Detection of Premature Ventricular Complexes using Semisupervised Autoencoders and Random Forests.

In this paper, we propose a technique for detection of premature ventricular complexes (PVC) based on information obtained from single-lead electrocardiogram (ECG) signals. A combination of semisupervised autoencoders and Random Forests models are used for feature extraction and PVC detection. The ECG signal is first denoised using Stationary Wavelet Transforms and denoising convolutional autoencoders. Following this, PVC classification is performed. Individual ECG beat segments along with features derived from three consecutive beats are used to train a hybrid autoencoder network to learn class-specific beat encodings. These encodings, along with the beat-triplet features, are then input to a Random Forests classifier for final PVC classification. Results: The performance of our algorithm was evaluated on ECG records in the MIT-BIH Arrhythmia Database (MITDB) and the St. Petersburg INCART Database (INCARTDB). Our algorithm achieves a sensitivity of 92.67% and a PPV of 95.58% on the MITDB database. Similarly, a sensitivity of 88.08% and a PPV of 94.76% are achieved on the INCARTDB database.

Deep Understanding of Breast Density Classification.

We have developed a deep learning architecture, DualViewNet, for mammogram density classification as well as a novel metric for quantifying network preference of mediolateral oblique (MLO) versus craniocaudal (CC) views in density classification. Also, we have provided thorough analysis and visualization to better understand the behavior of deep neural networks in density classification. Our proposed architecture, DualViewNet, simultaneously examines and classifies both MLO and CC views corresponding to the same breast, and shows best performance with a macro average AUC of 0.8970 and macro average 95% confidence interval of 0.8239-0.9450 obtained via bootstrapping 1000 test sets. By leveraging DualViewNet we provide a novel algorithm and quantitative comparison of MLO versus CC views for classification and find that MLO provides stronger influence in 1,187 out of 1,323 breasts.

Detection of atrial fibrillation using discrete-state Markov models and Random Forests.

In this paper, we present a fully automated technique for robust detection of Atrial Fibrillation (AF) episodes in single-lead electrocardiogram (ECG) signals using discrete-state Markov models and Random Forests. METHODS: The ECG signal is first preprocessed using Stationary Wavelet Transforms (SWT) for noise suppression, signal quality assessment and subsequent R-peak detection. Discrete-state Markov probabilities modelling transitions between successive RR intervals along with other statistical quantities derived from the RR-interval series constitute the feature set to perform AF classification using Random Forests. Further enhancement in AF detection is achieved by using a post-processing false positive suppression algorithm based on autocorrelation analysis of the RR-interval series. Datasets: The AF classifier was trained using the Physionet/Computing in Cardiology 2017 AF Challenge dataset and the Atrial Fibrillation Termination Database (AFTDB). The test datasets consist of the MIT-BIH Atrial Fibrillation Database (AFDB) and the MIT-BIH Arrhythmia Database (MITDB). RESULTS: Our algorithms achieved sensitivity, specificity and F-score values of 97.4%, 98.6% and 97.7% respectively on the AFDB dataset and 96.3%, 97.0% and 85.6% respectively on the MITDB dataset. It was also observed that inclusion of the false positive suppression step resulted in a 1.1% increase in specificity and a 4.0% increase in F-score for the MITDB dataset without any decrease in sensitivity. CONCLUSION: The proposed method of AF detection, combining Markov models and Random Forests, achieves high accuracy across multiple databases and demonstrates comparable or superior performance to several other state-of-the-art algorithms.

MAPGI: Accurate identification of anatomical landmarks and diseased tissue in gastrointestinal tract using deep learning.

Automatic detection of anatomical landmarks and diseases in medical images is a challenging task which could greatly aid medical diagnosis and reduce the cost and time of investigational procedures. Also, two particular challenges of digital image processing in medical applications are the sparsity of annotated medical images and the lack of uniformity across images and image classes. This paper presents methodologies for maximizing classification accuracy on a small medical image dataset, the Kvasir dataset, by performing robust image preprocessing and applying state-of-the-art deep learning. Images are classified as being or involving an anatomical landmark (pylorus, z-line, cecum), a diseased state (esophagitis, ulcerative colitis, polyps), or a medical procedure (dyed lifted polyps, dyed resection margins). A framework for modular and automatic preprocessing of gastrointestinal tract images (MAPGI) is proposed, which applies edge removal, contrast enhancement, filtering, color mapping and scaling to each image in the dataset. Gamma correction values are automatically calculated for individual images such that the mean pixel value for each image is normalized to 90 ±â€¯1 in a 0-255 pixel value range. Three state-of-the-art neural networks architectures, Inception-ResNet-v2, Inception-v4, and NASNet, are trained on the Kvasir dataset, and their classification performance is juxtaposed on validation data. In each case, 85% of the images from the Kvasir dataset are used for training, while the other 15% are reserved for validation. The resulting accuracies achieved using Inception-v4, Inception-ResNet-v2, and NASNet were 0.9845, 0.9848, and 0.9735, respectively. In addition, Inception-v4 achieved an average of 0.938 precision, 0.939 recall, 0.991 specificity, 0.938 F1 score, and 0.929 Matthews correlation coefficient (MCC). Bootstrapping provided NASNet, the worst performing model, a lower bound of 0.9723 accuracy on the 95% confidence interval.

Toward a Clinical Point of Care Tool for Managing Heart Failure.

Management of heart failure is a major challenge in health care. Optimal management of heart failure requires adherence to evidence-based clinical guidelines. The nearly 80-page guideline for heart failure management is very complex. As a result, clinical guidelines are difficult to implement and are adopted slowly by the medical community at large. In this paper we describe a heart failure treatment adviser system which automates the reasoning process required to comply with the heart failure management guideline. The system is able to correctly compute guideline- compliant treatment recommendations for a given patient. For each recommendation, justification is also given by the system. We illustrate the technical aspect of the implementation of the system as well as some primitive user interfaces associated with the system's core functionality. A simulated case is presented with system-generated recommendations and justifications.

RAMS: Remote and automatic mammogram screening.

About one in eight women in the U.S. will develop invasive breast cancer at some point in life. Breast cancer is the most common cancer found in women and if it is identified at an early stage by the use of mammograms, x-ray images of the breast, then the chances of successful treatment can be high. Typically, mammograms are screened by radiologists who determine whether a biopsy is necessary to ascertain the presence of cancer. Although historical screening methods have been effective, recent advances in computer vision and web technologies may be able to improve the accuracy, speed, cost, and accessibility of mammogram screenings. We propose a total screening solution comprised of three main components: a web service for uploading images and reviewing results, a machine learning algorithm for accepting or rejecting images as valid mammograms, and an artificial neural network for locating potential malignancies. Once an image is uploaded to our web service, an image acceptor determines whether or not the image is a mammogram. The image acceptor is primarily a one-class SVM built on features derived with a variational autoencoder. If an image is accepted as a mammogram, the malignancy identifier, a ResNet-101 Faster R-CNN, will locate tumors within the mammogram. On test data, the image acceptor had only 2 misclassifications out of 410 mammograms and 2 misclassifications out of 1,640 non-mammograms while the malignancy identifier achieved 0.951 AUROC when tested on BI-RADS 1, 5, and 6 images from the INbreast dataset.

An AI-Based Heart Failure Treatment Adviser System.

Management of heart failure is a major health care challenge. Healthcare providers are expected to use best practices described in clinical practice guidelines, which typically consist of a long series of complex rules. For heart failure management, the relevant guidelines are nearly 80 pages long. Due to their complexity, the guidelines are often difficult to fully comply with, which can result in suboptimal medical practices. In this paper, we describe a heart failure treatment adviser system that automates the entire set of rules in the guidelines for heart failure management. The system is based on answer set programming, a form of declarative programming suited for simulating human-style reasoning. Given a patient's information, the system is able to generate a set of guideline-compliant recommendations. We conducted a pilot study of the system on 21 real and 10 simulated patients with heart failure. The results show that the system can give treatment recommendations compliant with the guidelines. Out of 187 total recommendations made by the system, 176 were agreed upon by the expert cardiologists. Also, the system missed eight valid recommendations. The reason for the missed and discordant recommendations seems to be insufficient information, differing style, experience, and knowledge of experts in decision-making that were not captured in the system at this time. The system can serve as a point-of-care tool for clinics. Also, it can be used as an educational tool for training physicians and an assessment tool to measure the quality metrics of heart failure care of an institution.

Early Detection of Peak Demand Days of Chronic Respiratory Diseases Emergency Department Visits Using Artificial Neural Networks.

Chronic respiratory diseases, mainly asthma and chronic obstructive pulmonary disease (COPD), affect the lives of people by limiting their activities in various aspects. Overcrowding of hospital emergency departments (EDs) due to respiratory diseases in certain weather and environmental pollution conditions results in the degradation of quality of medical care, and even limits its availability. A useful tool for ED managers would be to forecast peak demand days so that they can take steps to improve the availability of medical care. In this paper, we developed an artificial neural network based classifier using multilayer perceptron with back propagation algorithm that predicts peak event (peak demand days) of patients with respiratory diseases, mainly asthma and COPD visiting EDs in Dallas County of Texas in the United States. The precision and recall for peak event class were 77.1% and 78.0%, respectively, and those for nonpeak events were 83.9% and 83.2%, respectively. The overall accuracy of the system is 81.0%.

Augmented reality for breast tumors visualization.

3D visualization of breast tumors are shown to be effective by previous studies. In this paper, we introduce a new augmented reality application that can help doctors and surgeons to have a more accurate visualization of breast tumors; this system uses a marker-based image-processing technique to render a 3D model of the tumors on the body. The model can be created using a combination of breast 3D mammography by experts. We have tested the system using an Android smartphone and a head-mounted device. This proof of concept can be useful for oncologists to have a more effective screening, and surgeons to plan the surgery.

A resource-efficient planning for pressure ulcer prevention.

Pressure ulcer is a critical problem for bed-ridden and wheelchair-bound patients, diabetics, and the elderly. Patients need to be regularly repositioned to prevent excessive pressure on a single area of body, which can lead to ulcers. Pressure ulcers are extremely costly to treat and may lead to several other health problems, including death. The current standard for prevention is to reposition at-risk patients every two hours. Even if it is done properly, a fixed schedule is not sufficient to prevent all ulcers. Moreover, it may result in nurses being overworked by turning some patients too frequently. In this paper, we present an algorithm for finding a nurse-effort optimal repositioning schedule that prevents pressure ulcer formation for a finite planning horizon. Our proposed algorithm uses data from a commercial pressure mat assembled on the beds surface and provides a sequence of next positions and the time of repositioning for each patient.

Apnea MedAssist: real-time sleep apnea monitor using single-lead ECG.

We have developed a low-cost, real-time sleep apnea monitoring system ''Apnea MedAssist" for recognizing obstructive sleep apnea episodes with a high degree of accuracy for both home and clinical care applications. The fully automated system uses patient's single channel nocturnal ECG to extract feature sets, and uses the support vector classifier (SVC) to detect apnea episodes. "Apnea MedAssist" is implemented on Android operating system (OS) based smartphones, uses either the general adult subject-independent SVC model or subject-dependent SVC model, and achieves a classification F-measure of 90% and a sensitivity of 96% for the subject-independent SVC. The real-time capability comes from the use of 1-min segments of ECG epochs for feature extraction and classification. The reduced complexity of "Apnea MedAssist" comes from efficient optimization of the ECG processing, and use of techniques to reduce SVC model complexity by reducing the dimension of feature set from ECG and ECG-derived respiration signals and by reducing the number of support vectors.

Real-time sleep quality assessment using single-lead ECG and multi-stage SVM classifier.

Sleep efficiency measures provide an objective assessment to gauge the quality of individual's sleep. In this study we present a home-based, automated and non-intrusive system that is based on Electrocardiogram (ECG) measurements and uses a multi-stage Support Vector Machines (SVM) classifier to measure three indices for sleep quality assessment per 30 s epoch segment: Sleep Efficiency Index, Delta-Sleep Efficiency Index and Sleep Onset Latency. This method provides an alternative to the intrusive and expensive Polysomnography (PSG) and scoring by Rechtschaffen and Kales visual method.

A patient-adaptive profiling scheme for ECG beat classification.

Recent trends in clinical and telemedicine applications highly demand automation in electrocardiogram (ECG) signal processing and heart beat classification. A patient-adaptive cardiac profiling scheme using repetition-detection concept is proposed in this paper. We first employ an efficient wavelet-based beat-detection mechanism to extract precise fiducial ECG points. Then, we implement a novel local ECG beat classifier to profile each patient's normal cardiac behavior. ECG morphologies vary from person to person and even for each person, it can vary over time depending on the person's physical condition and/or environment. Having such profile is essential for various diagnosis (e.g., arrhythmia) purposes. One application of such profiling scheme is to automatically raise an early warning flag for the abnormal cardiac behavior of any individual. Our extensive experimental results on the MIT-BIH arrhythmia database show that our technique can detect the beats with 99.59% accuracy and can identify abnormalities with a high classification accuracy of 97.42%.