Athina Tzovara

Background: Sleep–wake-circadian disorders are prevalent and contribute substantially to morbidity, mortality and reduced quality of life. Evaluation in real-world settings is essential to capture their complexity and association with long-term outcomes. Most sleep-wake registries are restricted to selected populations, lack systematic integration of objective sleep-wake-circadian assessments with the clinical data and provide limited follow-up. To address these limitations, we established the Bern Sleep-Wake Registry based on real-world data from a tertiary University Hospital sleep center. Methods: The registry includes data from patients referred to the multidisciplinary Sleep-Center at Inselspital, University Hospital Bern (founded in 1990), from the year 2000 onwards who underwent at least one objective assessment such as polysomnography, daytime sleep-wake recordings, vigilance tests, and actigraphy. The registry integrates clinical information, questionnaire scores, outputs of objective assessments and raw signal data, and is regularly updated to incorporate new patients and longitudinal follow-up information. Results: As of September 2025, records from over 11’000 patients were evaluated (median age: 50 years (interquartile range 36, 61); 65% male). The most common sleep-wake-circadian disorders were “sleep-related breathing disorders” (67%), “insomnia disorders” (15%) and “central disorders of hypersomnolence” (9%). Available assessments included more than 12’000 polysomnographies, 2’400 multiple sleep latency tests, 2’100 maintenance of wakefulness tests, 1’800 vigilance tests and 4’500 actigraphies. Conclusions: By longitudinally integrating multimodal deep clinical phenotyping with objective assessments and raw physiological signals, the Bern Sleep-Wake Registry provides a foundation for investigating disease trajectories, identifying novel biomarkers and developing data-driven approaches to precise diagnosis, prognosis and individualized treatment.

Background and Purpose: Central disorders of hypersomnolence (CDH) are, except for Narcolepsy Type 1 (NT1), difficult to diagnose and manage because of overlapping features and the lack of reliable biomarkers. Machine learning (ML) has the potential to improve diagnosis by detecting subtle physiological patterns and distinguishing between CDH subtypes. This review systematically explores current ML applications in CDH, assesses their limitations, and suggests future directions. Methods: Following PRISMA guidelines, MEDLINE, Embase, PsycINFO, IEEE Xplore, CINAHL, Web of Science, and Google Scholar (up to June 2025) were searched for studies using ML to classify or characterize CDH in adults. ML methods, data types, and diagnostic outcomes were extracted and analyzed. Results: Out of 3274 studies, 41 met the inclusion criteria (37 peer‐reviewed articles and 4 preprints). Data sources included neuroimaging (fMRI, PET), sleep assessments (MSLT, polysomnography), demographics, and standardized questionnaires. Supervised ML reliably identified known features, including early REM onset, hypocretin deficiency, and spectral EEG changes, showing strong performance for NT1 but limited generalizability across other CDH subtypes. Although many studies reported high accuracy, clinical relevance was often limited by rigid diagnostic labels that may not reflect the true complexity of CDH. Unsupervised learning uncovered heterogeneous phenotypes and exposed limitations in existing diagnostic labels. Conclusion: ML has the potential to improve CDH diagnosis. Deep learning models are promising for feature extraction; however, their black‐box nature and high data requirements hinder clinical application. Future advancements depend on large, diverse datasets, multimodal and longitudinal data, and close collaboration between clinicians and data scientists.

Aims.: Long-term adherence to cardiac rehabilitation (CR) remains low despite its proven benefits for secondary prevention. We explore the usability and engagement patterns of CUOREMA, a personalized mobile-health system integrating self-monitoring, wearable data, virtual coaching, and reinforcement learning-enhanced adaptive interventions to support lifestyle change during and after outpatient CR. Methods.: In a six-month two-center observational study (N = 53, Switzerland and France), patients in outpatient CR used the CUOREMA app. Usability was assessed via validated and solution-specific questionnaires. Engagement patterns were identified using k-means clustering on weekly app usage. Clusters were compared by demographics and psychological themes of personal goals. Health-related questionnaires and the 6-minute walk test (6MWT) were collected at baseline and follow-up visits. Results.: Three engagement clusters (dropout, sporadic, consistent) were identified and sup-ported by matching patterns across app components and smartwatch wearing time. Clusters were not associated with demographic factors. Psychological themes of personal goals sug-gested that intrinsic motivation was more common among sustained users, while extrinsic motivation was more frequent among early dropouts. User experience was rated positively. One center showed significant 6MWT improvement, though the study was not powered for clinical outcomes. However, attrition was high: only 19% of participants used the app on more than 100 days, and questionnaire response rates declined from 55% at baseline to 13% at six months, and health-related drop-out can not be ruled out. Conclusion.: Engagement variability driven by motivational rather than demographic fac-tors is a central challenge in digital CR. Tailoring interventions to individual motivational profiles may improve adherence. Graphical

Patients presenting to neurology clinics commonly have a complex history of comorbidities and partially documented health trajectories, making it essential to reliably extract comorbidity information from historical records. However, existing extraction methods, ranging from rule-based systems to classical machine learning (ML), often have limited accuracy, scalability, or adaptability across diverse documents. We present a large language model (LLM)–based framework for comorbidity extraction from diagnostic texts, capable of handling various prompt formats and textual sources such as patient history, comorbidities, and sleep assessments. The instruction fine-tuned Mistral-24B (Instruct-2501) model achieves 95% macro classification accuracy and 92% F1 score across six common classes of comorbidities, achieving strong performance that is competitive with metrics reported in prior clinical phenotyping and information extraction studies, while complementing recent transformer-based clinical NLP frameworks. The proposed method extracts comorbidities through a transparent hierarchical approach, thereby supporting clinical analysis and providing interpretable insights for disease modeling and personalized treatment planning in sleep medicine.

How can one trace the brain’s orderly directed signals amid a tangle of nerve fibers? Because direct access to actual brain signaling is rare in humans, the precise wiring diagrams for cortico-limbic communication during sleep and wake remain essentially unmapped, hampering progress in neuroscience. Now, a unique neurosurgical window on the human brain allows for electrically mapping cortical connections at the hospital, but studies so far have relied on average signals, masking the dynamic nature of signal flow across brain regions and vigilance states. To causally estimate signal flow, we repeatedly probed cortico-limbic networks with short-lived electrical pulses over days and assessed the variable fate of each transmitted signal on a single-trial basis. In the resulting openly available dataset, we characterized signaling probabilities and directionality across thousands of local and long-range cortico-limbic connections over days. Challenging established views, we found that limbic structures send twice as many signals as they receive, in both wakefulness and sleep. Our findings provide a fundamental framework for causally interpreting signal flow in the brain and formulating therapeutic strategies for brain network disorders.

Electrophysiological studies are profiting from multivariate pattern analysis methods. However, these mostly rely on machine-learning algorithms that assume consistent response latencies across trials and individuals. Deep learning provides high performance without such assumptions, but often at the cost of interpretability of learned features. Here, we evaluated how the interpretability of deep learning for electroencephalography (EEG) data is affected by preprocessing choices, the network’s architecture, and the way the learned features are extracted and visualized. We trained two convolutional neural networks (CNN): (1) ResNet, a residual network, and (2) EEGNet, which leverages spatiotemporal properties of EEG. We trained these networks to decode single-trial EEG responses to three different visual stimuli (visual dataset) and to the presence of a sound (auditory dataset). We then extracted and visualized learned features with two gradient-based techniques: saliency and gradient-weighted activation maps (GradCam). Results showed that EEGNet and ResNet performed at a similar level. Yet, visualization of learned features revealed that different architectures learn different aspects of the data. Between the two CNNs, EEGNet features had a higher similarity to the EEG data than ResNet features. Moreover, the latency and distribution of important electrodes varied depending on the visualization technique. GradCam provided features more similar to EEG data than those with saliency, emphasizing the impact of the feature extraction method on interpretability. Our results call for careful consideration of network architecture and feature visualization methods to improve interpretability, which is a crucial step for advancing the use of deep learning in EEG research.

Background: The diagnosis of sleep disorders presents a challenging landscape, characterized by the complex nature of their assessment and the often divergent views between objective clinical assessment and subjective patient experience. This study explores the interplay between these perspectives, focusing on the variability of individual perceptions of sleep quality and latency. Objective: Our primary goal was to investigate the alignment, or lack thereof, between subjective experiences and objective measures in the assessment of sleep disorders. Methods: To study this, we developed an aspect-based sentiment analysis method for clinical narratives: using large language models (Falcon 40B and Mixtral 8X7B), we are identifying entity groups of 3 aspects related to sleep behavior (day sleepiness, sleep quality, and fatigue). To phrases referring to these aspects, we are assigning sentiment values between 0 and 1 using a BERT-BiLSTM–based approach (accuracy 78%) and a fine-tuned GPT-2 sentiment classifier (accuracy 87%). Results: In a cohort of 100 patients with complete subjective (Karolinska Sleepiness Scale [KSS]) and objective (Multiple Sleep Latency Test [MSLT]) assessments, approximately 15% exhibited notable discrepancies between perceived and measured levels of daytime sleepiness. A paired-sample t test comparing KSS scores to MSLT latencies approached statistical significance ( t ₉₉ =2.456; P =.06), suggesting a potential misalignment between subjective reports and physiological markers. In contrast, the comparison using text-derived sentiment scores revealed a statistically significant divergence ( t ₉₉ =2.324; P =.047), indicating that clinical narratives may more reliably capture discrepancies in sleepiness perception. These results underscore the importance of integrating multiple subjective sources, with an emphasis on narrative free text, in the assessment of domains such as fatigue and daytime sleepiness—where standardized measures may not fully reflect the patient’s lived experience. Conclusions: Our method has potential in uncovering critical insights into patient self-perception versus clinical evaluations, which enables clinicians to identify patients requiring objective verification of self-reported symptoms.

The quality of sleep and its cognitive benefits rely on the cyclic alternance of two distinct sleep stages associated (REM) or not (NREM) with rapid-eye-movements. The ability to predict shifts in sleep stages could help design future interventions in sleep medicine, but it remains unknown how robust the NREM-REM sleep architecture may be for a given individual over many successive nights. We sought to characterize the individual variability and test the predictability of healthy human sleep recorded longitudinally over unprecedented durations (weeks). Based on ultra-long-term sub-scalp electroencephalographic recordings from a newly available, minimally invasive device, we characterized sleep cycles in eight healthy subjects over a median of 30 consecutive days. We first decomposed EEG signals into five frequency bands (δ, θ, α, σ and β) using a multi-taper time-frequency transform. Second, we quantified variability in sleep spectral composition and predictability in sleep stage transitions based on unsupervised and supervised learning methods, respectively. Using dynamic time warping, we quantified the dissimilarity (D) between pairs of nights, showing that it was lower within (D = 2.5 ± 0.7) than across subjects (D = 4.1 ± 0.5, P < 0.001). Further, we extracted archetypal sleep patterns, which are most representative of an individual’s NREM-REM spectral architecture. Based on the found interplay between δ and σ power bands modeled in a generalized linear model, we predicted transitions from NREM to REM two to four minutes in advance with high accuracy (area under the receiver operating characteristic curve = 0.88). Taken together, these results show that sleep is variable over consecutive nights in healthy subjects but that core dynamics in sleep oscillations are consistently shared across individuals. As a translational outlook, the predictability of certain sleep transitions affords the means to anticipate pathological symptoms specific of a given sleep stage.

Patients presenting to neurology clinics commonly have a complex history of comorbidities and partially documented health trajectories, making it essential to reliably extract comorbidity information from historical records. However, existing extraction methods, ranging from rule-based systems to classical machine learning (ML), often fall short in accuracy, scalability, or adaptability across diverse document types.In this study, we present a large language model (LLM)-based framework for comorbidity extraction from diagnostic texts, capable of handling various prompt formats and textual sources such as patient history, prior diagnoses, and structured sleep assessments. The fine-tuned Mistral-24B (Instruct-2501) model achieves 95% macro classification accuracy and 92% F1 measure across six common classes of comorbidities, substantially outperforming prior state-of-the-art approaches. The proposed method extracts comorbidities through a transparent hierarchical approach, thereby supporting clinical analysis and providing interpretable insights for disease modeling and personalized treatment planning in sleep medicine.