Detection of Typical Compensatory Movements during Autonomously Performed Exercises Preventing Low Back Pain (LBP)

With the growing number of people seeking medical advice due to low back pain (LBP), individualised physiotherapeutic rehabilitation is becoming increasingly relevant. Thirty volunteers were asked to perform three typical LBP rehabilitation exercises (Prone-Rocking, Bird-Dog and Rowing) in two categories: clinically prescribed exercise (CPE) and typical compensatory movement (TCM). Three inertial sensors were used to detect the movement of the back during exercise performance and thus generate a dataset that is used to develop an algorithm that detects typical compensatory movements in autonomously performed LBP exercises. The best feature combinations out of 50 derived features displaying the highest capacity to differentiate between CPE and TCM in each exercise were determined. For classifying exercise movements as CPE or TCM, a binary decision tree was trained with the best performing features. The results showed that the trained classifier is able to distinguish CPE from TCM in Bird-Dog, Prone-Rocking and Rowing with up to 97.7% (Head Sensor, one feature), 98.9% (Upper back Sensor, one feature) and 80.5% (Upper back Sensor, two features) using only one sensor. Thus, as a proof-of-concept, the introduced classification models can be used to detect typical compensatory movements in autonomously performed LBP exercises.

HARLEY: A Semi-Automated detection of foci in fluorescence images of yeast

Quantification of cellular structures in fluorescence microscopy data is a key means of understanding cellular function. Unfortunately, numerous cellular structures present unique challenges in their ability to be unbiasedly and accurately detected and quantified. In our studies on stress granules in yeast, users displayed a striking variation of up to 3.7-fold in foci calls and were only able to replicate their results with 62-78% correlation, when requantifying the same images. To facilitate consistent results we developed HARLEY (Human Augmented Recognition of LLPS Ensembles in Yeast), a customizable software for detection and quantification of stress granules in S.cerevisiae. After a brief model training on ~20 cells the detection of foci is fully automated and based on closed loops in intensity contours, constrained only by the a-priori known size of the features of interest. Since no shape is implied, this method is not limited to round features, as is often the case with other algorithms. Candidate features are annotated with a set of geometrical and intensity-based properties to train a kernel Support Vector Machine to recognize features of interest. The trained classifier is then used to create consistent results across datasets. HARLEY is aimed at users without technical expertise, allows for batch processing and is freely available, which should be of broad interest to users focused on analysis of microscopy data in yeast.

Developing an Antiviral Peptides Predictor with Generative Adversarial Network Data Augmentation

Motivation: New antiviral drugs are urgently needed because of emerging viral pathogens' increasing severity and drug resistance. Antiviral peptides (AVPs) have multiple antiviral properties and are appealing candidates for antiviral drug development. We developed a sequence-based binary classifier to identify whether an unknown short peptide has AVP activity. We collected AVP sequence data from six existing databases. We used a generative adversarial network to augment the number of AVPs in the positive training dataset and allow our deep convolutional neural network model to train on more data. Results: Our classifier achieved outstanding performance on the testing dataset compared with other state-of-the-art classifiers. We deployed our trained classifier on a user-friendly web server. Availability and implementation: AI4AVP is freely accessible at http://axp.iis.sinica.edu.tw/AI4AVP/

Combining Statistical Analysis and Machine Learning for EEG Scalp Topograms Classification

Incorporating brain-computer interfaces (BCIs) into daily life requires reducing the reliance of decoding algorithms on the calibration or enabling calibration with the minimal burden on the user. A potential solution could be a pre-trained decoder demonstrating a reasonable accuracy on the naive operators. Addressing this issue, we considered ambiguous stimuli classification tasks and trained an artificial neural network to classify brain responses to the stimuli of low and high ambiguity. We built a pre-trained classifier utilizing time-frequency features corresponding to the fundamental neurophysiological processes shared between subjects. To extract these features, we statistically contrasted electroencephalographic (EEG) spectral power between the classes in the representative group of subjects. As a result, the pre-trained classifier achieved 74% accuracy on the data of newly recruited subjects. Analysis of the literature suggested that a pre-trained classifier could help naive users to start using BCI bypassing training and further increased accuracy during the feedback session. Thus, our results contribute to using BCI during paralysis or limb amputation when there is no explicit user-generated kinematic output to properly train a decoder. In machine learning, our approach may facilitate the development of transfer learning (TL) methods for addressing the cross-subject problem. It allows extracting the interpretable feature subspace from the source data (the representative group of subjects) related to the target data (a naive user), preventing the negative transfer in the cross-subject tasks.

Quantifying Perceived Facial Asymmetry to Enhance Physician–Patient Communications

In cosmetic surgery, bridging the anticipation gap between the patients and the physicians can be challenging if there lacks objective and transparent information exchange during the decision-making and surgical process. Among all factors, facial symmetry is the most important for assessing facial attractiveness. The aim of this work is to promote communications between the two parties by providing a quadruple of quantitative measurements: overall asymmetry index (oAI), asymmetry vector, classification, and confidence vector, using an artificial neural network classifier to model people’s perception acquired from visual questionnaires concerning facial asymmetry. The questionnaire results exhibit a Cronbach’s Alpha value of 0.94 and categorize the respondents’ perception of each stimulus face into perceived normal (PN), perceived asymmetrically normal (PAN), and perceived abnormal (PA) categories. The trained classifier yields an overall root mean squared error < 0.01, and its result shows that the oAI is, in general, proportional to the degree of perceived asymmetry. However, there exist faces that are difficult to classify as either PN or PAN or either PAN or PA with competing confidence values. In such cases, oAI alone is not sufficient to articulate facial asymmetry. Assisting surgeon–patient conversations with the proposed asymmetry quadruple is advised to avoid or to mitigate potential medical disputes.

Graph convolutional network for protein sequence sampling and prediction using structural data

The increasing integration between protein engineering and machine learning has led to many interesting results. A problem still to solve is to evaluate the likelihood that a sequence will fold into a target structure. This problem can be also viewed as sequence prediction from a known structure.In the current work, we propose improvements in the recent architecture of Geometric Vector Perceptrons in order to optimize the sampling of sequences from a known backbone structure. The proposed model differs from the original in that there is: (i) no updating in the vectorial embedding, only in the scalar one, (ii) only one layer of decoding. The first aspect improves the accuracy of the model and reduces the use of memory, the second allows for training of the model with several tasks without incurring data leakage.We treat the trained classifier as an Energy-Based Model and sample sequences by sampling amino acids in a non-autoreggresive manner in the empty positions of the sequence using energy-guided criteria and followed by Monte Carlo optimization.We improve the median identity of samples from 40.2% to 44.7%.An additional question worth investigating is whether sampled and original sequences fold into similar structures independent of their identity. We chose proteins in our test set whose sampled sequences show low identity (under 30%) but for which our model predicted favorable energies. We used trRosetta server and observed that the predicted structures for sampled sequences highly resemble the predicted structures for original sequences, with an average TM score of 0.848.

Explainable Artificial Intelligence for Sarcasm Detection in Dialogues

Sarcasm detection in dialogues has been gaining popularity among natural language processing (NLP) researchers with the increased use of conversational threads on social media. Capturing the knowledge of the domain of discourse, context propagation during the course of dialogue, and situational context and tone of the speaker are some important features to train the machine learning models for detecting sarcasm in real time. As situational comedies vibrantly represent human mannerism and behaviour in everyday real-life situations, this research demonstrates the use of an ensemble supervised learning algorithm to detect sarcasm in the benchmark dialogue dataset, MUStARD. The punch-line utterance and its associated context are taken as features to train the eXtreme Gradient Boosting (XGBoost) method. The primary goal is to predict sarcasm in each utterance of the speaker using the chronological nature of a scene. Further, it is vital to prevent model bias and help decision makers understand how to use the models in the right way. Therefore, as a twin goal of this research, we make the learning model used for conversational sarcasm detection interpretable. This is done using two post hoc interpretability approaches, Local Interpretable Model-agnostic Explanations (LIME) and Shapley Additive exPlanations (SHAP), to generate explanations for the output of a trained classifier. The classification results clearly depict the importance of capturing the intersentence context to detect sarcasm in conversational threads. The interpretability methods show the words (features) that influence the decision of the model the most and help the user understand how the model is making the decision for detecting sarcasm in dialogues.

Analysis of Dynamic Interference Constraints in Cognitive Radio Cloud Networks

To analyze the security performance of a cloud network integrated with a cognitive radio. Cognitive Radio Cloud Network (CRCN) have a resilient potential for dynamic operation for energy saving. Integrating this Cognitive Radio Network (CRN) along with that of a cloud network produces great network exposure, spectrum usage and reduced power consumption. The problem arises for secure transmission and safe storage of huge amount of data in the cloud-based Cognitive Radio (CR) network. This paper, shows a study on preventing the CRCN from some advanced jamming techniques. An adversarial machine learning approach is introduced to launch jamming attacks on wireless communications and a defense strategy is presented. A cognitive transmitter uses a pre-trained classifier to predict the current channel status based on recent sensing results and decides whether to transmit or not, and then uses a Generative Adversarial Network (GAN), for generating synthetic data or false data, thus misleading the jammer to sense false data. The transmitter systematically selects when to take wrong actions and adapts the level of defense to mislead the jammer into making prediction errors and consequently increase its throughput. And to use the conventional cryptographic protocols for protecting the stored database in the cloud network from unauthorized users.

Machine learning-enabled cancer diagnostics with widefield polarimetric second-harmonic generation microscopy

The extracellular matrix (ECM) collagen undergoes major remodeling during tumorigenesis. However, alterations to the ECM are not widely considered in cancer diagnostics, due to mostly uniform appearance of collagen fibers in white light images of hematoxylin and eosin-stained tissue sections. Polarimetric second-harmonic generation (P-SHG) microscopy enables label-free visualization and ultrastructural investigation of non-centrosymmetric molecules, which, when combined with texture analysis, provides multiparameter characterization of tissue collagen. This paper demonstrates whole slide imaging of breast tissue microarrays using high-throughput widefield P-SHG microscopy. The resulting P-SHG parameters are used in classification to differentiate tumor tissue from normal with 94.2% accuracy and F1-score, and 6.3% false discovery rate. Subsequently, the trained classifier is employed to predict tumor tissue with 91.3% accuracy, 90.7% F1-score, and 13.8% false omission rate. As such, we show that widefield P-SHG microscopy reveals collagen ultrastructure over large tissue regions and can be utilized as a sensitive biomarker for cancer diagnostics and prognostics studies.

P009 Disease classifier and microbial dysbiosis index tools cross-predict various pathogenic conditions due to general microbial signal

Background Microbial dysbiosis is widely described in inflammatory bowel disease (IBD), and has been shown to predict IBD state. However, many other diseases including neuro-psychiatric, metabolic, and malignancies, most of which do not result in gut inflammation, are also linked with gut microbial alteration. Since most studies focus on a single disease, the extent of similarity between different diseases is usually not examined. Methods We reanalyzed raw sequencing data from 12,838 human gut V4 16Sseq samples, spanning 59 case-controls comparisons and 28 unique diseases. Novel statistical approach was applied to reduce the effect of the different cohorts; all samples were processed uniformly, and differentially expressed amplicon sequence variants (ASVs) were identified within each cohort. The resulting behavior (direction of change and effect size) of each ASV were then combined across all studies. We used random forest as our classifier and generated non-specific dysbiosis index (NSDI). Results For the disease prediction, each cohort was randomly subsampled to 23 healthy and 23 disease samples. Random forest classifier was trained on one disease/control cohort, and the trained classifier was then used to predict the status of a different disease/control cohort. Disease classifiers performed well in identifying many sick vs. healthy states but failed to differentiate between different diseases. For example, a classifier trained on IBD cohort classified relatively good also disease/control in lupus, schizophrenia, or Parkinson’s from different cohorts. We show this cross-identification is due to a large number of shared disease-associated bacteria and utilize these bacteria to define a novel non-specific dysbiosis index (NSDI). After, we identified 114 non-disease specific ASVs (86 up and 28 down regulated ASVs across diseases in comparison to controls), we calculate the per-sample NSDI by rank-transforming the bacteria within the sample and computing the normalized log ratio of the sum of the ranks of the 86 down and 28 up regulated ASVs. The resulting NSDI is shown to perform better than the previously published CD dysbiosis index (Gevers et al, 2014; PMID: 24629344) indicating that NSDI can successfully differentiate between most cases and controls across a wide variety of diseases. Conclusion A robust non-specific general response of the gut microbiome is detected across different diseases, some of which is shared with IBD. Classifiers trained on a single disease may identify the general non-specific signal and therefore care should be taken when interpreting the classifier predictions. Finally, our NSDI can be used to prioritize the per-sample degree of dysbiosis.