Physical and Engineering Sciences in Medicine最新文献

Autism spectrum disorder (ASD) is one of the major neurological symptoms affecting young children. Most neurological diseases are captured through speech, voice and changes in sbrain activity. Research leading to ASD diagnosis is done in different ways; still, the early ASD diagnosis is a complex task. Various co-occurring situations may hinder Automated ASD detection, and deep learners effectively tackle such issues and create a better design. Here, a novel automated autism detection approach is proposed employing a deep learning technique with the help of brain image. Initially, the brain images are garnered from the standard dataset links. These gathered images are employed for the pre-processing stage, which is accomplished by using contrast enhancement. Subsequently, the most noteworthy deep features are extracted from the image pre-processed using a multi-atlas-based residual network (MResNet). Finally, the detection process is carried out by influencing the adaptive cascaded attention long short term memory with bayesian learning (ACAL-BL), in which some of the hyperparameters are tuned optimally by the random fixed marine predators algorithm (RFMPA). The performance is examined under Python using various factors and contrasted with other classical models and the results show that our ACAL-BL achieved an FPR of 4.5%, representing relative improvements of 52%, 54%, 56%, 58%, and 60% compared to LSTM, CNN, ANN, auto encoder, and LSTM-Bayesian learning, respectively. Thus, the suggested technique has the tendency to exploit the outstanding results that aid clinical practitioners to diagnose the disease earlier.

Work-related musculoskeletal disorders represent a significant occupational health issue. These disorders encompass a range of conditions resulting from specific risk factors associate to manual material handling such as: intensity, repetition, and duration. Over the years, several observational methodologies have been developed to assess biomechanical risk, but their limits depend mainly on clinicians' subjective assessment. For this reason, wearable sensors coupled with artificial intelligence have recently been integrated in the occupational ergonomic field. This study aimed to develop a new technological methodology-based on machine learning algorithms and inertial wearable sensors-able to automatically discriminate biomechanical risk associated with lifting loads. Ten healthy volunteers were enrolled in this study performing specific weight-lifting tasks wearing two inertial measurement units on the sternum and lumbar region. The acquired inertial signals were appropriately processed to extract several features in the time-domain and frequency-domain which have been used as input to several machine learning algorithms. Excellent results in discriminating biomechanical risk classes were obtained reaching accuracies and areas under the receiver operating characteristic curve above 86% and 95%, respectively. In addition, the sternum emerged as the most informative body landmark, while the mean absolute value was identified as the most informative feature. Future investigations on a larger study population could confirm the potential of the proposed automatic procedure to be used in the workplace in combination with well-established methodologies.

Nowadays, monitoring the health of elderly people at home or patients at the hospital on a regular basis is becoming necessary. Unfortunately, peer-to-peer treatment may require a longer time based on the availability of the doctors. In addition, it is practically impossible to go to hospitals for health checkups almost every day of the week. Hence, this research proposes an idea that can automate these processes without decreasing efficiency and reducing manual labor by integrating a healthcare system with the cyber layer to execute the automation processes. Previous text and image recognition studies used different machine learning and deep learning algorithms. However, in this study, an optical character recognition method ‛YOLO V8' is used, which provides a faster detection speed than other methods. The target was to retrofit biomedical devices such as blood pressure monitoring machines, digital thermometers, etc. using image processing techniques. To train the'YOLOv8' model, we have utilized two distinct image datasets that we have developed. The model showed an accuracy of 99.5% in detecting areas of concern on medical devices. Later, for recognition of values of different parameters from those devices a Convolutional Neural Network model is used, which confirms real-time validation employing 1000 images from different medical equipment. An accuracy of 99.7% has been achieved using this method. In the future, other medical devices such as heart rate monitors, pulse oximeters, etc. can be included in this system.

Ovarian cancer is often diagnosed at advanced stages, with high-grade serous ovarian cancer (HGSOC) accounting for 70-80% of fatalities. Current predictive tools, limited by single-time-point data, fail to capture subtle temporal changes indicative of relapse. To evaluate the performance of OvarXNet, a novel deep learning framework integrating longitudinal PET/CT imaging and clinical data for early prediction of ovarian cancer relapse. This retrospective study included 58 advanced-stage HGSOC patients (mean age, 56 ± 10.4 years) who underwent [¹⁸F]FDG PET/CT scans from April 2019 to January 2025. Patients with uncontrolled diabetes or recent cancers were excluded. Each patient had a median of three PET/CT scans and associated clinical data. The OvarXNet framework combines 3D convolutional neural networks (CNNs) for volumetric feature extraction and bidirectional gated recurrent units for temporal analysis. Statistical analyses included area under the receiver operating characteristic curve (AUC), precision-recall (PR) metrics, and calibration plots. Fifty-eight patients (mean age 56 ± 10.4 years) contributed 1914 image sets post-augmentation. OvarXNet achieved an AUC of 0.92, outperforming single-time-point CNN (AUC: 0.84) and LSTM-based models (AUC: 0.89). PR analysis confirmed superior model performance (PR-AUC: OvarXNet > 0.90 vs. single-time-point CNN: 0.82). Calibration plots demonstrated robust probability estimates. Attention mechanisms highlighted time points with elevated CA-125 or progression-related clinical notes, enhancing interpretability. OvarXNet significantly improves early relapse prediction in advanced-stage HGSOC by leveraging longitudinal imaging and clinical data. The framework's accuracy and interpretability support its potential for guiding personalized treatment strategies.

Telehealthcare is an evolving area that typically employs cloud-connected wireless biomedical gadgets for diagnosis, monitoring, and prognosis of diseases. In such environment, data compression, transmission, security and processing effectiveness are key issues. This paper proposes a new method for the automated diagnosis of arrhythmia in an efficient and effective manner. The proposed technique fuses a combination of Level-Crossing Analog-Digital Converters (LCADCs), Enhanced Activity Selection Algorithm (EASA), Adaptive-Rate Filtering (ARF), and ID-CNN. The electrocardiogram (ECG) signal is sampled by using the level-crossing concept. The QRS based segmentation and ARF with lower tap filters are realized. The denoised segments, without any handcrafted features extraction, are classified with one dimensional (1-D) deep convolutional neural network (CNN). Comparison is performed with using statistically extracted features in combination with CNN, existing state-of-the-art classical methods for ECG classification, and recent advanced deep learning models. The goal is to reach an efficient method by attaining a real-time data size reduction, computationally efficient signal preconditioning and a lower latency accurate classification. Five clinically important classes of arrhythmias, collected from the MIT-BIH dataset, are used to examine its applicability. Our experimental results show a 4.2-times diminishing in the count of acquired samples, on average, compared to conventional fix-rate counterparts. Similarly, data dimension reduction results in a more than 7.2-times computational effectiveness of the post denoising stage over the conventional counterparts. Moreover, classification latency is also significantly reduced while still achieving an accuracy rate of 99%.

This study investigated the potential of combining baseline ¹⁸F-FDG PET tumor-to-liver ratio radiomics with demographic data, using machine learning, to classify lymphoma subtypes and differentiate between candidates for ABVD and R-CHOP therapy. Additionally, we assessed whether nodal radiomics alone is sufficient for treatment and subtype classification. We conducted a multi-center study involving 241 lymphoma patients, including 125 with Non-Hodgkin lymphoma (NHL) and 116 with Hodgkin lymphoma. Among these, 94 had high-grade NHL, whereas 110 had classical Hodgkin lymphoma. We utilized 107 radiomic features, along with demographic data, such as age, stage, gender, and weight, to develop predictive models for classifying lymphoma subtypes and selecting treatment regimens (ABVD vs. R-CHOP). Data harmonization was performed using ComBat, feature selection was done with SelectKBest, and three machine learning models (Logistic Regression, Random Forest, and XGBoost) were trained with hyperparameter tuning, followed by external validation. For the best model in each classifier on the external test, adding extra-nodal radiomic features improved performance for certain lymphoma subtypes. For NHL vs. HL, accuracy increased from 0.807-0.819, whereas NHL precision rose from 0.837-0.875. High-grade NHL precision improved notably from 0.821-0.962. In treatment classification, extra-nodal features boosted accuracy for R-CHOP from 0.783-0.839 and increased F1-scores for both R-CHOP and ABVD. This study demonstrated the promise of PET radiomics combined with demographic features for lymphoma classification and treatment decision-making. Overall, extra-nodal features enhanced high-grade NHL and treatment classification but had minimal impact on other lymphoma subtypes.

Mechanical thrombectomy (MT) is an emergency treatment for acute ischemic stroke (AIS) to remove a clot occluding a large cerebral vessel. Histological analysis on retrieved thrombi have shown that they are mainly composed of red blood cells (RBCs), platelets and fibrin, and the outcome of MT appears to be influenced by clot composition. Therefore, being able to predict clot composition from routine medical images used for AIS diagnosis could support the choice of interventional strategy. Along with that, finite element simulations of the MT procedure can help provide insights into the impact of the procedural choices, the vessels morphology and the clot characteristics on the MT outcome. To achieve this, a realistic representation of the involved structures is necessary. In this context, this work aimed to (i) develop a methodology for the analysis of routine radiological images aiming at inferring information about clot characteristics (position, length, and composition) and (ii) develop a semi-automatic pipeline to position the clot in the patient-specific reconstructed geometry to build a patient-specific model which could be the starting point for the in silico replica of the MT procedure. However, image analysis alone could not distinguish between white and mixed clots, while a distinction between red and non-red clots was possible. Consequently, histological analyses were used to assign the clot composition, and thus the mechanical properties, in the positioning simulation. The resulting patient-specific model showed a strong similarity with pre-interventional clinical images.

Seismocardiography (SCG) uses accelerometers to record cardiac-induced accelerations of the chest wall. Cardiorespiratory interactions cause changes in amplitude and morphology of the SCG signals. Accelerometers can also directly monitor respiration by tracking thoracic inclination. This study thoroughly investigated the influence of accelerometer placement on the monitoring accuracy of respiration and cardiorespiratory interactions from SCG signals. Simultaneous recordings acquired by 16 accelerometers and a respiration belt placed onto 9 subjects' chests were analyzed. Respiratory signals were estimated considering: (a) chest inclination, (b) amplitude modulation (AM) and (c) morphological changes of SCG signals for each sensor location. For the first time in literature, a continuous description of respiratory-induced changes in SCG morphology was obtained via a morphological similarity index (MSi). The performance of respiratory acts detection and inter-breath intervals (IBIs) estimation was evaluated against the concurrent reference respiration signal. High accuracy was achieved in all three kinds of respiratory signals, with average sensitivity and positive predictive value of 95.8% and 95.5% for chest inclination, 85.9% and 84.4% for AM, 94.3% and 95.7% for MSi. Moreover, IBIs measurements showed non-significant biases and limits of agreement of about ± 0.8 s for chest inclination and MSi, and ± 1 s for AM. Performance achieved by chest inclination and MSi appeared not much influenced by sensor location, while AM showed higher variations. Information on breathing and cardiorespiratory interactions can be accurately obtained via SCG on multiple sites on the chest.

The study of dielectric properties (DPs) reveals significant differences between normal and liver cancer tissues. Although the open-ended coaxial probe (OCP) method is widely used for measuring DPs, tumor infiltration depth affects the measurements, blurring dielectric thresholds and posing challenges for tissue identification based on DPs. This study combines DPs analysis with machine learning (ML) to achieve two key goals: (1) accurately distinguish tissue types, (2) reliably predict tumor infiltration depth. We simulated the DPs of liver cancer tissues at different infiltration depths, using a total of 90,000 samples with 181 frequency-point features. We evaluated the performance of common ML models, including artificial neural networks (ANN), support vector machines (SVM), and Bagging tree ensembles, and validated them using real tissue and phantom measurements. Additionally, the probe's detection depth was experimentally validated. Experimental results showed that all three ML models performed well in tissue identification and tumor infiltration depth prediction. SVM achieved the highest classification accuracy of 98.91%. For depth prediction, SVM and ANN yielded MAPE/RMSE of 0.1742/0.0673 and 0.1658/0.0730, respectively. The probe's effective detection range was 0.1-0.6 mm, essential for accurate measurement and prediction. The models also demonstrated strong performance in real tissue and phantom validations, with the Bagging ensemble achieving 100% classification accuracy and MAPE/RMSE of 0.1434/0.0614 for prediction. These findings confirm the method's reliability for precise tissue identification and infiltration depth estimation, supporting accurate tumor resection and improved patient outcomes.