Artur Piet

Background: Continuous glucose monitoring (CGM) can facilitate weight management and lower the risk of metabolic diseases by providing real-time feedback on glycemic responses, thereby enabling more informed lifestyle decisions. However, current CGM systems remain constrained by invasiveness, cost, and short sensor lifespan, limiting their practicality for guiding individualized postprandial low-glycemic diets. Objective: Extending earlier proof-of-concept findings, this study aimed to validate an interstitial glucose (IG) machine learning algorithm in real-world environments using multimodal, continuous data collected from wearable sensors and smartwatches. In the long term, we aim to embed a noninvasive, sensor-based algorithm for estimating tissue glucose within mobile health apps and postprandial low-glycemic diet frameworks to enable scalable, personalized prevention strategies. Methods: We conducted a 2-week study phase during which participants continuously wore 2 noninvasive sensor devices: a scientific sensor wristband (Empatica EmbracePlus) and a commercially available smartwatch (Fitbit Sense 2). As a reference measurement, an invasive CGM sensor (Abbott FreeStyle Libre 3) measured IG levels. For metabolic characterization, 1-point fasting blood and urine samples were collected, and deep phenotyping using state-of-the-art nuclear magnetic resonance spectroscopy and bioelectrical impedance analysis was performed. For participants with overweight and obesity, clinical standard parameters focusing on glucose metabolism were analyzed. Results: A total of 74 participants, 34 (46%) healthy controls and 40 (54%) metabolically at-risk (MR) individuals, simultaneously used an invasive CGM device together with 2 noninvasive wristbands over 2 weeks. Healthy controls were characterized by a mean age of 24.53 (SD 3.68) years and a mean BMI of 22.36 (SD 2.16) kg/m². In contrast, the MR cohort had a mean age of 55.38 (SD 15.08) years and a mean BMI of 35.38 (SD 4.91) kg/m². Furthermore, the MR cohort showed elevated fasting glucose levels (mean 107.56, SD 19.26 mg/dL), hemoglobin A _1c levels (mean 5.67%, SD 0.63%), and an increased homeostatic model assessment of insulin resistance index (mean 4.34, SD 3.49), indicating a disturbed glucose metabolism. The proposed long short-term memory network based on feature vectors obtained the best IG prediction performance, with an average root-mean-squared error of 21.04 (SD 8.32) mg/dL and 98.3% of predictions in zones A and B of the Clarke error grid analysis, representing a high level of predictive accuracy. In addition, based on data from a smartwatch, we achieved a comparable average root-mean-squared error of 23.49 (SD 11.46) mg/dL for the overall cohort. Conclusions: This study demonstrates that IG levels can be predicted from multimodal, noninvasive wearable sensor data using a machine learning approach under real-world conditions. While further validation in larger and more diverse cohorts is warranted, this approach represents a promising step toward accessible, personalized glycemic monitoring and dietary guidance as a preventive tool in mobile health apps.

Glucose monitoring is crucial for maintaining metabolic health, especially for individuals with prediabetes and type 2 diabetes. Wearable devices capable of continuous glucose monitoring offer a promising approach for tracking glycemic variability. This study explores the feasibility of combining non-invasive wearable technologies with machine learning to predict interstitial glucose levels. We utilize the publicly available Big-ideas-glycemic-wearable (BiGW) dataset, which comprises data from 16 prediabetic participants collected via the Dexcom G6 and Empatica E4 wristband. From these multimodal physiological signals including photoplethysmography, heart rate, skin temperature, and electrodermal activity, we engineer twenty-five predefined peak-domain features. The proposed stacked bidirectional gated recurrent unit (stacked bi-GRU) model achieves superior prediction performance, with a root mean square error of 13.11 milligrams per deciliter (mg/dL) and a mean absolute percentage error of 6.85% when food logs are incorporated. Clarke error grid analysis confirms the model's high clinical safety, with only 0.11% of predictions falling into Zone D. This proof-of-concept study presents a noninvasive framework that integrates multimodal sensor data with dietary context via a stacked bi-GRU model and peak-derived feature engineering, validating its potential as a low-burden tool for personalized glucose monitoring and nutrition management, and paving the way for next-generation metabolic health technologies.

Recent advancements in hardware technology have spurred a surge in the popularity and ubiquity of wearable sensors, opening up new applications within the medical domain. This proliferation has resulted in a notable increase in the availability of Time Series (TS) data characterizing behavioral or physiological information from the patient, leading to initiatives toward leveraging machine learning and data analysis techniques. Nonetheless, the complexity and time required for collecting data remain significant hurdles, limiting dataset sizes and hindering the effectiveness of machine learning. Data Augmentation (DA) stands out as a prime solution, facilitating the generation of synthetic data to address challenges associated with acquiring medical data. DA has shown to consistently improve performances when images are involved. As a result, investigations have been carried out to check DA for TS, in particular for TS classification. However, the current state of DA in TS classification faces challenges, including methodological taxonomies restricted to the univariate case, insufficient direction to select suitable DA methods and a lack of conclusive evidence regarding the amount of synthetic data required to attain optimal outcomes. This paper conducts a comprehensive survey and experiments on DA techniques for TS and their application to TS classification. We propose an updated taxonomy spanning across three families of Time Series Data Augmentation (TSDA): Random Transformation (RT), Pattern Mixing (PM), and Generative Models (GM). Additionally, we empirically evaluate 12 TSDA methods across diverse datasets used in medical-related applications, including OPPORTUNITY and HAR for Human Activity Recognition, DEAP for emotion recognition, BioVid Heat Pain Database (BVDB), and PainMonit Database (PMDB) for pain recognition. Through comprehensive experimental analysis, we identify the most optimal DA techniques and provide recommendations for researchers regarding the generation of synthetic data to maximize outcomes from DA methods. Our findings show that despite their simplicity, DA methods of the RT family are the most consistent in increasing performances compared to not using any augmentation.

Parkinson’s disease is characterized by motor and cognitive deficits. While previous work suggests a relationship between both, direct empirical evidence is scarce or inconclusive. Therefore, we examined the relationship between walking features and executive functioning in patients with Parkinson’s disease using state-of-the-art machine learning approaches. A dataset of 103 geriatric Parkinson inpatients, who performed four walking conditions with varying difficulty levels depending on single task walking and additional motor and cognitive demands, was analyzed. Walking features were quantified using an inertial measurement unit (IMU) system positioned at the patient’s lower back. The analyses included five imputation methods and four regression approaches to predict executive functioning, as measured using the Trail-Making Test (TMT). Multiple imputation by chained equations (MICE) in combination with support vector regression (SVR) reduce the mean absolute error by about 4.95% compared to baseline. Importantly, predictions solely based on walking features obtained with support vector regression mildly but significantly correlated with Δ−TMT values. Specifically, this effect was primarily driven by step time variability, double limb support time variability, and gait speed in the dual task condition with cognitive demands. Taken together, our data provide direct evidence for a link between executive functioning and specific walking features in Parkinson’s disease.

Cardiovascular diseases (CVDs) are chronic diseases associated with a high risk of mortality and morbidity. Early detection of CVD is crucial to initiating timely interventions, such as appropriate counseling and medication, which can effectively manage the condition and improve patient outcomes. This study introduces an innovative ontology-based model for the diagnosis of CVD, aimed at improving decision support systems in healthcare. We developed a database model inspired by ontology principles, tailored for the efficient processing and analysis of CVD-related data. Our model’s effectiveness is demonstrated through its integration into a web application, showcasing significant improvements in diagnostic accuracy and utility in resource-limited settings. Our findings indicate a promising direction for the application of artificial intelligence (AI) in early CVD detection and management, offering a scalable solution to healthcare challenges in diverse environments.

Type 2 diabetes mellitus (T2D) poses a significant global health challenge and demands effective self-management strategies, including continuous blood glucose monitoring (CGM) and lifestyle adaptations. While CGM offers real-time glucose level assessment, the quest for minimizing trauma and enhancing convenience has spurred the need to explore non-invasive alternatives for monitoring vital signs in patients with T2D. Objective: This systematic review is the first that explores the current literature and critically evaluates the use and reporting of non-invasive wearable devices for monitoring vital signs in patients with T2D. Methods: Employing the PRISMA and PICOS guidelines, we conducted a comprehensive search to incorporate evidence from relevant studies, focusing on randomized controlled trials (RCTs), systematic reviews, and meta-analyses published since 2017. Of the 437 publications identified, seven were selected based on predetermined criteria. Results: The seven studies included in this review used various sensing technologies, such as heart rate monitors, accelerometers, and other wearable devices. Primary health outcomes included blood pressure measurements, heart rate, body fat percentage, and cardiorespiratory endurance. Non-invasive wearable devices demonstrated potential for aiding T2D management, albeit with variations in efficacy across studies. Conclusions: Based on the low number of studies with higher evidence levels (i.e., RCTs) that we were able to find and the significant differences in design between these studies, we conclude that further evidence is required to validate the application, efficacy, and real-world impact of these wearable devices. Emphasizing transparency in bias reporting and conducting in-depth research is crucial for fully understanding the implications and benefits of wearable devices in T2D management.