Irbm最新文献

Objectives

Electrocardiogram (ECG) signals are beneficial for diagnosing cardiac diseases. The cardiac patients' life quality likely increases with continuous or long-period recording and monitoring of ECG signals, leading to better and early diagnosis of disease and heart attacks. However, continuous ECG recording necessitates high data rates and storage, which means high costs. Therefore, ECG compression is a handy concept that facilitates continuous monitoring of ECG signals. Deep neural networks open up new horizons for compression and also for ECG compression by providing high compression rates and quality. Although they bring constant compression ratios with better average quality, the compression quality of individual samples is not guaranteed, which may lead to misdiagnoses. This study aims to investigate the effect of compression quality on the diagnoses and to develop a deep neural network-based compression strategy that guarantees a quality-bound in return for varying compression ratios.

Materials and methods

The effect of the compression quality on the arrhythmia diagnoses is tested by comparing the performance of the deep learning-based ECG classifier on the original ECG recordings and the distorted recordings using a lossy compression algorithm with different compression error levels. Then, a compression error upper limit is calculated in terms of normalized percent root mean square difference (PRDN) error, which also coincides with the findings of the previous studies in the literature. Lastly, to enable deep learning in ECG compression, a single encoder-multi-decoder convolutional autoencoder architecture, and multiple quantization levels are proposed to guarantee a desired upper limit on the error rate.

Results

The efficiency of the proposed method is demonstrated on a popular benchmark data set for ECG compression methods using a transfer learning approach. The PRDN error is fixed to various values, and the average compression rates are reported. An average of $13.019:1$ compression is achieved for a 10% PRDN error rate, assessed as a fair quality threshold for reconstruction error. It has also been shown that the compression model has a runtime that can be run in real-time on wearable devices such as commercial smartwatches.

Conclusion

This study proposes a deep learning-based ECG compression algorithm that guarantees a desired upper limit on the compression error. This model may facilitate an eHealth solution for continuous monitoring of ECG signals of individuals, especially cardiac patients.

Objectives: The nociceptive flexion reflex (NFR) is used as a pseudo-objective measure of pain that is measured using electromyography (EMG). EMG signals can be analyzed using nonlinear methods to identify complex changes in physiological systems. Physiological complexity has been shown to allow a wider range of adaptable states for the system to deal with stressors. The purpose of this study was to examine changes in complexity and entropy of EMG signals from the biceps femoris during non-noxious stimuli and noxious stimuli that evoked the NFR before and after acute inflammation. Methods and Materials: Twelve healthy participants (25.17y ± 3.43) underwent the NFR protocol. EMG signal complexity was calculated using Hurst Exponent (H), determinism (DET), and recurrence rate (RR), and Sample Entropy (SampEn). Results: RR (∼200%), DET (∼70%), and H (∼35%) were higher and SampEn was reduced (∼50%) during the noxious stimulus that evoked the NFR compared to non-noxious stimuli. No significant differences were found for any of the complexity and entropy measures before and after exercise-induced inflammation ($p<0.05$). Reduced complexity (increased H, DET, and RR) and increased regularity (SampEn) reflect reduced adaptability to stressors. Conclusions: Nonlinear methods such as complexity and entropy measures could be useful in understanding how a healthy neuromuscular system responds to disturbances. The reductions in complexity following a noxious stimulus could reflect the neuromuscular system adapting to environmental conditions to prevent damage or injury to the body.

Objective

In ophthalmology, there is a need to explore the relationships between clinical parameters of the cornea and the corneal shape. This study explores the paradigm of machine learning with nonlinear regression methods to verify whether corneal shapes can effectively be predicted from clinical data only, in an attempt to better assess and visualize their effects on the corneal shape.

Methods

The dimensionality of a database of normal anterior corneal surfaces was first reduced by Zernike modeling into short vectors of 12 to 20 coefficients used as targets. The associated structural, refractive and demographic corneal parameters were used as predictors. The nonlinear regression methods were borrowed from the scikit-learn library. All possible regression models (method + predictors + targets) were pre-tested in an exploratory step and those that performed better than linear regression were fully tested with 10-fold validation. The best model was selected based on mean RMSE scores measuring the distance between the predicted corneal surfaces of a model and the raw (non-modeled) true surfaces. The quality of the best model's predictions was visually assessed thanks to atlases of average elevation maps that displayed the centroids of the predicted and true surfaces on a number of clinical variables.

Results

The best model identified was gradient boosting regression using all available clinical parameters to predict 16 Zernike coefficients. The predicted and true corneal surfaces represented in average elevation maps were remarkably similar. The most explicative predictor was the radius of the best-fit sphere, and departures from that sphere were mostly explained by the eye side and by refractive parameters (axis and cylinder).

Conclusion

It is possible to make a reasonably good prediction of the normal corneal shape solely from a set of clinical parameters. In so doing, one can visualize their effects on the corneal shape and identify its most important contributors.

Background and Objective

Utilizing artificial intelligence (AI), a clinical decision support system (CDSS), can help physicians anticipate possible complications of cirrhosis patients before prescribing more accurate treatments. This study aimed to establish a prototype of AI-CDSS modeling using electronic health records to predict five complications for cirrhosis patients who were controlled for oral antiviral drugs, lamivudine (LAM) or entecavir (ETV).

Methods

Our modeling attained a web-based AI-CDSS with four steps – data extraction, sample normalization, AI-enabled machine learning (ML), and system integration. We designed the extract-transform-load (ETL) procedure to filter the analytics features from a clinical database. The data training process applied 10-fold cross-validation to verify diverse ML models due to possible feature patterns with medications for predicting the complications. In addition, we applied both statistical means and standard deviations of the realistic datasets to create the simulative datasets, which contained sufficient and balanced data to train the most efficient models for evaluation. The modeling combined multiple ML methods, such as support vector machine (SVM), random forest (RF), extreme gradient boosting, naive Bayes, and logistic regression, for training fourteen features to generate the AI-CDSS's prediction functionality.

Results

The models achieving an accuracy of 0.8 after cross-validations would be qualified for the AI-CDSS. SVM and RF models using realistic data predicted jaundice with an accuracy of over 0.82. Furthermore, the SVM models using simulative data reached an accuracy of over 0.85 when predicting patients with jaundice. Our approaches implied that the simulative datasets based on the same distributions as that of the features in the realistic dataset were adequate for training the ML models. The RF model could reach an AUC of up to 0.82 for multiple complications by testing with the un-trained data. Finally, we successfully installed twenty models of the suitable ML methods in the AI-CDSS to predict five complications for cirrhosis patients prescribed with LAM or ETV.

Conclusions

Our modeling integrated a self-developed AI-CDSS with the approved ML models to predict cirrhosis complications for aiding clinical decision making.

Accurate and timely injection of insulin doses in accordance with the treatment protocol is very important in the follow-up of insulin-dependent diabetes patients. In this study, a new smart mobile apparatus (SMA) has been developed. The SMA can be attached to insulin pens and record and transfer data by detecting the patient's dose of insulin and the time at which it was provided. The SMA can detect the dose determined in the insulin pen through linear capacitive sensors. Electronic parts and sensor mechanism are located on the designed SMA body. The insulin pen's two-part mechanical construction of the body senses movement during dosage adjustment while also making sure the dose information is recorded in the control unit. The dose and time information recorded in the SMA internal memory are transmitted to the patient's smartphone via the developed mobile application. The developed SMA prototypes were evaluated by a team of doctors in a hospital setting for three months. As a result of the three-month study, it was observed that the insulin dose and administration times could be accurately sent to the smartphone application via SMA. The SMA was created in the laboratory environment and was prepared for pilot research with insulin-dependent diabetes patients in a hospital setting. It was observed that the SMA prototype successfully identified and recorded the dose and timing of the patient's self-administered insulin.

Introduction

Aging is associated with muscle decline, which alters both functional and anatomical properties of the neuromuscular system. These modifications can be reflected in high-density surface electromyography (HD-sEMG) signals. This study examines how age and sex impact the shape of the amplitude Probability Density Function (PDF) of HD-sEMG signals.

Materials and Methods

Monopolar HD-sEMG signals were collected from the Biceps Brachii in a cohort of 17 individuals: 10 women (mean age: 22.9 ± 3.6 years) and 7 men (mean age: 24.4 ± 2.5 years) in the younger group, and 10 women (mean age: 69.8 ± 4.8 years) and 7 men (mean age: 72.8 ± 2.7 years) in the elderly group. The recordings were conducted during an elbow flexion at both 20% and 40% maximum voluntary contraction. The signal amplitude was evaluated using root means square amplitude (RMSA) and the PDF shape of each HD-sEMG signal was assessed through skewness, excess Kurtosis, and robust functional statistics. These shape distance metrics evaluate the departure from Gaussianity related to muscle aging. a) We conducted a comparison study of the HD-sEMG PDF shapes between younger and elderly individuals. b) Evaluating differences between men and women. c) Considering monopolar and Laplacian electrode configurations that are sensitive to different muscle regions.

Results

A) The HD-sEMG PDFs of elderly subjects demonstrated a lower departure from Gaussianity than their younger counterparts. B) Women exhibited lower RMSA values than men, and, on average, a lower departure from Gaussianity whatever the age and contraction level C) Trends of departure from Gaussianity with contraction level, seems to be influenced by the electrode configuration. In fact, a decrease in Gaussianity departure is observed with monopolar recordings where an increase is observed with Laplacian one, clearly indicating different muscle region assessment.

Discussion

The findings highlight the influence of factors such aging, sex, contraction level and electrode montage on the shape of the HD-sEMG PDF, emphasizing the significance of using this descriptor for monitoring and better assessment of muscle aging.

Context

Deep learning algorithms have been widely used for cardiac image segmentation. However, most of these architectures rely on convolutions that hardly model long-range dependencies, limiting their ability to extract contextual information. Moreover, the traditional U-net architecture suffers from the difference of semantic information between feature maps of the encoder and decoder (also known as the semantic gap).

Material and method

To address this issue, a new network architecture relying on attention mechanism was introduced. Swin Filtering Blocks (SFB), that use Swin Transformer blocks in a cross-attention manner, were added between the encoder and the decoder to filter information coming from the encoder based on the feature map from the decoder. Attention was also employed at the lowest resolution in the form of a transformer layer to increase the receptive field of the network.

We conducted experiments to assess both generalization capability and to evaluate how training on all frames of the cardiac cycle rather than only the end-diastole and end-systole impacts strain and segmentation performances.

Results and conclusion

Visual inspection of feature maps suggested that Swin Filtering Blocks contribute to the reduction of the semantic gap. Performing attention between all patches using a transformer layer brought higher performance than convolutions. Training the model with all phases of the cardiac cycle resulted in slightly more accurate segmentations while leading to a more noticeable improvement for strain estimation. A limited decrease in performance was observed when testing on out-of-distribution data, but the gap widens for the most apical slices.

Background

Emotion assessment plays a vital role in understanding and enhancing various aspects of human life, from mental well-being and social interactions to decision-making processes. Electrodermal Activity (EDA) is widely used for emotion assessment, as it is highly sensitive to sympathetic nervous system activity. While numerous existing approaches are available for EDA-based emotion assessment, they often fall short in capturing the dynamic non-linear variations and time-varying characteristics of EDA. These limitations hinder their effectiveness in accurately classifying emotional states along the Arousal and Valence dimensions. This study aims to address these shortcomings by introducing Transition Network Analysis (TNA) as a novel approach to EDA-based emotion assessment.

Methods

To explore the dynamic non-linear variations in EDA and their impact on the classification of Arousal and Valence dimensions, we decomposed EDA data into its phasic and tonic components. The phasic information is represented over a transition network. From the transition network, we extracted seven features. These features were subsequently used for classification purposes employing four different machine learning classifiers: logistic regression, multi-layer perceptron, random forest, and support vector machine (SVM). The performance of each classifier was evaluated using Leave-One-Subject-Out cross-validation. The study evaluated the performance of these classifiers in characterizing emotional dimensions.

Results

The results of this research reveal significant variations in Degree Centrality and Closeness Centrality within the transition network features, enabling effective characterization of Arousal and Valence dimensions. Among the classifiers, the SVM achieved F1 scores of 71% and 72% for Arousal and Valence classification, respectively.

Significance

This study holds significant implications as it not only enhances our understanding of EDA's non-linear dynamics but also demonstrates the potential of TNA in addressing the limitations of existing techniques for EDA-based emotion assessment. The findings open exciting opportunities for the advancement of wearable EDA monitoring devices in naturalistic settings, bridging a critical gap in the field of affective computing. Furthermore, this research underlines the importance of recognizing the limitations in current EDA-based emotion assessment methods and suggests an innovative path forward in the pursuit of more accurate and comprehensive emotional state classification.

Objective

This study aims to evaluate the impact of image factors on the performance of deep learning models used for ophthalmic ultrasound image diagnosis.

Methods

A total of 3,373 ophthalmic ultrasound images are used to deeply evaluate the influence of image factors on the performance of deep learning classification models. Inceptionv3, Xception, and the fusion model Inceptionv3-Xception are used to explore how brightness, contrast, gain, noise, size, format, pseudo-color seven image-related factors affect the classification performance of the model.

Results

Inceptionv3-Xception has advantages in the recognition accuracy of various image factors. When the image brightness changes, the model's performance shows a downward trend (0.5 vs. 1 vs. 1.8, ACC 95.73 vs. 97.06 vs. 93.54, P < 0.05). When the image contrast changes, the model's performance is comparable (0.5 vs. 1 vs. 1.2, ACC 96.23 vs. 96.95 vs. 97.45, P > 0.05). When the image gain drops to 50 dB, the model's accuracy decreases significantly (50 dB vs. 105 dB, ACC 96.49 vs. 97.57, P < 0.05). When Gaussian noise is added to the image, the model's performance gradually decreases (0.02 vs. 0, ACC 89.48vs97.06, P < 0.05). When the image size drops to 25% of the original image, the model's performance decreases significantly (25% vs. 100%, ACC 93.18 vs. 97.06, P < 0.01). When the image format changes, the model's recognition accuracy is similar (JPG vs. BMP vs. PNG, ACC 96.98 vs. 97.06 vs. 97.06, P > 0.05). The accuracy of the model in recognizing pseudo-color images decreases significantly compared to grayscale images (grayscale vs. pseudo-color, ACC 35.96 vs. 97.06).

Conclusion

These results indicate that image quality greatly influences the model training process, and acquiring high-quality images is an important prerequisite for high recognition performance of the model. This study offers valuable insights for the improvement of other robust deep learning models for ophthalmic ultrasound image recognition.

Objectives

Segmentation of cranial bones in magnetic resonance images (MRIs) is a challenging and indispensable task to study neonatal brain development and injury. This paper presents a new approach for creating subject-specific probability maps of the scalp, skull and cerebrospinal fluid (CSF) from retrospective bimodal (MR and CT) images acquired from neonates in the gestational age range of 39 to 42 weeks. These maps are subsequently employed for the segmentation of cranial bones in cerebral MRIs from neonates in the same age range.

Material and methods

Retrospective MR and CT of neonates with normal head in the gestational age range of 39-42 weeks were preprocessed, segmented semi-automatically and employed as atlas data. For an input MR image acquired from a subject under study, a preprocessing stage and three main processing blocks were performed: First, subject-specific head and intracranial templates and CSF probability map were created using retrospective MR atlas data. Second, the CT atlas data were coregistered to MR templates and the resulted deformation matrices were fed to the next block to create subject-specific scalp and skull probability maps. Finally, some novel performance measures were presented to evaluate the performance of subject-specific CSF, scalp and skull probability maps for skull and intracranial segmentation in neonatal MRIs.

Results

The subject-specific probability maps were employed for brain tissue extraction and compared with two public methods such as Brain Extraction Tool (BET) and Brain Surface Extractor (BSE). They were also applied for cranial bone extraction. Then, the similarity in shape between the frontal and occipital sutures (which had been reconstructed from segmented cranial bones) and the ground truth landmarks was evaluated. For this purpose, modified versions of the Dice similarity coefficient (DSC) were used. Finally, a retrospective bimodal (MR-CT) data acquired from a neonate within a short time interval was used for evaluation. After co-alignment of the two images, the DSC and modified Hausdorff distance (MHD) were used to compare the similarity of cranial bones in the MR and CT images.

Conclusion

Significant improvements were achieved compared to conventional methods which rely solely on MR image intensities. These advancements hold promise for enhancing neurodevelopmental studies in neonates. The algorithm for creating subject-specific atlases is publicly accessible through a graphical user interface at medvispy.ee.kntu.ac.ir.