MEG-Derived Symptom-Sensitive Biomarkers with Long-Term Test-Retest Reliability

Neuroelectric measures derived from human magnetoencephalographic (MEG) recordings hold promise as aides to diagnosis and treatment monitoring and targeting for chronic sequelae of traumatic brain injury (TBI). This study tests novel MEG-derived regional brain measures of tonic neuroelectric activation for long-term test-retest reliability and sensitivity to symptoms. Resting state MEG recordings were obtained from a normative cohort (CamCAN, baseline: n = 613; mean 16-month follow-up: n = 245) and a chronic symptomatic TBI cohort (TEAM-TBI, baseline: n = 62; mean 6-month follow-up: n = 40). The MEG-derived neuroelectric measures were corrected for the empty-room contribution using a random forest classifier. The mean 16-month correlation between baseline and 16-month follow-up CamCAN measures was 0.67; test-retest reliability was markedly improved in this study compared with previous work. The TEAM-TBI cohort was screened for depression, somatization, and anxiety with the Brief Symptom Inventory and for insomnia with the Insomnia Severity Index and was assessed via adjudication for six clinical syndromes: chronic pain, psychological health, and oculomotor, vestibular, cognitive, and sleep dysfunction. Linear classifiers constructed from the 136 regional measures from each TEAM-TBI cohort member distinguished those with and without each symptom, p < 0.0003 for each, i.e., the tonic regional neuroelectric measures of activation are sensitive to the presence/absence of these symptoms and clinical syndromes. The novel regional MEG-derived neuroelectric measures obtained and tested in this study demonstrate the necessary and sufficient properties to be clinically useful, i.e., good test-retest reliability, sensitivity to symptoms in each individual, and obtainable using automatic processing without human judgement or intervention.

The Geometry of Representational Drift in Natural and Artificial Neural Networks

Neurons in sensory areas encode/represent stimuli. Surprisingly, recent studies have suggest that, even during persistent performance, these representations are not stable and change over the course of days and weeks. We examine stimulus representations from fluorescence recordings across hundreds of neurons in the visual cortex using in vivo two-photon calcium imaging and we corroborate previous studies finding that such representations change as experimental trials are repeated across days. This phenomenon has been termed "representational drift". In this study we geometrically characterize the properties of representational drift in the primary visual cortex of mice in two open datasets from the Allen Institute and propose a potential mechanism behind such drift. We observe representational drift both for passively presented stimuli, as well as for stimuli which are behaviorally relevant. Across experiments, the drift most often occurs along directions that have the most variance, leading to a significant turnover in the neurons used for a given representation. Interestingly, despite this significant change due to drift, linear classifiers trained to distinguish neuronal representations show little to no degradation in performance across days. The features we observe in the neural data are similar to properties of artificial neural networks where representations are updated by continual learning in the presence of dropout, i.e. a random masking of nodes/weights, but not other types of noise. Therefore, we conclude that a potential reason for the representational drift in biological networks is driven by an underlying dropout-like noise while continuously learning and that such a mechanism may be computational advantageous for the brain in the same way it is for artificial neural networks, e.g. preventing overfitting.

Markerless Video Analysis of Spontaneous Bodily Movements in 4-Month-Old Infants Predicts Autism-like Behavior in 18-Month-Olds.

Early intervention is now considered the core treatment strategy for autism spectrum disorders (ASD). Thus, it is of significant clinical importance to establish a screening tool for the early detection of ASD in infants. To achieve this goal, in a longitudinal design, we analysed spontaneous bodily movements of 4-month-old infants and assessed their ASD-like behaviours at 18 months of age. Infants at high risk for ASD at 18 months of age exhibited less rhythmic and weaker bodily movement patterns at 4 months of age than low-risk infants. When the observed bodily movement patterns were submitted to a machine learning-based analysis, linear and non-linear classifiers successfully predicted ASD-like behaviour at 18 months of age based on the bodily movement patterns at 4 months of age, at the level acceptable for practical use. This suggests the utility of the proposed method for the early screening of infants at risk for ASD.

Deep Learning Features Encode Interpretable Morphologies within Histological Images

Convolutional neural networks (CNNs) are revolutionizing digital pathology by enabling machine learning-based classification of a variety of phenotypes from hematoxylin and eosin (H&E) whole slide images (WSIs), but the interpretation of CNNs remains difficult. Most studies have considered interpretability in a post hoc fashion, e.g. by presenting example regions with strongly predicted class labels. However, such an approach does not explain the biological features that contribute to correct predictions. To address this problem, here we investigate the interpretability of H&E-derived CNN features (the feature weights in the final layer of a transfer-learning-based architecture), which we show can be construed as abstract morphological genes (“mones”) with strong independent associations to biological phenotypes. We observe that many mones are specific to individual cancer types, while others are found in multiple cancers especially from related tissue types. We also observe that mone-mone correlations are strong and robustly preserved across related cancers. Importantly, linear mone-based classifiers can very accurately separate 38 distinct classes (19 tumor types and their adjacent normals, AUC=97.1%±2.8% for each class prediction), and linear classifiers are also highly effective for universal tumor detection (AUC=99.2%±0.12%). This linearity provides evidence that individual mones or correlated mone clusters may be associated with interpretable histopathological features or other patient characteristics. In particular, the statistical similarity of mones to gene expression values allows integrative mone analysis via expression-based bioinformatics approaches. We observe strong correlations between individual mones and individual gene expression values, notably mones associated with collagen gene expression in ovarian cancer. Mone-expression comparisons also indicate that immunoglobulin expression can be identified using mones in colon adenocarcinoma and that immune activity can be identified across multiple cancer types, and we verify these findings by expert histopathological review. Our work demonstrates that mones provide a morphological H&E decomposition that can be effectively associated with diverse phenotypes, analogous to the interpretability of transcription via gene expression values.

Deep Learning Features Encode Interpretable Morphologies within Histological Images

ABSTRACTConvolutional neural networks (CNNs) are revolutionizing digital pathology by enabling machine learning-based classification of a variety of phenotypes from hematoxylin and eosin (H&E) whole slide images (WSIs), but the interpretation of CNNs remains difficult. Most studies have considered interpretability in a post hoc fashion, e.g. by presenting example regions with strongly predicted class labels. However, such an approach does not explain the biological features that contribute to correct predictions. To address this problem, here we investigate the interpretability of H&E-derived CNN features (the feature weights in the final layer of a transfer-learning-based architecture), which we show can be construed as abstract morphological genes (“mones”) with strong independent associations to biological phenotypes. We observe that many mones are specific to individual cancer types, while others are found in multiple cancers especially from related tissue types. We also observe that mone-mone correlations are strong and robustly preserved across related cancers. Importantly, linear mone-based classifiers can very accurately separate 38 distinct classes (19 tumor types and their adjacent normals, AUC=97.1% ± 2.8% for each class prediction), and linear classifiers are also highly effective for universal tumor detection (AUC=99.2% ± 0.12%). This linearity provides evidence that individual mones or correlated mone clusters may be associated with interpretable histopathological features or other patient characteristics. In particular, the statistical similarity of mones to gene expression values allows integrative mone analysis via expression-based bioinformatics approaches. We observe strong correlations between individual mones and individual gene expression values, notably mones associated with collagen gene expression in ovarian cancer. Mone-expression comparisons also indicate that immunoglobulin expression can be identified using mones in colon adenocarcinoma and that immune activity can be identified across multiple cancer types, and we verify these findings by expert histopathological review. Our work demonstrates that mones provide a morphological H&E decomposition that can be effectively associated with diverse phenotypes, analogous to the interpretability of transcription via gene expression values.

Research on an Adaptive Neural Network K-Pixel Adversarial Example Generation Algorithm

Neural network technology has achieved good results in many tasks, such as image classification. However, for some input examples of neural networks, after the addition of designed and imperceptible perturbations to the examples, these adversarial examples can change the output results of the original examples. For image classification problems, we derive low-dimensional attack perturbation solutions on multidimensional linear classifiers and extend them to multidimensional nonlinear neural networks. Based on this, a new adversarial example generation algorithm is designed to modify a specified number of pixels. The algorithm adopts a greedy iterative strategy, and gradually iteratively determines the importance and attack range of pixel points. Finally, experiments demonstrate that the algorithm-generated adversarial example is of good quality, and the effects of key parameters in the algorithm are also analyzed.

About Methods for Classifying Hidden Language Concepts in Specialized Texts Involving Pseudoinverse, Clustering and Data Grouping

This paper discusses the problems of analysis of hidden language concepts in scientific texts in the Ukrainian language, using methods of text mining, dimensionality reduction, grouping of features and linear classifiers. A corpus of scientific texts and dictionaries, as well as stop words and affixes, has been formed for processing specialized texts. The resulting texts were analyzed and converted into text frequency-inverse document frequency (TF-IDF) feature representation. In order to process the feature vector, we propose to use methods of dimensionality rteduction of the data, in particular, the algorithm for the synthesis of linear systems and Karunen – Loeve transform and grouping of features: T-stochastic grouping of nearest neighbors (T-SNE). A series of experiments were performed on test examples, in particular, for the determination of informational density in the text and classification by keywords in specialized texts using the method of random samples consensus (RANSAC). A method of classification of hidden language concepts was proposed, making use of clustering methods (K-means). As a result of the experiment, the structure of the classifier of hidden language concepts was obtained in structured texts was obtained, which gained a relatively high recognition accuracy (97 – 99 %) using such linear classification algorithms: decision trees and extreme gradient boost machine. The stability of the proposed method is investigated by using the perturbation of the original data by a variational autoencoder, test runs shown that sparse autocoder reduces the mean square error, but the separation band decreases, which affects the convergence of the classification algorithm. In further research, we propose to apply other methods of analysis of structured texts and ways to improve the separability of specialized texts with similar authorial styles and different topic using a proposed set of parameters. Keywords: text processing, language concepts, pseudoinverse, clusterization, methods of data groupings.

The effect of mobile camera selection on the capacity to predict water turbidity

Recently, cameras of mobile phones emerged as an alternative for quantifying water turbidity. Most of these studies lack a strategy to determine the water turbidity for new samples, focusing mainly on one particular device. Nevertheless, widespread use of these approaches requires a predictive capacity on out-of-the-sample images acquired in devices of different capabilities. We studied the influence of mobile device camera sensors on the predictive performance of water turbidity for non-previously observed turbid images. For this, a reference database with turbid images acquired for different mobile devices was constructed. A machine learning method based on image quality measures and linear classifiers (least squares and LASSO) was proposed to perform predictions on each mobile device. Relative accuracy and precision were evaluated. Results suggest that these approaches may provide accurate predictions reaching most than 80% of relative accuracy with high test-retest reliability (&gt;.99). Nevertheless, our results also indicate that the predictive performance levels dropped in low capacity quality sensors. Therefore, despite the high performance that can be reached using these approaches, widespread use on multiple mobile devices may require further developments of low-quality sensors and a better understanding of their operative ranges.