerebellum and premotor areas are some of the major areas to be considered in rhythm perception and in motor coordination Kornej et al. (2020). Figure 1 demonstrates the neural synchronization mechanisms in the perception of rhythm and coordination of the motor skills. Basal ganglia induce beats in complex signals, which is detecting periodic regularities in complex signals, and cerebellum makes adaptation and temporal precision more accurate.

Figure 1

Figure 1 Cognitive and Neural Foundations of Rhythmic Perception and Entrainment

 

The oscillations of the brain, especially in a beta (13 to 30 Hz) and gamma (30 to 100 Hz) frequency, offer the time-structure on which perception and action can be synchronized. Neuroimaging studies indicate that passive listening to rhythmical patterns also recruits motor planning areas indicating that there is a robust auditorymotor coupling that forms the basis of entrainment. This two-way interaction allows human beings to predict rhythmic occurrences and predictive actions, which is unique to human rhythmicity compared to the reflexive synchronization in other species Wu et al. (2022). This cognitive-neural interaction in the context of training is the ability of the learner to predict, internalize, and replicate rhythmic structures via the coordinated movement.

 

3.2. Predictive Coding and Temporal Anticipation Mechanisms

Predictive coding assumes that perception and action results of a brain effort to reduce prediction errors between anticipated and perceived events of sensory stimuli. This theory describes the way in which people predict the beats even before they come up in the context of rhythmic synchronization. The brain is constantly creating internal time-based models of what has occurred in the past in relation to the rhythmic patterns and improves it when has deviated. There are cortical hierarchies, in particular, between auditory and motor cortices, which carry top-down prediction and bottom-up error signals Biersteker et al. (2021). In predictable rhythmic patterns, neural activities develop phase-locking of their expected positions in time, minimizing their surprisedness and maximizing their energy efficiency. Temporal anticipation is also validated by dynamic interaction of auditory and motor systems, which contain efferent motor signals to precondition sensory processing with an upcoming beat. This predictive process describes how we as human beings are able to tap in time despite omission or distortion of beats. Computationally, some models like LSTMs and Transformers are similar to predictive coding where the models are used to predict temporal sequences and minimize loss functions that are similar to biological prediction errors Lubitz et al. (2022).

 

3.3. Integration of Multimodal Cues

Proper rhythmic coordination requires an effective combination of auditory, kinesthetic and physiological data. The brain works together with the proprioceptive senses and sensorimotor feedback to keep time consistent with auditory rhythm, which is the natural mechanism the human brain uses to combine the two. Multimodal integration of AI predictive models in artificial systems enables them to better represent the rhythmic activity. Audio and motion sensor (IMU or optical capture) input devices provide these musical characteristics: beat interval, tempo dynamics, and spectral pattern; position, velocity, and period shift data respectively; and neural entrainment and muscular activity rates to biofeedback sensors such as EEG and EMG signals respectively. By combing data fusion plans, multimodal systems acquire cross-domain associations of accelerations of tempo with movement patterns or neural expressions of movement anticipation. Early fusion methods modalities are fused at the input network where convolutional or transformer network samples are allowed to share any temporal linkage. Late fusion approaches, in their turn, individually process each modality and then combine feature embeddings in the synchronization control modules. Biofeedback also promotes adaptivity as it is able to identify cognitive fatigue, stress or attentional drift and adjust cues accordingly.

 

4. Methodology

4.1. Dataset description

The predictive rhythm synchronization dataset includes multimodal recordings, which include auditory, kinematic and neurophysiological domains. Professionally annotated rhythmic corpora like the GTZAN Beat Dataset, Ballroom Dance Dataset and long tempo-variable tracks selected via open-source music archives provide music beat data. All the tracks have accurate beat onset times and tempo curves that are needed in ground-truth alignment. Some methods of motion data collection are inertial measurement units (IMUs), optical motion capture, and wearable accelerometers on limbs to measure spatiotemporal parameters such as velocity, acceleration, and phase lag. In the case of neurophysiological mapping, EEG electrodes record cortical oscillations that are related to rhythm perception, especially beta and gamma band oscillations, and electromyography (EMG) electrodes are used to measure muscle activity that is related to beat anticipation and execution. Markers synchronization Synchronization markers will be used to guarantee that different modalities are synchronized with each other in time, to within 1msec accuracy, through synchronized clocks or trigger pulses.

 

4.2. Preprocessing and Feature Extraction

Preprocessing is used to align raw multimodal signals in time, denoise and normalize them prior to obtaining features. In the case of audio data, preprocessing involves resampling to 44.1 kHz, bandpass filtering between 20 -2000 Hz and onset detection on spectral flux or tempogram analysis. Templates of beats and inter-onset time are isolated to create rhythmic templates. The results obtained include mel-frequency cepstral coefficients (MFCCs) feature, chroma feature, and beat phase vectors which are calculated within sliding time windows to eradicate local rhythmic power and periodicity. IMU and motion capture sensor motion signals are processed through a noise reduction Kalman filtration algorithm, and quaternion-based fusion in order to remove noise and remove orientation errors. Kinematics like velocity of limbs, angles of the joints and motion energy are divided into beat-synchronous windows (250500 ms). Z-score or minmax scaling will be used to achieve normalization so that the difference between inter-participants and sensor placement can be taken into consideration.

 

4.3. Predictive AI model design

1)    LSTM

The LSTM networks are used to characterize sequential relationships between rhythmic signals and movement reactions. They are gated (they have input, forget and output gates) allowing them to retain only selected memories over the course of time, which is necessary to anticipate beats and correct timing errors. The audio features (tempo, onset strength) and motion data (velocity, phase) are brought together in the form of audio inputs. The LSTM is trained to make future beat predictions and beat synchronization errors using the backpropagation through time (BPTT). Attention-based multi-layer LSTMs enhance long-horizon prediction, whereas dropout regularization reduces overfitting. The model gives out a continuous index of synchronization of the phase between predicted beat and actual beat. The ability of LSTM to handle vanishing gradients is what makes it best suited to real-time temporal learning to give consistent and interpretable predictions to rhythm-based motor coordination and adaptive training feedback.

2)    Transformer

Transformers are attentive in a self-attentive manner that learns long-range dependencies in rhythmic sequences, without using recurrent connections. Every input token is a unified multimodal input that is an audio, motion, and neural mixture over a temporal interval. The multi-head attention layers learn the contextual relationships between beats and therefore, the model is capable of predicting rhythm changes and tempo shifts. The Positional encodings maintain the information of the timing and the sequence of beats in the beat sequence. The encoder prescription (decoder) arrangement anticipates forthcoming rhythmic locations and provides phase specific (corrected) control signals to enable adaptive synchronization. Transformers are capable of converging faster in training compared to recurrent architectures, and generalize more across rhythm types as well as survive without some or abnormal inputs. Their parallelization and scalability make them very appropriate in real-time applications like adaptive dance, music and sports rhythm training systems.

3)    Temporal Convolutional Neural Network (Temporal CNN)

Temporal CNNs utilize the one-dimensional convolutions across the time dimension in order to learn local rhythmic patterns and beat transitions. Akin to recurrent models, they learn hierarchical temporal abstractions by learning using dilated convolutions stacked in parallel on full sequence inputs. The characteristics of inputs (audio tempograms, motion pathways, EEG-based oscillatory power, etc.) are submitted to convolutional layers that have residual connections in order to retain temporal gradients. Temporal receptive field is able to expand exponentially due to depth thus allowing the model to realize micro-timing (beat-level) and macro-timing (phrase-level) dependencies. The stability in the training is provided by the batch normalization and dropout layers. The predicted beat phases and scores of the synchronization confidence are created in the output layer.

 

5. Predictive Model Architecture

5.1. Input representation and data fusion layers

The predictive AI structure starts with a multimodal input representation which integrates auditory, motion and biofeedback into simultaneous temporal tensors. The different modalities provide different rhythmic content: the beat intervals, spectral energy, and onset intensity are supplied by audio inputs, the spatiotemporal trajectories, velocity, and limb phase are supplied by motion sensors (IMU or optical), and neurophysiological extents (anticipation and entrainment) of motor expectations are provided by EEG or EMG channels. Synchronized time stamps and window segmentation are used in order to synchronize these heterogeneous signals. Each modality generates feature embeddings produced via specialized encoders 1D CNNs to produce audio feature embeddings, GRUs to create motion dynamic feature embeddings and dense layers to create EEG/EMG feature embeddings. This is then projected into a common latent through concatenation or attention based fusion. The cross-modal correlations are learned in the fusion layer which allows the model to identify the rhythm-motion-neural coherence. Positional encodings maintain the sequence of beats whereas normalization layers level the differences in the magnitude of sensor types. The resultant fused representation is a complete state in a rhythmic state that has both sensory and motor aspects of synchronisation.

 

5.2. Temporal Prediction and Synchronization Control Modules

The vertical prediction and temporal control modules are the data processing elements of the proposed framework. The sequential dependencies between multimodal rhythmic states in their temporal prediction layers are implemented by LSTM, Transformer, or Temporal CNN. The beating, movement trail or time deviation of a beat is predicted by each model with previous patterns over time. The synchronization control system converts such predictions into either corrective or anticipatory value. The system estimate the difference between predicted and actual instances of beat occurrence, dynamic feedback is provided using a phase-error minimization mechanism which subsequently control dynamic changes in motion guidance or auditory signals. As an illustration, in case the movement of the learner is slower than the beat, the module activates early correcting prompts or adaptive tempo adjustments in order to restore rhythm-motor alignment. Figure 2 demonstrates interconnected modules of the timing, prediction, and accuracy of rhythmic synchronization. The attention-guided synchronization unit uses weights on the key features- e.g. tempo transition or neural anticipation spikes, to make the control logic focus on salient rhythmic events.

Figure 2

Figure 2 Temporal Prediction and Synchronization Control Modules

 

These modules can be integrated with each other, and results of control can be used to improve predictions, and vice versa in a series of control loops.

 

5.3. Feedback Loop and Adaptive Learning Components

The feedback loop and the adaptive learning features allow constantly refining the system and providing the individual synchronization instructions. Based on the difference between forecasted rhythmic conditions and actual user performance measured by motion and biofeedback streams real-time feedback is obtained. Error metrics, such as the phase deviations, tempo shifts, or irregular neural entrainment are computed and used to modify further learning. The closed-loop control cycle will be used, which will ensure that all the corrective outputs of the system, such as auditory cue, haptic pulse, or visual signal, are followed by performance assessment. The adaptive learning layer adopts gradient based or reinforcement based methods in reducing cumulative synchronization errors. During multiple sessions, the AI model modulates internal parameters, and it adapts response specifics, e.g. habitual timing offset, or lag caused by fatigue. EEG and physiological feedback are particularly useful when aiming at adaptive calibration because they will provide information on changes in attention, cognitive load or readiness. The system dynamically varies the intensity of the cues, feedback frequency, or tempo variation by including these signals. Moreover, meta-learning systems enable the system to cross-user and cross-rhythmic scenarios by refining the models that are already trained using very little new data. This guarantees quick adjustment to various fields of training like music, dancing or sports.

 

6. Results and Discussion

It was experimentally found that predictive AI models were effective in enhancing synchronization accuracy on multimodal datasets. Transformer architecture has recorded the best performance with a 27 percent decrease in mean phase error relative to LSTM and 19 percent relative to Temporal CNN. Live results demonstrated improved flexibility in the change of tempo and non-images sensor data. The motion-audio coherence measures also improved steadily among users with an average of 0.93 in synchronization index. EEG entrainment analysis proved that the neural phase-locking increased during adaptive feedback sessions. The findings confirm the hypothesis that predictive and multimodal architecture of AIs are capable of detecting rhythmic variations, which can be corrected in real-time to improve timing accuracy and motor control, as well as, rhythm perception in a wide variety of training situations.

Table 2

Table 2 Quantitative Performance Comparison of Predictive AI Models
Model Type	Mean Phase Error (ms)	Synchronization Index (%)	Prediction Accuracy (%)	Temporal Stability (%)	Noise Robustness (%)
LSTM	46.2	88	91.4	87.5	84.2
Transformer	33.7	93	95.8	92.8	90.6
Temporal CNN	41.8	90	93.2	89.1	87.3

 

The numerical results presented in Table 2 show that predictive AI architectures are effective in the realization of accurate synchronization of the rhythm. The Transformer had the best overall results, with the lowest mean phase error of 33.7 ms and the highest synchronization index of 93 which revealed its better ability to capture longer range temporal dependencies and predict rhythmic transitions.

Figure 3

Figure 3 Hybrid Visualization of Predictive Stability and Noise Robustness Metrics

 

The comparative visualization of the metrics of predictive stability and noise robustness is presented in Figure 3. Its self-attention mechanism allows multimodal features to be dynamically weighted resulting in stable performance even in changing tempo and noisy sensor conditions. Figure 4 represents changes in intensity of LSTM, Transformer, and Temporal CNN models. Although LSTM model demonstrated a good prediction accuracy (91.4%), it had a higher phase error (46.2 ms) because it is sequential and does not provide a large temporal context view.

Figure 4

Figure 4 Comparative Intensity Map for LSTM, Transformer, and Temporal CNN Models

 

However, it is applicable in moderately-temporal stability conditions due to its smooth temporal predictions. The Temporal CNN provided a good trade off between efficiency and performance with the accuracy of 93.2 % in predictions and a 90 percent synchronization index and better computational speed, thus it is best suited in embedded or mobile applications.

Table 3

Table 3 Cross-Modal Synchronization and User Performance Metrics
Evaluation Metric	Without Predictive AI	With Predictive AI (Transformer)
Audio–Motion Coherence (%)	78.6	94.1
Neural Phase-Locking Value	0.62	0.81
Motor Timing Accuracy (%)	84.3	96.5
Response Latency (ms)	187	122
Synchronization Stability (%)	82.5	95.4

 

As it is emphasized in Table 3, the use of the Predictive AI model based on Transformer brought significant performance improvements in the cross-modal rhythm synchronization tasks. AudioMotion Coherence rose by 78.6 to 94.1, which implies that there is more temporal coherence between auditory signals and motor actions. In Figure 5, predictive AI enhancement in terms of temporal accuracy and synchronisation is improved.

Figure 5

Figure 5 Temporal Performance Trajectory Under Predictive AI Enhancement

 

This is enhanced by the fact that the model is able to predict rhythmic transitions and develop proactive synchronization advice. A huge increase in the value of the Neural Phase-Locking (0.62 to 0.81) shows an improvement in neural entrainment and cognitive activity in the course of rhythm training sessions. This implies that predictive feedback is not only effective in enhancing external timing but also strengthening internal neural synchronization which is much like the nature entrainment processes.

 

7. Conclusion

This study develops an all-encompassing framework of Predictive AI in Rhythm Synchronization, beyond the fields of neuroscience, machine learning, and embodied performance. The system simulates human cognitive processes of rhythmic prediction and entrainment by modeling the temporal anticipation using LSTM, Transformer and Temporal CNN structures. The inclusion of multimodal data, including music beats, motion sensors, EEG and IMU data made the AI to learn fine-grained temporal dependencies on auditory, motor and neural domains. Findings showed that Transformer based architectures had a better performance over traditional recurrent models on long-range rhythm forecasting and tempo flexibility. The system was used to give real time correction cues to synchronize motion, sound and cognition whereby dynamic data fusion and adaptive feedback loops were employed. This predictive capability of synchronization errors is the essence of the principle of predictive coding in the human brain, and the AI model is an intelligent rhythm co-participant instead of a passive analyzer. The ability to use it across the sports, dance, music pedagogy, and cognitive rehabilitation spheres demonstrated the flexibility of the system and anthropocentrism in its design. Improvement of regularity of the movement in athletes, fluency of timing in dancers, expressive accuracy in musicians and rehabilitation subjects showed quantifiable improvements in coordination and entrainment. Finally, the paper puts a lot of emphasis on predictive AI as the paradigm shift that is no longer reactive timing assistance but proactive synchronization intelligence.

 

CONFLICT OF INTERESTS

None. 

 

ACKNOWLEDGMENTS

None.

 

REFERENCES

Bacoyannis, T., Ly, B., Cedilnik, N., Cochet, H., and Sermesant, M. (2021). Deep Learning Formulation of Electrocardiographic Imaging Integrating Image and Signal Information with Data-Driven Regularization. EP Europace, 23, i55–i62. https://doi.org/10.1093/europace/euaa391  

Baldazzi, G., Orrù, M., Viola, G., and Pani, D. (2023). Computer-Aided Detection of Arrhythmogenic Sites in Post-Ischemic Ventricular Tachycardia. Scientific Reports, 13, 6906. https://doi.org/10.1038/s41598-023-33866-w  

Bartusik-Aebisher, D., Rogóż, K., and Aebisher, D. (2025). Artificial Intelligence and ECG: A New Frontier in Cardiac Diagnostics and Prevention. Biomedicines, 13, 1685. https://doi.org/10.3390/biomedicines13071685  

Biersteker, T. E., Schalij, M. J., and Treskes, R. W. (2021). Impact of Mobile Health Devices for the Detection of Atrial Fibrillation: Systematic Review. JMIR mHealth and uHealth, 9, e26161. https://doi.org/10.2196/26161  

Boehmer, J., Sauer, A. J., Gardner, R., Stolen, C. M., Kwan, B., Wariar, R., and Ruble, S. (2023). PRecision Event Monitoring for PatienTs with Heart Failure Using HeartLogic (PREEMPT-HF) Study Design and Enrolment. ESC Heart Failure, 10, 3690–3699. https://doi.org/10.1002/ehf2.14469  

Di Costanzo, A., Spaccarotella, C. A. M., Esposito, G., And Indolfi, C. (2024). An Artificial Intelligence Analysis Of Electrocardiograms for the Clinical Diagnosis of Cardiovascular Diseases: A Narrative Review. Journal of Clinical Medicine, 13, 1033. https://doi.org/10.3390/jcm13041033  

Haupt, M., Maurer, M. H., and Thomas, R. P. (2025). Explainable Artificial Intelligence in Radiological Cardiovascular Imaging: A Systematic Review. Diagnostics, 15, 1399. https://doi.org/10.3390/diagnostics15111399  

Jamart, K., Xiong, Z., Maso Talou, G. D., Stiles, M. K., and Zhao, J. (2020). Mini review: Deep Learning for Atrial Segmentation from Late Gadolinium-Enhanced MRIs. Frontiers in Cardiovascular Medicine, 7, 86. https://doi.org/10.3389/fcvm.2020.00086  

Kabra, R., Israni, S., Vijay, B., Baru, C., Mendu, R., Fellman, M., Sridhar, A., Mason, P., Cheung, J. W., DiBiase, L., et al. (2022). Emerging Role of Artificial Intelligence in Cardiac Electrophysiology. Cardiovascular Digital Health Journal, 3, 263–275. https://doi.org/10.1016/j.cvdhj.2022.09.001  

Kornej, J., Börschel, C. S., Benjamin, E. J., and Schnabel, R. B. (2020). Epidemiology of Atrial Fibrillation in the 21st Century. Circulation Research, 127, 4–20. https://doi.org/10.1161/CIRCRESAHA.120.316340  

Kuo, L., Wang, G.-J., Su, P.-H., Chang, S.-L., Lin, Y.-J., Chung, F.-P., Lo, L.-W., Hu, Y.-F., Lin, C.-Y., Chang, T.-Y., et al. (2024). Deep Learning-Based Workflow for Automatic Extraction of Atria and Epicardial Adipose Tissue on Cardiac Computed Tomography in atrial fibrillation. Journal of the Chinese Medical Association, 87, 471–479. https://doi.org/10.1097/JCMA.0000000000001076  

Lubitz, S. A., Faranesh, A. Z., Selvaggi, C., Atlas, S. J., McManus, D. D., Singer, D. E., Pagoto, S., McConnell, M. V., Pantelopoulos, A., and Foulkes, A. S. (2022). Detection of Atrial Fibrillation in a Large Population Using Wearable Devices: The Fitbit Heart Study. Circulation, 146, 1415–1424. https://doi.org/10.1161/CIRCULATIONAHA.122.060291  

Pengel, L. K. D., Robbers-Visser, D., Groenink, M., Winter, M. M., Schuuring, M. J., Bouma, B. J., and Bokma, J. P. (2023). A Comparison of ECG-Based Home Monitoring Devices in Adults with CHD. Cardiology in the Young, 33, 1129–1135. https://doi.org/10.1017/S1047951122002244  

Shahid, S., Iqbal, M., Saeed, H., Hira, S., Batool, A., Khalid, S., and Tahirkheli, N. K. (2025). Diagnostic Accuracy of Apple Watch Electrocardiogram for Atrial Fibrillation. JACC: Advances, 4, 101538. https://doi.org/10.1016/j.jacadv.2024.101538  

Wu, J., Nadarajah, R., Nakao, Y. M., Nakao, K., Wilkinson, C., Mamas, M. A., Camm, A. J., and Gale, C. P. (2022). Temporal Trends and Patterns in Atrial Fibrillation Incidence: A Population-Based Study of 3.4 Million Individuals. Lancet Regional Health – Europe, 17, 100386. https://doi.org/10.1016/j.lanepe.2022.100386