Qi Wang

Digital twins for physiological forecasting require robust mechanisms for continuously assimilating streaming data and adapting to patient-specific dynamics. Threshold-triggered model updates can restore predictive performance after substantial drift, but their reactive nature limits continuous adaptation. We propose a proactive digital twin framework that couples a pretrained PatchTST foundation model with an Actor–Critic reinforcement learning (RL) agent for residual correction. Using a 30-minute look-back window and recent residual history, the agent generates bounded corrections (±0.05) to the first-step baseline forecast, with corrected predictions recursively propagated over a 15- step horizon. We evaluate both an RL-only workflow and a hybrid workflow combining RL correction with threshold-triggered fine-tuning using minute-by-minute continuous glucose monitoring data from 10 critically ill patients with sepsis. Across all patients, the pretrained RL agent improves forecasting accuracy without patient-specific retraining. Both workflows reduce trajectory-averaged mean absolute percentage error, ensemble relative error, and maximum errors compared with the uncorrected baseline and previously developed reactive digital twin models. These findings support residual reinforcement learning as a lightweight mechanism for continuous, real-time data assimilation that preserves the foundation model’s global predictive prior while enabling patient-specific forecast adaptation, with potential applications in critical care decision support.

We present a digital twin framework for real-time glucose monitoring and forecasting in septic patients admitted to intensive care units (ICUs). The framework integrates advanced machine learning models trained on continuous glucose measurements with a dynamic digital twin workflow that enables rapid deployment to individual patients, ongoing model assessment, updating with newly acquired data, and personalized predictive decision support. Built on a foundation model—a pretrained time-series transformer—the digital twin adapts as new patient data become available and generates rolling near-term glucose forecasts in real time. To evaluate adaptability and computational efficiency, we deployed the pretrained model to ten septic patients and compared multiple personalization strategies, including zero-shot inference, linear probing, full fine-tuning, and staged fine-tuning. The results show that the model can be initialized and personalized for a new patient within seconds on a standard laptop while maintaining accurate glucose forecasts across varying data conditions. These findings demonstrate the feasibility of real-time digital twin personalization in resource-constrained, high-acuity clinical environments and highlight the potential of digital twins as scalable, AI-enabled platforms for continuous physiological monitoring, clinical decision support, and individualized treatment design in the ICU.

We formulate the problem of estimating hydraulic conductivity from measured hydraulic head data in groundwater flow within the unified framework of parameter estimation using the extended Kalman filter (EKF) and least-squares optimization. We first formulate the EKF and its variants as problems of least-squares optimization. We then show that the successive linear estimator can be interpreted as a variant of the EKF, with its initial estimate obtained through co-Kriging. Furthermore, by showing that the EKF is equivalent to a one-step Newton iteration for solving a least-squares problem, we reformulate the hydraulic conductivity estimation problem as an associated least-squares optimization problem. Finally, we present an algorithm for estimating hydraulic conductivity from unsaturated soil data.

This manuscript reports a mixed‐spacer architecture for resolving the conductivity–swelling–stability tradeoff in ether‐free anion‐exchange membranes (AEMs). Unlike conventional spacer engineering, which usually relies on a single spacer length, the mixed‐spacer approach provides an additional degree of freedom for regulating membrane properties. Using mixed‐spacer engineering in polycarbazoles with mixed short and long spacer side chains, we show that hydroxide transport and membrane durability can be jointly optimized through control of hydration, ionic connectivity, and local cation solvation, rather than through ion exchange capacity alone. A series of membranes with varied ratios of short ( C4 ) and long ( C6 ) alkyl spacers were synthesized via acid‐catalyzed Friedel–Crafts polymerization followed by quaternization. The optimized membrane achieves a hydroxide conductivity of 184 mS cm ⁻¹ at 80°C, outperforming both short‐spacer and long‐spacer counterparts. Membranes rich in C4 spacers exhibit weaker alkaline stability, whereas those containing dominant C6 spacers show excellent stability. In a fuel cell device, the membrane delivers a peak power density of 2.51 W cm ⁻² , among the highest reported for polycarbazole AEMs. Mechanistic understanding was achieved through both experimental and computational investigations. These results demonstrate that mixed‐spacer molecular engineering is a promising and effective strategy for designing high‐performance AEMs.

This manuscript reports a mixed‐spacer architecture for resolving the conductivity–swelling–stability tradeoff in ether‐free anion‐exchange membranes (AEMs). Unlike conventional spacer engineering, which usually relies on a single spacer length, the mixed‐spacer approach provides an additional degree of freedom for regulating membrane properties. Using mixed‐spacer engineering in polycarbazoles with mixed short and long spacer side chains, we show that hydroxide transport and membrane durability can be jointly optimized through control of hydration, ionic connectivity, and local cation solvation, rather than through ion exchange capacity alone. A series of membranes with varied ratios of short ( C4 ) and long ( C6 ) alkyl spacers were synthesized via acid‐catalyzed Friedel–Crafts polymerization followed by quaternization. The optimized membrane achieves a hydroxide conductivity of 184 mS cm ⁻¹ at 80°C, outperforming both short‐spacer and long‐spacer counterparts. Membranes rich in C4 spacers exhibit weaker alkaline stability, whereas those containing dominant C6 spacers show excellent stability. In a fuel cell device, the membrane delivers a peak power density of 2.51 W cm ⁻² , among the highest reported for polycarbazole AEMs. Mechanistic understanding was achieved through both experimental and computational investigations. These results demonstrate that mixed‐spacer molecular engineering is a promising and effective strategy for designing high‐performance AEMs.

We formulate the problem of estimating hydraulic conductivity from measured hydraulic-head data in groundwater flow within the unified framework of parameter estimation using the extended Kalman filter (EKF) and least-squares optimization. We first formulate the EKF and its variants as problems of least-squares optimization. We then show that the successive linear estimator can be interpreted as a variant of the EKF, with its initial estimate obtained through Co-Kriging. Furthermore, by showing that the EKF is equivalent to a one-step Newton iteration for solving a least-squares problem, we reformulate the hydraulic-conductivity estimation problem as an associated least-squares optimization problem. Finally, we present an algorithm for estimating hydraulic conductivity from unsaturated soil data.

Karst hydrological processes are driven by seasonal precipitation and evaporation and modulated by anthropogenic activities within highly heterogeneous aquifers. These processes produce spatiotemporally differentiated and magnitude-dependent responses that remain difficult to simulate and interpret. This study develops an LSTM-Transformer Decoders (LSTM-TDs) framework that links discharge-difference prediction with attention-based hydroclimatic sensitivity identification. To characterize discharge-change dynamics, the LSTM encodes historical first-order discharge differences into a discharge-change state, which then serves as the query for two spatiotemporal Transformer Decoder pathways separately attending to precipitation and evaporation inputs. The resulting attention structures are used to identify meteorological relevance across space, lag, and forcing magnitude. Applied to the Niangziguan Springs in North China, LSTM-TDs outperformed the evaluated benchmark models, with NSE = 0.85, RMSE = 0.34 m³/s, MAE = 0.26 m³/s, and MAPE = 3.88%. Relative to the non-differenced version, the discharge-difference design increased testing NSE from 0.79 to 0.85 and reduced MAPE by 19.8%. Attention-based sensitivity identification showed hydrologically plausible spatial and temporal patterns. Precipitation relevance was mainly concentrated in the main recharge–runoff area with relatively high groundwater levels, evaporation relevance was stronger in soil-moisture-rich areas, and dominant lags were consistent with cross-correlation-derived response times. The model-derived sensitivity turning points were approximately 42.0 mm for precipitation and 130.2 mm for evaporation, and their spatial variations were consistent with differences in infiltration partitioning, terrain-controlled redistribution, and aquifer storage conditions. These results indicate that the LSTM-TDs framework can support spring discharge prediction while providing interpretable diagnostics of hydroclimatic sensitivity in karst systems.

BackgroundSchizophrenia is a chronic mental disorder that requires lifelong treatment. Medication nonadherence contributes to poor medical outcomes; the ability to predict a patient’s adherence trajectory cannot be understated in forestalling the negative effects of nonadherence. This study aimed to identify antipsychotic adherence latent classes among patients diagnosed with schizophrenia or schizoaffective disorder and evaluate predictors of adherence latent classes. Healthcare utilization, including hospitalizations and emergency department visits across adherence trajectories, was also evaluated.MethodsThis study utilized South Carolina Medicaid claims for beneficiaries diagnosed with schizophrenia or schizoaffective disorder from 2014 to 2020 using a latent class mixture modeling framework. This consists of two submodels: a membership model that categorizes patients into adherence classes and a longitudinal model that characterizes adherence trajectories over time within each adherence class.ResultsA total of 2,870 eligible patients were included in the study, with 1,707 males and 1,163 females. Four distinct adherence trajectories were identified: most adherent (56% of the population), early drop-off (12%), initial adherence followed by an increase (8%), least adherent (24%). The impact of predictors and frequency of healthcare utilization outcomes varied across classes. Time consistently demonstrated a significant impact on trajectory shapes.ConclusionFour distinct trajectories for adherence were identified among South Carolina’s Medicaid beneficiaries. Significant patient characteristics that determined class memberships and adherence trajectories varied across classes.

We present a digital twin framework for real-time glucose monitoring and forecasting in septic patients in intensive care units (ICUs). The framework combines advanced machine learning models trained on continuous glucose measurements with a dynamic transfer-learning workflow that enables rapid deployment to individual patients and supports personalized, adaptive, and predictive clinical decision-making. Built on a foundation model—a pretrained time-series transformer—the digital twin continuously updates its parameters as new patient data arrive and produces rolling near-term forecasts in real time. To assess adaptability and computational efficiency, we deployed the pretrained model to ten septic patients and evaluated multiple retraining strategies, including zero-shot inference, linear probing, and full and staged fine-tuning. Results show that the model can be initialized and personalized for a new patient within seconds on a standard laptop while achieving accurate glucose forecasts under varying data conditions. These findings demonstrate the feasibility of real-time model personalization in resource-constrained, high-acuity environments and highlight the potential of digital twins as scalable, AI-enabled platforms for continuous physiological monitoring, clinical decision support, and individualized treatment design in the ICU.