Long short-term memory (LSTM) recurrent neural network for muscle activity detection

Identification of Escherichia coli strains using MALDI-TOF MS combined with long short-term memory neural networks

The current study aims to develop a new technique for the precise identification of Escherichia coli strains, utilizing matrix-assisted laser desorption ionization time-of-flight mass spectrometry (MALDI-TOF MS) combined with a long short-term memory (LSTM) neural network. A total of 48 Escherichia coli strains were isolated and cultured on tryptic soy agar medium for 24 hours for the generation of MALDI-TOF MS spectra. Eight hundred MALDI-TOF MS spectra were obtained per strain, resulting in a database of 38,400 spectra. Fifty percent of the data was utilized for LSTM neural network training, with fine-tuned parameters for strain-level identification. The other half served as the test set to assess model performance. Traditional PCA dimension reduction of MALDI-TOF MS spectra indicated 47 out of 48 strains to be unclassifiable. In contrast, the LSTM neural network demonstrated remarkable efficacy. After 20 training epochs, the model achieved a loss value of 0.0524, an accuracy of 0.999, a precision of 0.985, and a recall of 0.982. When tested on the unseen data, the model attained an overall accuracy of 92.24%. The integration of MALDI-TOF MS and LSTM neural network markedly enhances the identification of Escherichia coli strains. This innovative approach offers an effective and accurate tool for MALDI-TOF MS-based strain-level identification, thus expanding the analytical capabilities of microbial diagnostics.

Deep Learning Model Coupling Wearable Bioelectric and Mechanical Sensors for Refined Muscle Strength Assessment

Muscle strength (MS) is related to our neural and muscle systems, essential for clinical diagnosis and rehabilitation evaluation. Although emerging wearable technology seems promising for MS assessment, problems still exist, including inaccuracy, spatiotemporal differences, and analyzing methods. In this study, we propose a wearable device consisting of myoelectric and strain sensors, synchronously acquiring surface electromyography and mechanical signals at the same spot during muscle activities, and then employ a deep learning model based on temporal convolutional network (TCN) + Transformer (Tcnformer), achieving accurate grading and prediction of MS. Moreover, by combining with deep clustering, named Tcnformer deep cluster (TDC), we further obtain a 25-level classification for MS assessment, refining the conventional 5 levels. Quantification and validation showcase a patient's postoperative recovery from level 3.2 to level 3.6 in the first few days after surgery. We anticipate that this system will importantly advance precise MS assessment, potentially improving relevant clinical diagnosis and rehabilitation outcomes.

Review of electromyography onset detection methods for real-time control of robotic exoskeletons

BACKGROUND

Electromyography (EMG) is a classical technique used to record electrical activity associated with muscle contraction and is widely applied in Biomechanics, Biomedical Engineering, Neuroscience and Rehabilitation Robotics. Determining muscle activation onset timing, which can be used to infer movement intention and trigger prostheses and robotic exoskeletons, is still a big challenge. The main goal of this paper was to perform a review of the state-of-the-art of EMG onset detection methods. Moreover, we compared the performance of the most commonly used methods on experimental EMG data.

METHODS

A total of 156 papers published until March 2022 were included in the review. The papers were analyzed in terms of application domain, pre-processing method and EMG onset detection method. The three most commonly used methods [Single (ST), Double (DT) and Adaptive Threshold (AT)] were applied offline on experimental intramuscular and surface EMG signals obtained during contractions of ankle and knee joint muscles.

RESULTS

Threshold-based methods are still the most commonly used to detect EMG onset. Compared to ST and AT, DT required more processing time and, therefore, increased onset timing detection, when applied on experimental data. The accuracy of these three methods was high (maximum error detection rate of 7.3%), demonstrating their ability to automatically detect the onset of muscle activity. Recently, other studies have tested different methods (especially Machine Learning based) to determine muscle activation onset offline, reporting promising results.

CONCLUSIONS

This study organized and classified the existing EMG onset detection methods to create consensus towards a possible standardized method for EMG onset detection, which would also allow more reproducibility across studies. The three most commonly used methods (ST, DT and AT) proved to be accurate, while ST and AT were faster in terms of EMG onset detection time, especially when applied on intramuscular EMG data. These are important features towards movement intention identification, especially in real-time applications. Machine Learning methods have received increased attention as an alternative to detect muscle activation onset. However, although several methods have shown their capability offline, more research is required to address their full potential towards real-time applications, namely to infer movement intention.

Wrist EMG Improves Gesture Classification for Stroke Patients

Electromyography (EMG) is a popular human-machine interface for hand gesture control of assistive and rehabilitative technology. EMG can be used to estimate motor intent even when an individual cannot physically move due to weakness or paralysis. EMG is traditionally recorded from the extrinsic hand muscles located in the forearm. However, the wrist has become an increasingly attractive recording location for commercial applications as EMG sensors can be integrated into wrist-worn wearables (e.g., watches, bracelets). Here we explored the impact that recording EMG from the wrist, instead of the forearm, has on stroke patients with upper-limb hemiparesis. We show that EMG signal-to-noise ratio is significantly worse at the paretic wrist relative to the paretic forearm and non-paretic wrist. Despite this, we also show that the ability to classify hand gestures from EMG was significantly better at the paretic wrist relative to the paretic forearm. Our results also provide guidance as to the ideal gestures for each recording location. Namely, single-digit gestures appeared easiest to classify from both forearm and wrist EMG on the paretic side. These results suggest commercialization of wrist-worn EMG would benefit stroke patients by providing more accurate EMG control in a more widely adopted wearable formfactor.

A Light-Weight Artificial Neural Network for Recognition of Activities of Daily Living

Human activity recognition (HAR) is essential for the development of robots to assist humans in daily activities. HAR is required to be accurate, fast and suitable for low-cost wearable devices to ensure portable and safe assistance. Current computational methods can achieve accurate recognition results but tend to be computationally expensive, making them unsuitable for the development of wearable robots in terms of speed and processing power. This paper proposes a light-weight architecture for recognition of activities using five inertial measurement units and four goniometers attached to the lower limb. First, a systematic extraction of time-domain features from wearable sensor data is performed. Second, a small high-speed artificial neural network and line search method for cost function optimization are used for activity recognition. The proposed method is systematically validated using a large dataset composed of wearable sensor data from seven activities (sitting, standing, walking, stair ascent/descent, ramp ascent/descent) associated with eight healthy subjects. The accuracy and speed results are compared against methods commonly used for activity recognition including deep neural networks, convolutional neural networks, long short-term memory and convolutional-long short-term memory hybrid networks. The experiments demonstrate that the light-weight architecture can achieve a high recognition accuracy of 98.60%, 93.10% and 84.77% for seen data from seen subjects, unseen data from seen subjects and unseen data from unseen subjects, respectively, and an inference time of 85 μs. The results show that the proposed approach can perform accurate and fast activity recognition with a reduced computational complexity suitable for the development of portable assistive devices.

Application of a deep generative model produces novel and diverse functional peptides against microbial resistance

Antimicrobial resistance could threaten millions of lives in the immediate future. Antimicrobial peptides (AMPs) are an alternative to conventional antibiotics practice against infectious diseases. Despite the potential contribution of AMPs to the antibiotic's world, their development and optimization have encountered serious challenges. Cutting-edge methods with novel and improved selectivity toward resistant targets must be established to create AMPs-driven treatments. Here, we present AMPTrans-lstm, a deep generative network-based approach for the rational design of AMPs. The AMPTrans-lstm pipeline involves pre-training, transfer learning, and module identification. The AMPTrans-lstm model has two sub-models, namely, (long short-term memory) LSTM sampler and Transformer converter, which can be connected in series to make full use of the stability of LSTM and the novelty of Transformer model. These elements could generate AMPs candidates, which can then be tailored for specific applications. By analyzing the generated sequence and trained AMPs, we prove that AMPTrans-lstm can expand the design space of the trained AMPs and produce reasonable and brand-new AMPs sequences. AMPTrans-lstm can generate functional peptides for antimicrobial resistance with good novelty and diversity, so it is an efficient AMPs design tool.

Classification of elderly pain severity from automated video clip facial action unit analysis: A study from a Thai data repository

Data from 255 Thais with chronic pain were collected at Chiang Mai Medical School Hospital. After the patients self-rated their level of pain, a smartphone camera was used to capture faces for 10 s at a one-meter distance. For those unable to self-rate, a video recording was taken immediately after the move that causes the pain. The trained assistant rated each video clip for the pain assessment in advanced dementia (PAINAD). The pain was classified into three levels: mild, moderate, and severe. OpenFace^© was used to convert the video clips into 18 facial action units (FAUs). Five classification models were used, including logistic regression, multilayer perception, naïve Bayes, decision tree, k-nearest neighbors (KNN), and support vector machine (SVM). Out of the models that only used FAU described in the literature (FAU 4, 6, 7, 9, 10, 25, 26, 27, and 45), multilayer perception is the most accurate, at 50%. The SVM model using FAU 1, 2, 4, 7, 9, 10, 12, 20, 25, and 45, and gender had the best accuracy of 58% among the machine learning selection features. Our open-source experiment for automatically analyzing video clips for FAUs is not robust for classifying pain in the elderly. The consensus method to transform facial recognition algorithm values comparable to the human ratings, and international good practice for reciprocal sharing of data may improve the accuracy and feasibility of the machine learning's facial pain rater.

Bi-directional long short term memory-gated recurrent unit model for Amharic next word prediction

The next word prediction is useful for the users and helps them to write more accurately and quickly. Next word prediction is vital for the Amharic Language since different characters can be written by pressing the same consonants along with different vowels, combinations of vowels, and special keys. As a result, we present a Bi-directional Long Short Term-Gated Recurrent Unit (BLST-GRU) network model for the prediction of the next word for the Amharic Language. We evaluate the proposed network model with 63,300 Amharic sentence and produces 78.6% accuracy. In addition, we have compared the proposed model with state-of-the-art models such as LSTM, GRU, and BLSTM. The experimental result shows, that the proposed network model produces a promising result.

Muscle Co-Contraction Detection in the Time-Frequency Domain

Background: Muscle co-contraction plays a significant role in motion control. Available detection methods typically only provide information in the time domain. The current investigation proposed a novel approach for muscle co-contraction detection in the time–frequency domain, based on continuous wavelet transform (CWT). Methods: In the current study, the CWT-based cross-energy localization of two surface electromyographic (sEMG) signals in the time–frequency domain, i.e., the CWT coscalogram, was adopted for the first time to characterize muscular co-contraction activity. A CWT-based denoising procedure was applied for removing noise from the sEMG signals. Algorithm performances were checked on synthetic and real sEMG signals, stratified for signal-to-noise ratio (SNR), and then validated against an approach based on the acknowledged double-threshold statistical algorithm (DT). Results: The CWT approach provided an accurate prediction of co-contraction timing in simulated and real datasets, minimally affected by SNR variability. The novel contribution consisted of providing the frequency values of each muscle co-contraction detected in the time domain, allowing us to reveal a wide variability in the frequency content between subjects and within stride. Conclusions: The CWT approach represents a relevant improvement over state-of-the-art approaches that provide only a numerical co-contraction index or, at best, dynamic information in the time domain. The robustness of the methodology and the physiological reliability of the experimental results support the suitability of this approach for clinical applications.

Comparison of different algorithms based on TKEO for EMG change point detection

OBJECTIVE

A significant challenge in surface electromyography (EMG) is the accurate identification of onset and offset of muscle activation while maintaining high real-time performance. Teager-Kaiser energy operator (TKEO) is widely used in muscle activity monitoring systems because of its computational simplicity and strong real-time performance. However, in contrast to TKEO ontology, few studies have examined how well the energy operator variants from multiple fields perform in conditioning EMG signals. This paper aims to investigate the role of the energy operator and its variants in EMG change point detection by a threshold detector.

APPROACH

To compare the stability and accuracy of TKEO and its variants for EMG change point detection, the EMG data of extensor carpi radialis longus and flexor carpi radialis were acquired from twenty participants operating a controller under normal and disturbed conditions, and EMG change point detection was performed by four energy operators and their rectified versions.

MAIN RESULTS

Based on the "standard" change points collected by the controller, the detection results were evaluated by three evaluation indexes: detection rate, F1 Score, and accuracy. The experimental results show that the multiresolution energy operator (MTEO) and the TKEO with rectified (abs-TKEO) are more suitable for EMG change point detection.

SIGNIFICANCE

This paper compared the effect of the energy operator and its variants on a threshold-based EMG change point detector. The experimental results in this paper can provide a reference for the selection of EMG signal conditioning methods to improve the detection performance of the EMG change point detector.

Machine Learning for Detection of Muscular Activity from Surface EMG Signals

Background: Muscular-activity timing is useful information that is extractable from surface EMG signals (sEMG). However, a reference method is not available yet. The aim of this study is to investigate the reliability of a novel machine-learning-based approach (DEMANN) in detecting the onset/offset timing of muscle activation from sEMG signals. Methods: A dataset of 2880 simulated sEMG signals, stratified for signal-to-noise ratio (SNR) and time support, was generated to train a hidden single-layer fully-connected neural network. DEMANN’s performance was evaluated on simulated sEMG signals and two different datasets of real sEMG signals. DEMANN was validated against different reference algorithms, including the acknowledged double-threshold statistical algorithm (DT). Results: DEMANN provided a reliable prediction of muscle onset/offset in simulated and real sEMG signals, being minimally affected by SNR variability. When directly compared with state-of-the-art algorithms, DEMANN introduced relevant improvements in prediction performances. Conclusions: These outcomes support DEMANN’s reliability in assessing onset/offset events in different motor tasks and the condition of signal quality (different SNR), improving reference-algorithm performances. Unlike other works, DEMANN’s adopts a machine learning approach where a neural network is trained by only simulated sEMG signals, avoiding the possible complications and costs associated with a typical experimental procedure, making this approach suitable to clinical practice.