Weakly Supervised Wearable AI Tracks Continuous Functional States in Diabetes Study

In an observational study investigating computational approaches to continuous digital health tracking, biomedical researchers at Chengdu University of Traditional Chinese Medicine have developed a weakly supervised deep-learning framework named MFSI-Net that reconstructs continuous, study-specific trajectories of physiological functional regulation in adults with type 2 diabetes mellitus. Published in the journal Frontiers in Physiology, the research addresses a core methodological bottleneck in modern wearable computing: consumer biosensors collect dense, longitudinal physiological data around the clock, yet clinical experts can only provide sparse, intermittent evaluations during isolated consultations. To bridge this temporal resolution gap, the investigators formulated an integrated analytical pipeline combining fold-contained dependency-aware feature selection, Gaussian process regression for continuous soft-label generation, and a hybrid temporal neural architecture that unites recurrent and self-attention mechanisms. Evaluated across ninety participants over a comprehensive twenty-eight-day continuous tracking protocol, the framework achieved close internal agreement with model-derived soft labels, registering a root mean squared error of 0.17 and a Pearson correlation coefficient of 0.72. The study authors explicitly emphasize that these quantitative performance metrics represent internal technical consistency within their model-supervised framework rather than direct diagnostic agreement with an external clinical gold standard, offering a reproducible framework for studying continuous trajectories synthesized from sparse clinical ratings.

The clinical motivation underlying the investigation centers on the pervasive autonomic and metabolic disturbances that frequently complicate type 2 diabetes mellitus. Diabetic autonomic neuropathy represents a widespread, debilitating sequela stemming from persistent hyperglycemia, microvascular damage, and progressive neurodegenerative injury throughout the autonomic nervous system. Within this patient population, autonomic impairment rarely exists in isolation; it frequently intersects with fragmented sleep architecture, disrupted rest-activity rhythms, and systemic glycemic fluctuations. While conventional clinical evaluation relies upon distinct, episodic diagnostic tools—such as standardized cardiovascular autonomic reflex testing, formal laboratory polysomnography, actigraphy arrays, and venous blood panels—these specialized modalities capture isolated snapshots across disparate environments rather than ongoing functional dynamics during daily life. Continuous sensor technology offers an attractive pathway to capture real-time physiological indicators across naturalistic living conditions. However, raw data generated by commercial-grade smartwatches and connected rings reflect heterogeneous device-level estimates rather than direct measurements of unified clinical states. The authors designed MFSI-Net to explore whether multimodal wearable streams could mathematically capture a continuous composite index, termed the Multi-System Functional State Index, without claiming to replace standardized clinical diagnostic assessments or discovering novel autonomous biological mechanisms.

To rigorously evaluate the framework in a clinical cohort, the investigators conducted a prospective observational trial enrolling ninety adult outpatients with confirmed type 2 diabetes mellitus diagnosed according to World Health Organization criteria. The enrolled participants, comprising forty-two men and forty-eight women with a mean age of 53.6 years and a median disease duration of seven years, were required to maintain a stable antidiabetic pharmacological regimen for at least three months prior to study entry while possessing the cognitive and physical capacity to navigate smartphone-based logging applications. Over the twenty-eight-day continuous monitoring window, participants were outfitted with a commercial sensing suite consisting of a factory-calibrated SIBIONICS continuous glucose monitor recording interstitial glucose concentrations every five minutes, a Tuya smartwatch tracking heart rate and movement, and a Freeda smart ring capturing peripheral blood oxygen saturation and skin temperature variations. Across the cohort, median effective wear duration reached twenty-six days. Signal completeness post-preprocessing differed by physiological stream, reaching 94.6 percent for heart-rate variability, 91.3 percent for sleep architecture, 89.7 percent for physical activity, and 78.2 percent for nocturnal skin temperature, where minor signal dropouts occurred primarily due to temporary displacement of smart-ring sensors during sleep.

Supervisory clinical labels were generated through an extensive, blinded expert evaluation procedure conducted twice daily, during morning windows between 07:00 and 09:00 and evening windows between 20:00 and 22:00. Two attending physicians with more than ten years of clinical experience in traditional Chinese internal medicine independently evaluated each participant without access to wearable biosensor streams, continuous glucose trajectories, or preceding assessment scores. Each clinical encounter integrated a structured seven-item clinician assessment covering mood, energy levels, sleep quality, appetite, bowel function, cognitive focus, and somatic discomfort, with each domain scored on an ordinal scale from zero to three to yield a composite symptom score ranging from zero to twenty-one. To capture dominant regulatory dynamics, physicians classified patients into three directional categories: under-regulation (−1), indicating pronounced fatigue, low energy, poor appetite, and disrupted sleep; normal regulation (0), indicating balanced functional status without overt deficit or excess; and over-regulation (+1), reflecting agitation, mental restlessness, insomnia, physical tension, or heat-type discomfort. Controlled-light tongue photography provided corroborating evidence for borderline cases. The two raters achieved high inter-rater agreement, demonstrated by a pre-consensus Cohen kappa of 0.81, with senior arbitration required in only 12.9 percent of instances.

Signal preprocessing transformed the continuous multi-channel biosensor streams into structured tabular feature matrices while enforcing strict partition boundaries to eliminate data leakage. Continuous glucose monitoring data were aligned to a five-minute temporal grid and intentionally quarantined exclusively for downstream exploratory glycemic associations, ensuring that glucose values never served as input features for MFSI-Net. Wearable data streams from the smartwatch and smart ring were resampled to one-minute intervals using cubic-spline interpolation for short gaps under five minutes, while statistical outliers exceeding three median absolute deviations across rolling two-hour windows were Winsorized. Within non-overlapping five-minute windows, the team extracted between 250 and 300 candidate features spanning time-domain heart-rate variability metrics like SDNN and RMSSD, frequency-domain spectral powers including high-frequency and low-frequency components, respiratory-heart coupling indices, multi-day sleep architecture parameters, physical activity transitions, and peripheral skin-temperature amplitudes. To capture complex temporal dynamics without manual feature engineering, the team also incorporated a one-dimensional convolutional neural network trained under a contrastive learning paradigm with SimCLR augmentations on twenty-four-hour acceleration and temperature sequences, outputting a compact sixteen-dimensional representation of rest-activity rhythm embeddings.

To distill this high-dimensional feature set down to a stable, informative subset without introducing statistical bias, the investigators implemented a rigorous three-stage fold-contained feature-screening protocol nested inside outer leave-one-subject-out validation folds. In the initial stage, stability selection applied LASSO regression across one hundred bootstrap resamples of eighty percent of training participants, retaining only candidate variables demonstrating non-zero coefficients in at least eighty percent of resamples, which contracted the feature space from roughly three hundred down to 108 variables. In the second stage, the PC-stable conditional-independence discovery algorithm applied partial-correlation Fisher Z transformations to remove conditionally independent variables given observed training-set covariates, narrowing the active subset to sixty-one features. In the third stage, an extreme gradient-boosted decision-tree ensemble fitted to the training fold calculated global SHAP importance values, isolating the minimal set of features accounting for ninety-six percent of cumulative predictive importance, yielding a final core subset of thirty-four features. Subsequent post-hoc exploratory hierarchical and K-means clustering of this core subset identified four functional groupings: autonomic regulation (11 features), sleep-emotion modulation (9 features), circadian stability (7 features), and peripheral circulation-activity (7 features), which the authors define strictly as descriptive statistical clusters rather than validated physiological modules.

Because clinical evaluations took place only twice per day while wearable sensors gathered continuous streams, the research team employed Gaussian process regression to interpolate between the discrete tri-class labels, generating continuous, time-resolved soft supervision targets. The Gaussian process specification adopted a zero-mean function paired with a composite covariance kernel uniting a radial basis function kernel for flexible local smoothness and a periodic kernel with a fixed twenty-four-hour period to reflect diurnal cycles. Optimized purely on training participants via log-marginal likelihood maximization with randomized L-BFGS-B restarts, the resulting Gaussian process posterior supplied a continuous posterior mean trajectory alongside a time-varying posterior standard deviation. This model-based posterior uncertainty reached its lowest values in close proximity to actual expert assessments and widened during unobserved nocturnal intervals, peaking between 02:00 and 06:00 with an overall cohort mean of 0.14 and a maximum observed standard deviation of 0.38. The authors explicitly stress that this posterior variance reflects mathematical uncertainty under the chosen kernel assumptions rather than verified clinical margin of error, functioning as a continuous training target to supervise subsequent temporal deep-learning architectures.

The MFSI-Net deep-learning architecture processes sequential five-minute feature vectors through an offline, bidirectional temporal pipeline designed to capture both immediate physiological fluctuations and broader daily rhythms. The network first feeds screened feature vectors into a two-layer bidirectional gated recurrent unit with 128 hidden units, generating representations of local temporal context across intervals of minutes to hours. These recurrent hidden states are subsequently passed to a four-layer Transformer encoder utilizing eight attention heads, learnable positional encodings, and a 256-dimensional feed-forward layer to integrate contextual information across multi-day sequence spans. A fully connected linear projection layer with a hyperbolic tangent activation produces the final scalar continuous output constrained to the range of −1 to +1. MFSI-Net was trained using the Adam optimizer with weight decay, early stopping, and mini-batches of sixty-four windows against a compound loss function. This objective balanced mean squared error against the Gaussian-process soft labels with an auxiliary daily consistency penalty weighted by a fixed lambda parameter of 0.5, which was pre-specified on an independent fifteen-patient pilot cohort to encourage the network’s daily average predictions to align with original physician labels.

Rigorous leave-one-subject-out cross-validation across all ninety participants confirmed that MFSI-Net tracked Gaussian-process soft labels with strong technical fidelity, producing a mean absolute error of 0.13 alongside the root mean squared error of 0.17 and Pearson r of 0.72. Systematic architectural ablation experiments demonstrated that removing or replacing specific components reduced tracking accuracy: a two-layer GRU alone yielded an RMSE of 0.21 and an r of 0.58; a four-layer bidirectional GRU achieved an RMSE of 0.19 and an r of 0.63; a pure Transformer operating on raw features achieved an RMSE of 0.20 and an r of 0.59; and a bidirectional GRU paired with simple additive attention yielded an RMSE of 0.18 and an r of 0.67. Similarly, omitting individual screening layers degraded performance, with bypassing all feature selection causing the Pearson correlation to drop sharply to 0.54. In a temporal label-availability sensitivity test—where models were trained strictly on the initial eighty percent of monitoring days and evaluated against held-out physician labels across the final twenty percent—MFSI-Net maintained a continuous Pearson r of 0.61 and attained a multi-class categorization accuracy of 73.8 percent with an overall F1-score of 72.3 percent, demonstrating stability under restricted supervisory availability.

Mean individual MFSI scores correlated inversely with overall scores on the seven-item clinical symptom scale (Spearman rank correlation of −0.68, remaining at −0.61 after statistical adjustment for age, sex, and diabetes duration). Furthermore, non-parametric group comparisons showed marked separation between patients classified under different expert labels: median MFSI scores were −0.42 for under-regulation, +0.08 for normal regulation, and +0.31 for over-regulation, supported by a significant Jonckheere-Terpstra trend test. In parallel evaluations with continuous glucose monitor recordings, participants exhibiting under-regulation demonstrated significantly lower glycemic time in range (62.4 percent versus 71.8 percent for normal regulation) and higher glycemic coefficient of variation (31.2 percent versus 24.7 percent), associations that remained statistically significant after adjusting for baseline body mass index, glycated hemoglobin, and diabetes duration. Despite these promising statistical correlations, the investigators carefully delineate the study’s boundaries: the single-center design, modest sample size, post-hoc nature of feature clusters, and offline bidirectional structure mean the framework cannot currently serve as an acute real-time alert system, diagnostic replacement, or definitive measure of underlying metabolic physiology without prospective multi-center clinical validation.

Subject of Research: Continuous functional-state estimation from multisource wearable signals in type 2 diabetes

Article Title: MFSI-Net: a weakly supervised framework for continuous estimation of expert-annotated functional states from multisource wearable signals in type 2 diabetes

Article References: Yang, T., Xu, Y., & Wen, C. (2026). MFSI-Net: a weakly supervised framework for continuous estimation of expert-annotated functional states from multisource wearable signals in type 2 diabetes. Frontiers in Physiology, 17, Article 1948298. https://doi.org/10.3389/fphys.2026.1948298

Image Credits: AI Generated

DOI: 10.3389/fphys.2026.1948298

Keywords: type 2 diabetes, wearable devices, MFSI-Net, functional state assessment, weakly supervised learning, Gaussian process regression, GRU-Transformer, heart rate variability, skin temperature, continuous glucose monitoring, stability selection, PC-stable algorithm, SHAP feature importance, smartwatch sensing, smart ring sensors, accelerometry, observational study, digital health