Information maximization-based clustering of histopathology images using deep learning.

Pancreatic cancer is one of the most adverse diseases and it is very difficult to treat because the cancer cells formed in the pancreas intertwine themselves with nearby blood vessels and connective tissue. Hence, the surgical procedure of treatment becomes complicated and it does not always lead to a cure. Histopathological diagnosis is the usual approach for cancer diagnosis. However, the pancreas remains so deep inside the body that experts sometimes struggle to detect cancer in it. Computer-aided diagnosis can come to the aid of pathologists in this scenario. It assists experts by supporting their diagnostic decisions. In this research, we carried out a deep learning-based approach to analyze histopathology images. We collected whole-slide images of KPC mice to implement this work. The pancreatic abnormalities observed in KPC mice develop similar histological features to human beings. We created random patches from whole-slide images. Then, a convolutional autoencoder framework was used to embed these patches into an integrated latent space. We applied 'information maximization', a deep learning clustering technique to cluster the identical patches in an unsupervised manner since our dataset does not have annotation. Moreover, Uniform manifold approximation and projection, a nonlinear dimension reduction technique was utilized to visualize the embedded patches in a 2-dimensional space. Finally, we calculated a few internal cluster validation metrics to determine the optimal cluster set. Our work concentrated on patch-based anomaly detection in the whole slide histopathology images of KPC mice.

On QSAR-based cardiotoxicity modeling with the expressiveness-enhanced graph learning model and dual-threshold scheme.

Introduction: Given the direct association with malignant ventricular arrhythmias, cardiotoxicity is a major concern in drug design. In the past decades, computational models based on the quantitative structure-activity relationship have been proposed to screen out cardiotoxic compounds and have shown promising results. The combination of molecular fingerprint and the machine learning model shows stable performance for a wide spectrum of problems; however, not long after the advent of the graph neural network (GNN) deep learning model and its variant (e.g., graph transformer), it has become the principal way of quantitative structure-activity relationship-based modeling for its high flexibility in feature extraction and decision rule generation. Despite all these progresses, the expressiveness (the ability of a program to identify non-isomorphic graph structures) of the GNN model is bounded by the WL isomorphism test, and a suitable thresholding scheme that relates directly to the sensitivity and credibility of a model is still an open question. Methods: In this research, we further improved the expressiveness of the GNN model by introducing the substructure-aware bias by the graph subgraph transformer network model. Moreover, to propose the most appropriate thresholding scheme, a comprehensive comparison of the thresholding schemes was conducted. Results: Based on these improvements, the best model attains performance with 90.4% precision, 90.4% recall, and 90.5% F1-score with a dual-threshold scheme (active: <1µM; non-active: >30µM). The improved pipeline (graph subgraph transformer network model and thresholding scheme) also shows its advantages in terms of the activity cliff problem and model interpretability.

On finding natural antibiotics based on TCM formulae.

CONTEXT: Novel kinds of antibiotics are needed to combat the emergence of antibacterial resistance. Natural products (NPs) have shown potential as antibiotic candidates. Current experimental methods are not yet capable of exploring the massive, redundant, and noise-involved chemical space of NPs. In silico approaches are needed to select NPs as antibiotic candidates. OBJECTIVE: This study screens out NPs with antibacterial efficacy guided by both TCM and modern medicine and constructed a dataset aiming to serve the new antibiotic design. METHOD: A knowledge-based network is proposed in this study involving NPs, herbs, the concepts of TCM, and the treatment protocols (or etiologies) of infectious in modern medicine. Using this network, the NPs candidates are screened out and compose the dataset. Feature selection of machine learning approaches is conducted to evaluate the constructed dataset and statistically validate the im- portance of all NPs candidates for different antibiotics by a classification task. RESULTS: The extensive experiments prove the constructed dataset reaches a convincing classification performance with a 0.9421 weighted accuracy, 0.9324 recall, and 0.9409 precision. The further visu- alizations of sample importance prove the comprehensive evaluation for model interpretation based on medical value considerations.

Sleep Staging Framework with Physiologically Harmonized Sub-Networks.

Sleep screening is an important tool for both healthcare and neuroscientific research. Automatic sleep scoring is an alternative to the time-consuming gold-standard manual scoring procedure. Recently there have seen promising results on automatic stage scoring by extracting spatio-temporal features via deep neural networks from electroencephalogram (EEG). However, such methods fail to consistently yield good performance due to a missing piece in data representation: the medical criterion of the sleep scoring task on top of EEG features. We argue that capturing stage-specific features that satisfy the criterion of sleep medicine is non-trivial for automatic sleep scoring. This paper considers two criteria: Transient stage marker and Overall profile of EEG features, then we propose a physiologically meaningful framework for sleep stage scoring via mixed deep neural networks. The framework consists of two sub-networks: feature extraction networks, constructed in consideration of the physiological characteristics of sleep, and an attention-based scoring decision network. Moreover, we quantize the framework for potential use under an IoT setting. For proof-of-concept, the performance of the proposed framework is demonstrated by introducing multiple sleep datasets with the largest comprising 42,560 h recorded from 5,793 subjects. From the experiment results, the proposed method achieves a competitive stage scoring performance, especially for Wake, N2, and N3, with higher F1 scores of 0.92, 0.86, and 0.88, respectively. Moreover, the feasibility analysis of framework quantization provides a potential for future implementation in the edge computing field and clinical settings.

Cancer Subtyping via Embedded Unsupervised Learning on Transcriptomics Data.

Cancer is one of the deadliest diseases worldwide. Accurate diagnosis and classification of cancer subtypes are indispensable for effective clinical treatment. Promising results on automatic cancer subtyping systems have been published recently with the emergence of various deep learning methods. However, such automatic systems often overfit the data due to the high dimensionality and scarcity. In this paper, we propose to investigate automatic subtyping from an unsupervised learning perspective by directly constructing the underlying data distribution itself, hence sufficient data can be generated to alleviate the issue of overfitting. Specifically, we bypass the strong Gaussianity assumption that typically exists but fails in the unsupervised learning subtyping literature due to small-sized samples by vector quantization. Our proposed method better captures the latent space features and models the cancer subtype manifestation on a molecular basis, as demonstrated by the extensive experimental results.

Feasibility Analysis of Symbolic Representation for Single-Channel EEG-Based Sleep Stages.

Sleep screening based on the construction of sleep stages is one of the major tool for the assessment of sleep quality and early detection of sleep-related disorders. Due to the inherent variability such as inter-users anatomical variability and the inter-systems differences, representation learning of sleep stages in order to obtain the stable and reliable characteristics is runoff for downstream tasks in sleep science. In this paper, we investigated feasibility of the EEG-based symbolic representation for sleep stages. By combining the Latent Dirichlet Allocation topic model and comparing with different feature extraction methods, the work proved the feasibility of multi-topics representation for sleep stages and physiological signals.

A Real-Time Portable IoT System for Telework Tracking.

Telework has become a universal working style under the background of COVID-19. With the increased time of working at home, problems, such as lack of physical activities and prolonged sedentary behavior become more prominent. In this situation, a self-managing working pattern regulation may be the most practical way to maintain worker's well-being. To this end, this paper validated the idea of using an Internet of Things (IoT) system (a smartphone and the accompanying smartwatch) to monitor the working status in real-time so as to record the working pattern and nudge the user to have a behavior change. By using the accelerometer and gyroscope enclosed in the smartwatch worn on the right wrist, nine-channel data streams of the two sensors were sent to the paired smartphone for data preprocessing, and action recognition in real time. By considering the cooperativity and orthogonality of the data streams, a shallow convolutional neural network (CNN) model was constructed to recognize the working status from a common working routine. As preliminary research, the results of the CNN model show accurate performance [5-fold cross-validation: 0.97 recall and 0.98 precision; leave-one-out validation: 0.95 recall and 0.94 precision; (support vector machine (SVM): 0.89 recall and 0.90 precision; random forest: 0.95 recall and 0.93 precision)] for the recognition of working status, suggesting the feasibility of this fully online method. Although further validation in a more realistic working scenario should be conducted for this method, this proof-of-concept study clarifies the prospect of a user-friendly online working tracking system. With a tailored working pattern guidance, this method is expected to contribute to the workers' wellness not only during the COVID-19 pandemic but also take effect in the post-COVID-19 era.

The feasibility of predicting impending malignant ventricular arrhythmias by using nonlinear features of short heartbeat intervals.

BACKGROUND AND OBJECTIVE: Malignant ventricular arrhythmias (MAs) occur unpredictably and lead to emergencies. A new approach that uses a timely tracking device e.g., photoplethysmogram (PPG) solely to predict MAs would be irreplaceably valuable and it is natural to expect the approach can predict the occurrence as early as possible. METHOD: We assumed that with an appropriate metric based on signal complexity, the heartbeat interval time series (HbIs) can be used to manifest the intrinsic characteristics of the period immediately precedes the MAs (preMAs). The approach first characterizes the patterns of preMAs by a new complexity metric (the refined composite multi-scale entropy). The MAs detector is then constructed by checking the discriminability of the MAs against the sinus rhythm and other prevalent arrhythmias (atrial fibrillation and premature ventricular contraction) of three machine-learning models (SVM, Random Forest, and XGboost). RESULTS: Two specifications are of interest: the length of the HbIs needed to delineate the preMAs patterns sufficiently (lspec) and how long before the occurrence of MAs will the HbIs manifest specific patterns that are distinct enough to predict the impending MAs (tspec). Our experimental results confirmed the best performance came from a Random-Forest model with an average precision of 99.99% and recall of 88.98% using a HbIs of 800 heartbeats (the lspec), 108 seconds (the tspec) before the occurrence of MAs. CONCLUSION: By experimental validation of the unique pattern of the preMAs in HbIs and using it in the machine learning model, we showed the high possibility of MAs prediction in a broader circumstance, which may cover daily healthcare using the alternative sensor in HbIs monitoring. Therefore, this research is theoretically and practically significant in cardiac arrest prevention.

A Novel CNN-Based Framework for Classification of Signal Quality and Sleep Position from a Capacitive ECG Measurement.

The further exploration of the capacitive ECG (cECG) is hindered by frequent fluctuations in signal quality from body movement and changes in sleep position. The processing framework must be fundamentally adapted to make full use of this signal. Therefore, we propose a new signal-processing framework that determines the signal quality for short signal segments (2 and 4 seconds) using a multi-class classification model (qua_model) based on a convolutional neural network (CNN). We built another independent deep CNN classifier (pos_model) to classify the sleep position. In the validation, 12 subjects were recruited for a 30-minute experiment, which required the subjects to lie on a bed in different sleeping positions. The short segments, classified as clear (C1 class) by the qua_model, were used to determine sleep positions with the pos_model. In 10-fold cross-validation, the qua_model for signals of 4-second length could recognize the signal of the C1 class at a 0.99 precision and a 0.99 recall; the pos_model could recognize the supine sleep position, the left, and right lateral sleep positions at a 0.99 averaged precision and a 0.99 averaged recall. Given the amount of data accumulated per night and the instability in the signal quality, this fully automatic processing framework is indispensable for a personal healthcare system. Therefore, this study could serve as an important step for cECG technique trying to explore the cECG for unconstrained heart monitoring.

The Feasibility of Arrhythmias Detection from A Capacitive ECG Measurement Using Convolutional Neural Network.

Capacitive ECG (cECG) can measure the cardiac electrical signal via capacitive coupling between electrodes and skin. This unconstrained measurement is suitable for personal heart monitoring; however, the instability in the quality of the signal hinders a further use of the signal. To use the cECG for heart monitoring, an adapted framework that could automatically classify the signal into clear cECG, arrhythmias and noise signal is a prerequisite. In view of this problem, the conventional quality estimation method using predefined features based on R-peak detection is not suitable for this unconstrained measurement of cECG. In this study, we examine the feasibility of arrhythmias detection from the cECG measurement using a convolutional neural network (CNN) model. The malignant ventricular tachycardia (VT) and ventricular fibrillation (VF) do not have the Q-R-S waveforms and therefore may be easily classified as the noise. Hence, in this study, we used the cECG signals that have 3 classes in quality (C1: clear signal; C2: blurry signal with significant R peak and N: noise) and the arrhythmias signals (VT, VF, and atrial fibrillation) from open databases to train the classification model. 13 subjects were recruited in an experiment for the cECG data collection in the Nara Institute of Science and Technology. As a result, the CNN model could recognize C1 and AF signal with over 0.98 recalls and precisions; whereas the recall and precision of VT and VF are lower scores and the lower scores were caused mainly by the similarity between VT and VF. Given the results of the CNN model, this CNN-based framework can accurately label the C1, AF, and malignant ventricular arrhythmias (VT and VF) signals. Further stratification of the C2, VT, and VF will further enhance the use of the cECG measurement.