Computers in biology and medicine最新文献

Magnetic hyperthermia therapy (MHT) in glioblastoma requires accurate modeling of nanoparticle transport and heat deposition across highly heterogeneous tumor regions. Traditional numerical approaches remain limited by high computational cost and sensitivity to complex tumor geometry, reducing their suitability for rapid clinical evaluation. To address these challenges, we introduce a genetically optimized physics-informed neural network (GA-PINN) that directly solves the bioheat transfer equation, while its governing parameters are dynamically coupled to nanoparticle transport, Darcy flow, and Arrhenius damage kinetics. Unlike prior PINN implementations, our approach integrates automatic genetic tuning of learning rates and loss-term weights, ensuring balanced convergence across coupled physics. Furthermore, tumor-focused collocation sampling uniquely enhances resolution of steep gradients near injection sites, a critical feature for patient-specific modeling. Results show that single-port injection restricts heating to the necrotic core, yielding central temperatures of 40 °C, 42.5 °C, and 48 °C for nanoparticle doses of 2.5, 5, and 10 kg/m³, respectively, but produces <10% necrosis in the viable rim. Increasing the magnetic field amplitude-frequency product to 8.4 × 10⁸ A/(m.s) raises peak temperatures to ∼47 °C and significantly accelerates damage accumulation. A multi-port injection strategy improves peripheral nanoparticle coverage, elevates rim temperatures to ∼41.5 °C, and reduces the dispersion index by more than 30%, indicating markedly more uniform ablation. These findings demonstrate that GA-PINN provides a stable, efficient, and physics-consistent surrogate for MHT, enabling rapid assessment and optimization of dosing conditions, magnetic field parameters, and multi-site injection strategies for patient-specific treatment planning.

Laparoscopy has revolutionised surgery, with faster recovery and less trauma for patients. However, extensive training is needed to gain the required skills, with said training ideally involving simulation-based strategies. A shortage of human expert coaches results in a lack of training opportunities in many jurisdictions. An Artificial Intelligence (AI) coach that can replicate the feedback from a human expert can be utilised to make quality surgical training more accessible. To be effective, this AI coach must classify complex movements into meaningful surgical gestures and then assess the trainee's performance on several parameters. We apply model soup, a recently developed machine learning approach, to improve gesture recognition and surgical score regression from kinematic simulator data. We also apply model soup to improve surgical gesture recognition from video data. By mixing several different 'ingredient' models into a final 'soup' model, performance is increased by 2.84-7.25 percentage points across three surgical tasks over state-of-the-art methods. These improved accuracies bring us closer to a fully autonomous laparoscopic surgery coach, with the potential to dramatically increase the availability of quality training, especially in environments where the availability of teachers and coaches is insufficient.

Attention Deficit Hyperactivity Disorder (ADHD) is a neurodevelopmental condition affecting mood, anxiety, learning, and sleep. Electroencephalogram (EEG) signals assist diagnosis, but challenges include complexity, nonlinearity, non-stationarity, overlapping patterns, and limited feature interpretability. To address these issues, Optimized Complex-Valued Spatio-Temporal Graph Convolutional Networks for ADHD Detection in Pediatric EEG Signals (c) is proposed. Input signals are collected from an EEG dataset for ADHD and an EEG dataset of children with Learning Disabilities (LD). Preprocessing is performed using a Multi-Window Savitzky-Golay Filter (MWSGF) to remove noise and artifacts, followed by the Synchro Transient Extracting Transform (STET) for extracting EEG channel features. These features are input into a Complex-Valued Spatio-Temporal Graph Convolutional Network (CSTGCN), classifying signals into ADHD or No-ADHD, and LD or No-LD. Red-Billed Blue Magpie Optimization Algorithm (RBBMO) is employed to optimize the network weights. Implemented in Python, the proposed CSTGCN-ADHD-EEG framework achieves 99.46% accuracy and 98.32% precision, outperforming existing models.

Subaortic stenosis, a heart disease characterised by a narrowing of the left ventricular outflow tract, is frequently caused by the presence of a subaortic membrane (SAM) located at the aortic valve inlet. This anatomical obstruction leads to significant haemodynamic alterations and leaflets fluttering, whose mechanisms are not yet fully understood. This research investigates, through computer simulations, the SAM's haemodynamic impact and the mechanism behind leaflets fluttering. A mono-physics fluid-structure interaction approach, based on the meshless smoothed particle hydrodynamics method, was employed. This approach represents both blood and structures with particles without defining interfaces, efficiently capturing large deformations and dynamic phenomena. Two common types of SAMs were investigated - a discrete thin SAM layer (flexible) and a thick fibromuscular ridge SAM (stiff) - and compared with a healthy aortic valve. Projected dynamic valve area (PDVA) was used as a reference parameter to quantify leaflet oscillation. While the PDVA in the healthy aortic valve stabilised at 283 mm² without oscillation, both pathological cases exhibited self-sustained periodic fluctuations. In the presence of discrete thin SAM layer, the mean PDVA decreased by 3% compared to the healthy control. This reduction was more pronounced for thick fibromuscular ridge configurations, where the mean PDVA was 9% lower than the healthy case. Notably, stiffer SAM configurations more than doubled the oscillation amplitude (from 3.12 mm² to 6.77 mm²) and increased the oscillation frequency by 8% relative to flexible membranes. Vortices dynamics was analysed, determining the phases of their formation, growth and migration. Through the analysis of velocity, vorticity, and shear stress maps, this study provides critical insights into the origin of fluttering and its influence over these key haemodynamic parameters. Findings demonstrate that the oscillatory leaflet motion is the result of vortices formation and shedding. The stiffness of the SAM significantly modulates the fluttering behaviour. While structural damage and haematological complications were not directly simulated, the identified oscillations represent haemodynamic conditions associated in literature with such pathologies. The observed alterations in wall shear stress magnitude and direction provide a physical basis for the mechanical environment that could contribute to endothelial cell dysfunction in the presence of SAM.

Electrocardiogram (ECG) reconstruction from reduced-lead configurations is essential for improving patient comfort and enabling wearable cardiac monitoring. Traditional reconstruction models (whether generic, population-specific, or patient-specific) require large datasets and extensive computational resources, limiting their practicality in real-world applications. This study investigates transfer learning as a strategy to overcome these limitations by enabling efficient personalisation of generic ECG reconstruction models. Three pipelines were evaluated: a linear regression model using leads I, II, V2, a wave-masked linear regression model using leads I, II, V3 (WMLR), and an ensemble of feed-forward neural networks. Generic models were trained on 10,000 normal ECG records from the CODE-15% dataset and fine-tuned using patient-specific data from the PTB-XL dataset. Performance was assessed using multiple metrics, including Pearson correlation, Dynamic Time Warping, Percentage Root-mean-square Difference, morphology similarity, and spectral similarity, across time intervals up to two years post-personalisation. Results show that transfer learning significantly improves reconstruction accuracy compared to generic models, with personalised models maintaining stable performance over extended periods. WMLR consistently outperformed other pipelines in correlation and morphological fidelity, demonstrating the continued relevance of linear approaches for resource-efficient deployment. While performance declined in cases where ECG morphology changed over time (like from normal to infarction), accuracy remained statistically acceptable, highlighting the robustness of transfer learning-based personalisation. These findings highlight the potential of transfer learning to enable scalable, long-term ECG monitoring systems that combine accuracy, adaptability, and computational efficiency.

This work introduces a machine learning enhanced finite element model for the study of blood flow in a stretching artery with the presence of sulfonated polystyrene nanoparticles (NSPS). The analysis deals with the coupling of the blood dynamics with nanoscale additives, which show significant promise for biomedical applications, such as drug delivery, oncological therapy, and targeted treatment based on magnetohydrodynamic mechanisms. The influence of nondimensional parameters (magnetic field strength, porosity, chemical reaction rate, and viscous dissipation) is studied to validate the usefulness of NSPS particles in clinical applications. Governing equations are solved by the finite element method, and numerical simulations are performed in the Python programming language to determine velocity, temperature, and concentration field. A multilayer perceptron regression surrogate is then trained on the finite element data and gives a predictive accuracy of 97.3%. on the numerical reference. This hybrid FEM-ML-based framework provides a fast computational tool for parameter space exploration, which helps to reduce the dependency on repeated, computationally expensive, finite element simulations. Numerical experimentation proves that the improvement of the magnetic parameter results in an approximate 18%. reduction in axial velocity, while improved porosity increases the drug penetration depth by approximately 11%. For hyperthermia-based treatments, localized heating could be used to release the drug from NSPS, whereas in pH-sensitive scenarios, the acidic tumor environment triggers automatic drug release. This approach ensures high localized action of the drugs, with the least amount of side effects on the healthy tissues.

Accurate infectious-disease forecasts are essential for timely public health decision-making. In this study, we develop a hybrid modeling framework that combines compartmental models with Long Short-Term Memory (LSTM) networks to estimate a key time-varying epidemiological parameter as a case study for leptospirosis in Thailand. Our framework uses an LSTM-ODE model trained on environmental covariates (rainfall, flooding, and temperature) and infected human cases to infer the transmission rate, which shows strong seasonal and environmental dependencies. The results demonstrate that including flooding, temperature, and human cases improves the prediction of infected individuals (MSE = 35.41). Our findings suggest that the integrated hybrid framework offers a more precise solution by improving the estimation of a key epidemiological parameter. The model accommodates multiple input features and, once trained, enables inference suitable for forecasting. Its ability to generate predictions using environmental covariates, particularly when epidemiological surveillance data are incomplete or delayed.

Background: Colorectal cancer (CRC) remains among the leading causes of cancer-related mortality worldwide. The TiaoPi AnChang Decoction (TPACD), a traditional Chinese herbal formula, has shown potential anti-CRC activity in preclinical studies (in vitro and in vivo); however, its active chemical basis and key target interactions remain unclear. The complex, multicomponent nature of TPACD poses significant challenges for systematically identifying and optimizing compounds.

Purpose: This study aimed to establish an intelligent molecular discovery framework that integrates a decoction-inspired genetic algorithm (EGA) with proximal policy optimization (PPO)-based reinforcement learning to simulate the dynamic transformation and recombination processes underlying traditional herbal decoctions and to generate and prioritize in silico candidate molecules for further anti-CRC evaluation.

Study design: A parallel molecular generation strategy was developed, comprising three algorithmic pathways, namely, the basic molecular generation algorithm (BMGA), enhanced molecular generation algorithm (EMGA), and intelligent molecular generation algorithm (IMGA), which represent progressive levels of structural diversity, biological relevance, and optimization capacity.

Methods: The EGA simulated the selection-transformation-recombination principles of decoction through adaptive mutation, fragment recombination, and a diversity-preserving selection procedure. PPO-based reinforcement learning further refined the molecular properties via a reward-guided exploration of the chemical space. The BMGA enabled baseline molecular generation; the EMGA incorporated affinity-based selection; and the IMGA achieved multiobjective optimization, integrating drug likeness, novelty, and toxicity filtering. The candidate molecules were validated by affinity prediction, QED scoring, and molecular docking.

Results: All three methods generated chemically valid and pharmacologically enhanced molecules relative to the original TPACD components. The EMGA provided improved biological affinity and exploration efficiency, whereas the IMGA achieved the highest overall performance level, with superior QED and docking stability, and predicted the CRC inhibition potential of the molecules.

Conclusion: The proposed EGA-PPO framework effectively connects traditional decoction principles with modern AI-based drug design. The tri-algorithmic system (BMGA-EMGA-IMGA) provides a scalable and interpretable strategy for de novo molecular discovery, offering promising leads for acquiring natural product-derived CRC therapeutics.

Because of their improved thermal conductivity, nanofluids are very useful in heat exchangers, solar collectors, and electronics cooling, among other cooling and heat transfer systems. Ternary hybrid nanofluids (THNF), which consist of three dissimilar types of nanoparticles, provide additional enhancement and multifunctional capabilities in fields including refrigeration systems, drug delivery, and enhanced oil recovery because of their adaptable thermophysical characteristics. In the current analysis, heat and mass transport in a tri-hybrid nanofluid flowing over a Riga plate is investigated using a generalized Fourier-Fick heat and mass flux model with variable relaxation times. Further, the influence of velocity and thermal slip boundary conditions, variable viscosity and thermal conductivity are considered. Through appropriate similarity transformations, the governing partial differential equations (PDEs) are changed into a system of ordinary differential equations (ODEs). The BVP4C solver in MATLAB is then used to numerically solve the resulting ODE system. The graphical results are obtained for ternary hybrid nanofluid (CNT-Al₂O₃-graphene) against various parameters. From the figure, it is perceived that enhancement values of velocity slip and variable viscosity parameter reduce the fluid velocity. Further, the variable thermal and solute relaxation parameters reduce the concentration and temperature fields.