To address the limitations of conventional physics-informed neural network (PINN) in handling hemodynamic boundary constraints, an improved hard boundary-constrained PINN (HBC-PINN) framework was proposed to achieve precise prediction of blood flow fields within stenotic arteries.

An idealized stenosed vessel geometry model was established and computational fluid dynamic simulation was performed to obtain a validation dataset. Appropriate boundary dependent trial functions were designed according to the hard constraint method to embed the flow boundary conditions into the network output. Thus, an HBC-PINN model with the hard boundary constraint method was constructed to predict the velocity field and pressure field of stenosed blood flow. Meanwhile, an original PINN model with the soft constraint method was also built for comparison. By evaluating the accuracy of the two models on the validation dataset, the capability of the HBC-PINN model to simulate hemodynamics without using any labeled data for training was verified.

The effectiveness of the HBC-PINN method in predicting hemodynamic parameters in stenosed blood flow tasks was validated. The relative L₂ errors of the flow velocity and pressure predicted by the HBC-PINN in two different stenosis scenarios were both lower than 0.5%, representing an improvement of over 48.8% in accuracy compared to the original PINN model. Additionally, the prediction accuracy of the transverse velocity also increased by more than 35.4%.

Implementing hard constraints on boundary conditions in the PINN modeling process can effectively improve the prediction accuracy of hemodynamic parameters and the efficiency of model solving.

To achieve non-invasive and precise prediction of mean arterial pressure (MAP) based on a fully convolutional neural network (FCNN).

A high-precision blood pressure data acquisition system compliant with international metrological standards was used in conjunction with the ‘gold standard’ auscultation method to collect blood pressure and pulse waveform data from patients. True MAP values were derived via Gaussian fitting of pulse waveform data, constructing a traceable dataset. The FCNN was applied to this dataset to develop a novel MAP prediction method. Additionally, the predictive accuracy of the FCNN was compared with linear regression and conventional empirical formulas.

The mean squared errors (MSE) for MAP prediction using the FCNN, linear regression, and empirical formulas were 19.76, 21.40, and 30.97, respectively. The coefficients of determination (R²) were 0.90, 0.89, and 0.84, and the prediction accuracies were 0.90, 0.89, and 0.85, respectively.

By using systolic blood pressure, diastolic blood pressure, age, and arm circumference as input parameters, the FCNN-based MAP prediction method significantly reduces the bias of empirical formulas. This approach not only improves the accuracy of hemodynamic boundary condition acquisition but also contributes to refining the metrological traceability system of non-invasive blood pressure measurement.

By applying the long short-term memory (LSTM) neural network model and using lower body landmark coordinates obtained from a markerless motion capture system as inputs, to estimate ground reaction force (GRF) curves during running.

The video images and GRF data of 59 amateur runners during running were collected by the markerless motion capture system and three-dimensional (3D) force plates. The LSTM model was established, and the 3D coordinates of 11 lower body landmarks, obtained via the Theia3D markerless system, were used as inputs to estimate the 3D GRF curves during the stance of running. The estimation performance was evaluated using correlation coefficients r, root mean square error (RMSE), and normalized root mean square error (nRMSE) by comparing LSTM model estimation and force plate measurement. Statistical parametric mapping was used to analyze differences in GRF curves estimated by the LSTM model and measured by the force plate, while paired t-tests were used to assess differences in GRF characteristics between model estimation and actual measurement.

A strong correlation (r>0.85, P<0.001) and lower error (RMSE<0.3 body weight, nRMSE<15%) was found between the LSTM model estimation and actual measurements. No significant difference was found in GRF curve intervals between LSTM model estimation and actual measurements. There was no significant difference in GRF characteristics between LSTM model estimation and actual measurements (P>0.05).

Based on the LSTM model, the 3D GRF curves can be effectively estimated by lower body landmark coordinates obtained from the makerless motion capture system, thereby acquiring the highly accurate GRF characteristics. The LSTM model developed in this study can be used to monitor injury risks during running in outdoor environments.

To explore the impact of vision impairment (VI) on the gait of hemiplegic patients, assess their walking ability and fall risks, and provide a basis for developing effective rehabilitation strategies.

Thirty hemiplegic patients were enrolled and stratified by the severity of visual acuity impairment into three groups (unimpaired, mildly impaired, and severely impaired). The gait data of patients under uncorrected vision were collected using the Qualisys motion capture system and the Kistler three-dimensional force platform, and the balance ability of patients was assessed simultaneously. Subsequently, the gait and assessment data were statistically analyzed to compare inter-group differences.

Compared with the visually unimpaired group, significant differences in step length, symmetry, and walking speed were observed in hemiplegic patients of the mild visual impairment group and severe visual impairment group. As VI increased, gait abnormalities became more pronounced, with a longer double-limb support phase, a longer swing phase of the affected limb, and a shorter single-limb support phase of the affected limb in the gait cycle. Compared with the visually unimpaired group, significant differences in center of pressure (COP) and COP symmetry were found between the mild visual impairment group and severe visual impairment group, with gait abnormalities intensifying. The Berg balance scale (BBS) scores showed that there was a significant difference between the visually unimpaired group and severe visual impairment group, indicating that the group with visual impairment had poorer balance ability.

VI has a significant negative impact on the gait and walking ability of hemiplegic patients. This study emphasizes the importance of focusing on the impact of VI in the rehabilitation of hemiplegic patients, with regular vision assessments and personalized interventions being conducted, which are of great significance in enhancing patients' walking quality.

To elucidate the regulatory effects of titanium surface modification on the immune function of immature dendritic cells (imDCs), different crystalline nanomorphologies were constructed on titanium surface to investigate the mechanobiological response of imDCs to nanomorphologies with different crystalline phases.

Nanomorphologies with different crystalline phases were constructed on the titanium surface by anodic oxidation and calcination. The changes of the cytoskeleton F-actin, cell adhesion and morphology of imDCs cultured on nanomorphologies with different crystalline phases were observed by fluorescence staining. The relative gene expression of adhesion molecules was detected by quantitative real-time PCR. The migration behaviors of imDCs were observed using real-time live-cell imaging, and the membrane fluidity was detected by fluorescence polarization.

Nanomorphologies with different crystalline phases, namely amorphous phase, anatase and rutile, were obtained on the titanium surface by anodic oxidation and calcination. The cytoskeleton of imDCs on nanomorphologies with different crystalline phases was remodeled. The spreading area of cells on anatase crystalline phase was relatively small, which was (353.3±148.5) μm². The number of adherent cells was the largest, which was 587±132. The expression of adhesion molecules such as CD11a, integrin β2, ICAM1, and VCAM1 were also increased in cells which cultured on anatase crystalline phase. The imDCs cultured on anatase crystalline phase were equipped with strong migration ability. The accumulative migration distance was (383.6±177.7) μm, and the Euclidean migration distance was (51.82±50.13) μm. The membrane fluidity was relatively weak, and the fluorescence polarization was 0.348 5±0.041 8.

imDCs can respond to nanomorphologies with different crystalline phases on the titanium surface and exhibit different biomechanical behaviors. The results might provide a theoretical basis for the design of titanium biomaterials with immunomodulatory functions.

To analyze the effects and differences of two veno-arterial extracorporeal membrane oxygenation (VA-ECMO) cannulation methods and subsequent left ventricular unloading on cardiac function and hemodynamics.

The lumped parameter model (LPM) of VA-ECMO integrated with the cardiovascular system in the MATLAB/Simulink environment was extended to simulate and analyze the changes in ventricular function and blood flow in the heart failure patient model under central VA-ECMO or peripheral VA-ECMO support. The effects of using arterial vasodilators or a left atrial drainage cannula on left ventricular function under central VA-ECMO support at a pump flow rate of 3 L/min were compared.

Under central VA-ECMO or peripheral VA-ECMO support, left ventricular pressure and volume increased, and stroke volume and ventricular work decreased. Both arterial vasodilators and the left atrial drainage cannula could reduce left ventricular pressure and volume. Arterial vasodilators additionally increased stroke volume and improved left ventricular ejection fraction from 11.6% to 19.5%.

Both VA-ECMO cannulation methods provide effective circulatory support in the heart failure patient model, with similar effects on ventricular function. Under central VA-ECMO support, arterial vasodilators can improve left ventricular function more effectively than the left atrial drainage cannula.

The biological characteristics and action mechanisms underlying the excellent performance of skeletal muscles were studied through experiments to provide a scientific basis for the development of flexible actuators with performance comparable to that of skeletal muscles.

A frog skeletal muscle sample was contracted by applying electrical stimulation, and then tensile load was applied to it to analyze the relationship between the driving properties (such as contraction length and output force) of skeletal muscle and its structure from three aspects: skeletal muscle dimensions, tendon, and epimysium.

The contraction lengths of these skeletal muscle samples were approximately 28.92% and 20% under unloaded conditions and under 50% of their maximum output force, respectively. When the load on the skeletal muscles did not exceed 20% of their maximum output force, they also exhibited the property of rapid reduction (approximately 1.25 s). The active tendon increased contraction by approximately 19.68% compared with the inactive tendon, and the integrity of the epimysium protected the force transfer efficiency of skeletal muscles.

By simulating the structural and biomechanical properties of skeletal muscles, flexible actuators can achieve better driving performance, thus greatly promoting the development of bionic robots.