Hyo Kyung Lee

Background: Cardiovascular disease (CVD) remains the leading cause of death worldwide, yet participation in cardiac rehabilitation (CR) remains suboptimal. Wrist-worn wearables may support fully remote or hybrid CR through activity monitoring and individualized feedback, but the effectiveness and consistency of benefits remain uncertain. Objective: This study aimed to evaluate the effectiveness of wrist-worn wearable–guided CR delivered through fully remote or hybrid programs initiated during phase II or III CR in patients with CVD. Methods: A total of 8 databases were searched for randomized controlled trials published from January 1, 2015, through August 7, 2026. Eligible trials enrolled adults with CVD and integrated a wrist-worn wearable into fully remote or hybrid CR. Studies focused primarily on heart failure or using wearables only for outcome measurement were excluded. Primary outcomes were peak oxygen consumption (VO2), 6-minute walk distance (6MWD), and daily steps. Random-effects meta-analyses used mean differences (MDs), except for quality of life, which used standardized MDs. Between-study variance was estimated using the DerSimonian-Laird method, pooled-effect CIs used the Knapp-Hartung adjustment, and 95% prediction intervals (PIs) used the Nagashima-Noma-Furukawa parametric bootstrap method. Risk of bias was assessed with Cochrane Risk of Bias 2. Subgroup and univariable meta-regression analyses explored study and intervention characteristics. Results: A total of 18 trials involving 1993 randomized participants, reported in 20 publications, were included; 17 contributed to meta-analysis. Pooled estimates favored wearable-guided CR for peak VO2 (10 studies; 591 participants; MD 1.69, 95% CI 0.46-2.92, 95% PI –2.21 to 5.39 mL/kg/min), 6MWD (9 studies; 1094 participants; MD 31.03, 95% CI 7.66-53.92, 95% PI –62.11 to 123.18 m), and daily steps (8 studies; 474 participants; MD 1062.8, 95% CI 383.5-1825.0, 95% PI –503.0 to 2860.0 steps/day). All 3 PIs included the null, indicating uncertainty about consistency across future comparable settings. No clear pooled effects were observed for anaerobic threshold, BMI, waist circumference, grip strength, or quality of life. Subgroup differences were not statistically significant. In univariable meta-regression, a higher proportion of female participants was associated with greater peak VO2 improvement, although this aggregate-level finding was exploratory. Conclusions: Wrist-worn wearable–guided CR delivered through fully remote or hybrid programs may improve average exercise capacity and physical activity in patients with CVD. However, PIs included the null, indicating uncertainty about consistency across future settings. These programs may extend activity monitoring, individualized feedback, and behavioral support beyond center-based CR, but their multicomponent nature prevents attributing effects to wearables alone. Larger, rigorously designed trials with standardized protocols and outcomes should identify the patients and program components most likely to benefit. Future research should also evaluate long-term adherence, safety, equity, and implementation within routine multidisciplinary CR across diverse health systems.

BACKGROUND Cardiovascular disease (CVD) remains the leading cause of death worldwide, yet participation in cardiac rehabilitation (CR) remains suboptimal. Wrist-worn wearables may support fully remote or hybrid CR through activity monitoring and individualized feedback, but the effectiveness and consistency of benefits remain uncertain. OBJECTIVE This study aimed to evaluate the effectiveness of wrist-worn wearable–guided CR delivered through fully remote or hybrid programs initiated during phase II or III CR in patients with CVD. METHODS A total of 8 databases were searched for randomized controlled trials published from January 1, 2015, through August 7, 2026. Eligible trials enrolled adults with CVD and integrated a wrist-worn wearable into fully remote or hybrid CR. Studies focused primarily on heart failure or using wearables only for outcome measurement were excluded. Primary outcomes were peak oxygen consumption (VO2), 6-minute walk distance (6MWD), and daily steps. Random-effects meta-analyses used mean differences (MDs), except for quality of life, which used standardized MDs. Between-study variance was estimated using the DerSimonian-Laird method, pooled-effect CIs used the Knapp-Hartung adjustment, and 95% prediction intervals (PIs) used the Nagashima-Noma-Furukawa parametric bootstrap method. Risk of bias was assessed with Cochrane Risk of Bias 2. Subgroup and univariable meta-regression analyses explored study and intervention characteristics. RESULTS A total of 18 trials involving 1993 randomized participants, reported in 20 publications, were included; 17 contributed to meta-analysis. Pooled estimates favored wearable-guided CR for peak VO2 (10 studies; 591 participants; MD 1.69, 95% CI 0.46-2.92, 95% PI –2.21 to 5.39 mL/kg/min), 6MWD (9 studies; 1094 participants; MD 31.03, 95% CI 7.66-53.92, 95% PI –62.11 to 123.18 m), and daily steps (8 studies; 474 participants; MD 1062.8, 95% CI 383.5-1825.0, 95% PI –503.0 to 2860.0 steps/day). All 3 PIs included the null, indicating uncertainty about consistency across future comparable settings. No clear pooled effects were observed for anaerobic threshold, BMI, waist circumference, grip strength, or quality of life. Subgroup differences were not statistically significant. In univariable meta-regression, a higher proportion of female participants was associated with greater peak VO2 improvement, although this aggregate-level finding was exploratory. CONCLUSIONS Wrist-worn wearable–guided CR delivered through fully remote or hybrid programs may improve average exercise capacity and physical activity in patients with CVD. However, PIs included the null, indicating uncertainty about consistency across future settings. These programs may extend activity monitoring, individualized feedback, and behavioral support beyond center-based CR, but their multicomponent nature prevents attributing effects to wearables alone. Larger, rigorously designed trials with standardized protocols and outcomes should identify the patients and program components most likely to benefit. Future research should also evaluate long-term adherence, safety, equity, and implementation within routine multidisciplinary CR across diverse health systems.

Intensive care unit (ICU) data are highly irregular, heterogeneous, and temporally fragmented, posing challenges for generalizable clinical prediction. We present PULSE-ICU, a self-supervised foundation model that learns event-level ICU representations from large-scale EHR sequences without resampling or manual feature engineering. A unified embedding module encodes event identity, continuous values, units, and temporal attributes, while a Longformer-based encoder enables efficient modeling of long trajectories. PULSE-ICU was fine-tuned across 18 prediction tasks, including mortality, intervention forecasting, and phenotype identification, achieving strong performance across task types. External validation on eICU, HiRID, and P12 showed substantial improvements with minimal fine-tuning, demonstrating robustness to domain shift and variable constraints. These findings suggest that foundation-style modeling can improve data efficiency and adaptability, providing a scalable framework for ICU decision support across diverse clinical environments.

Irregular sampling and high missingness are intrinsic challenges in modeling time series derived from electronic health records (EHRs),where clinical variables are measured at uneven intervals depending on workflow and intervention timing. To address this, we propose VITAL, a variable-aware, large language model (LLM) based framework tailored for learning from irregularly sampled physiological time series. VITAL differentiates between two distinct types of clinical variables: vital signs, which are frequently recorded and exhibit temporal patterns, and laboratory tests, which are measured sporadically and lack temporal structure. It reprograms vital signs into the language space, enabling the LLM to capture temporal context and reason over missing values through explicit encoding. In contrast, laboratory variables are embedded either using representative summary values or a learnable [Not measured] token, depending on their availability. Extensive evaluations on the benchmark datasets from the PhysioNet demonstrate that VITAL outperforms state of the art methods designed for irregular time series. Furthermore, it maintains robust performance under high levels of missingness, which is prevalent in real world clinical scenarios where key variables are often unavailable.

Acute kidney injury (AKI) presents a public health challenge with profound short and long-term morbidity and mortality. Early prediction and severity identification of AKI are crucial for improving clinical outcomes through timely interventions and efficient resource allocation. Previous studies have predominantly focused on serum creatinine, neglecting the significance of urine output, which, combined with the delayed rise in serum creatinine post-AKI onset, hinders the timely detection of AKI. To address these shortcomings, we propose a novel multi-task learning approach incorporating a continuous urine output monitoring strategy, predicting AKI onset and stage within 6-hour intervals up to 48 hours. Our model exhibits strong performance with the area under the receiver operating characteristic curve of 99.3% and an area under the precision-recall curve of 99.0% for predicting AKI within 48 hours. Also, our model is able to capture overall disease trends perfectly for 35.7% of the AKI cohort and 94.8% of the disease-free cohort. The proposed approach enhances clinical applicability, providing insights into disease dynamics.

Shock is a life-threatening condition characterized by generalized circulatory failure, which can have devastating consequences if not promptly treated. Thus, early prediction and continuous monitoring of physiological signs are essential for timely intervention. While previous machine learning research in clinical settings has primarily focused on predicting the onset of deteriorating events, the importance of monitoring the ongoing state of a patient's condition post-onset has often been overlooked. In this study, we introduce a novel analytical framework for a prognostic monitoring system that offers hourly predictions of shock occurrence within the next 8 hours preceding its onset, along with forecasts regarding the likelihood of shock continuation within the subsequent hour post-shock occurrence. We categorize the patient's physiological states into four cases: pre-shock (non-shock or shock within the next 8 hours) and post-shock onset (continuation or improvement of shock within the next hour). To effectively predict these cases, we adopt supervised contrastive learning, enabling differential representation in latent space for training a predictive model. Additionally, to extract effective contrastive embeddings, we incorporate a feature tokenizer transformer into our approach. Our framework demonstrates improved predictive performance compared to baseline models when utilizing contrastive embeddings, validated through both internal and external datasets. Clinically, our system significantly improved early detection by identifying shock on average 6 hours before its onset. This framework not only provides early predictions of shock likelihood but also offers real-time assessments of shock persistence risk, thereby facilitating early prevention and evaluation of treatment effectiveness.

As delayed treatment of septic shock can lead to an irreversible health state, timely identification of septic shock holds immense value. While numerous approaches have been proposed to build early warning systems, these approaches primarily focus on predicting the future risk of septic shock, irrespective of its precise onset timing. Such early prediction systems without consideration of timeliness fall short in assisting clinicians in taking proactive measures. To address this limitation, we establish a timely warning system for septic shock with data-task engineering, a novel technique regarding the control of data samples and prediction targets. Leveraging machine learning techniques and the real-world electronic medical records from the MIMIC-IV (Medical Information Mart for Intensive Care) database, our system, TEW3S (Timely Early Warning System for Septic Shock), successfully predicted 94% of all shock events with one true alarm for every four false alarms and a maximum lead time of 8 hours. This approach emphasizes the often-overlooked importance of prediction timeliness and may provide a practical avenue to develop a timely warning system for acute deterioration in hospital settings, ultimately improving patient outcomes.

Parkinson’s Disease (PD) is a chronic condition, with extensive research on initial medication selection for treatment, but limited guidance on long-term medication management. This study aims to identify optimal medication adjustment strategies based on patient clusters, focusing on either maximizing time spent in favorable health states or minimizing time spent in unfavorable ones while avoiding adverse effects. To guide treatment, we developed decision models using prescription dosages converted into standardized units for various medications. Using data from the Parkinson’s Progression Markers Initiative, we employed a multivariate time-series clustering approach to capture symptom progression dynamics. This analysis identified four distinct clusters: two representing desirable and undesirable states, and two highlighting motor-focused and non-motor-focused failures across multiple domains. We developed two separate Markov Decision Process (MDP) models to address these dual objectives, which were then integrated into a comprehensive framework that suggests optimal actions and cautions against risky ones for each patient state. This model provides valuable insights for clinical decision-making by offering flexible guidance on adjusting medication intensity rather than prescribing specific medication types, enhancing its applicability in clinical practice.

In the emergency department, accurately identifying patients who will require intensive care unit admission is crucial for effective care and resource allocation. Existing methods, such as triage judgments based on Emergency Severity Index, may not fully capture the dynamic nature of patient conditions, potentially overlooking critical risks. Therefore, it is essential to detect sequential deterioration in patients with initially underestimated triage scores. This study addresses the challenge of identifying patients who, despite being initially triaged as low-acuity, later require intensive care unit admission due to clinical deterioration. We propose an anomaly detection approach to continuously monitor patients for signs of deterioration. To achieve this, we employed a sequential anomaly detection model integrated with Long Short-Term Memory to capture temporal relationships within the data. Given the limitations of existing metrics for assessing clinical relevance, we validated the performance of our model by measuring its ability to correctly identify cases and its effectiveness in providing timely and actionable alerts to healthcare providers. Our model achieved a recall of 0.8299 and a precision of 0.8206, continuously alerting healthcare providers 36% of the time throughout the patient’s stay, with an average lead time of 3.13 hours.