We consider the use of machine learning for convenient, inexpensive monitoring of heart health status using a wrist-worn photoplethysmography (PPG) sensor. Patients' health status is assessed via a standard questionnaire-based measure of symptom severity previously shown to provide information relevant to survival. In a pilot study yielding a nearly classbalanced collection of 198 data records from 15 participants who have been diagnosed with heart failure (heart insufficiency), we attain a macro-averaged classification accuracy of 0.73 using a random forest classifier to predict binarized symptom severity from heart rate variability (HRV) and physical activity metrics extracted from the PPG signals. Our results support the possibility of reducing hospital readmission rates through inhome monitoring, while limiting the frequency of inconvenient and expensive in-person assessments in a clinical setting.

We use a$\beta$-variational autoencoder regularized by a denoising autoencoder to learn latent representations of sleep electroencephalography (EEG) timefrequency spectrograms in a self-supervised manner. Our approach enables the discovery of sleep EEG spectral features unbiased by extrinsic sleep stage labels. We provide sample results that demonstrate the effectiveness of this approach in identifying EEG features that exhibit a variety of patterns of specialization in time and frequency without supervision, including features resembling sleep spindles and theta bursts that occur in sleep stages N2 and N3.

Centered kernel alignment (CKA), also known as centered kernel-target alignment, is useful as a similarity measure between kernels and as a kernel-based similarity measure between feature representations. We prove that CKA based on a Gaussian RBF kernel converges to linear CKA in the large-bandwidth limit. The result relies on mean-centering of the feature maps and on a Hilbert-Schmidt Independence Criterion (HSIC) identity. We show that convergence onset is sensitive to the geometry of the feature representations, and that a notion of representation eccentricity, ρ, constrains the bandwidth range for which Gaussian CKA can differ noticeably from linear CKA. Our experimental results suggest that Gaussian bandwidths less than ρ should be selected in order to enable nonlinear modeling.

We develop a deep convolutional neural network (CNN) that performs sleep stage classification from human sleep EEG and EOG signals. We build and compare several alternative CNN architectures, using nested cross-validation to evaluate performance on unseen patients for different hyperparameter values. Performance after model selection exceeds human expert inter-scorer agreement, even if only a single EEG input channel is used, and is competitive with the state of the art. We further investigate the effect of additional pre-sleep and post-sleep waking data in the training samples. We use t-SNE to visualize the responses of the internal layers of our network, enabling a more complete picture of the development of sleep stage differentiation with layer depth than do activation maximization approaches.

Convolutional neural networks (CNN) have demonstrated state-of-the-art classification results in image categorization, but have received comparatively little attention for classification of one-dimensional physiological signals. We design a deep CNN architecture for automated sleep stage classiffication of human sleep EEG and EOG signals. The CNN proposed in this paper amply outperforms recent work that uses a different CNN architecture over a single-EEG-channel version of the same dataset. We show that the performance gains achieved by our network rely mainly on network depth, and not on the use of several signal channels. Performance of our approach is on par with human expert inter-scorer agreement. By examining the internal activation levels of our CNN, we find that it spontaneously discovers signal features such as sleep spindles and slow waves that figure prominently in sleep stage categorization as performed by human experts.

We introduce a novel ensemble learning framework for supervised classification. Our proposed framework, mixed bagging, is a form of bootstrap aggregating (bagging) in which the sampling process takes into account the classification hardness of the training instances. The classification hardness, or simply hardness, of an instance is defined as the probability that the instance will be misclassified by a classification model built from the remaining instances in the training set. We incorporate instance hardness into the bagging process by varying the sampling probability of each instance based on its estimated hardness. Bootstraps of differing hardness can be created in this way by over-representing, under-representing and equally representing harder instances. This results in a diverse committee of classifiers induced from the bootstraps, whose individual outputs can be aggregated to achieve a final class prediction. We propose two versions of mixed bagging – one where the bootstraps are grouped as easy, regular or hard, with all bootstraps in one group having the same hardness; and the other where the hardness of bootstraps change gradually from one iteration to the next. We have tested our system on 47 publicly available binary classification problems using C4.5 Decision Trees of varying depth as base learners. We find that the proposed mixed bagging methods perform better than traditional bagging and weighted bagging (wagging) regardless of the base learner. The proposed method also outperforms AdaBoost when the base learner consists of deeper decision trees. We examine the results of mixed bagging in terms of bias-variance decomposition and find that mixed bagging is better than AdaBoost at reducing variance and better than traditional bagging at reducing inductive bias.

Human sleep exhibits characteristic patterns during the course of a full night, exemplified by alternation of deep sleep (stage N3/4) and light sleep (stages N1 and N2 occasionally interrupted by REM sleep). However, individual variations in this pattern occur. This paper uncovers a coarse classification of such sleep patterns into types described by varying balances among all-night summary variables such as sleep efficiency and the fraction of sleep period time spent in each of the sleep stages N1, N2, N3/4, and REM. Unsupervised expectation-maximization (EM) clustering is used over data obtained from 244 all-night polysomnographic sleep studies, revealing several naturally occurring sleep type clusters. It is found that sleep efficiency plays a major role in differentiating among sleep types, with time in deep sleep further refining the sleep type classification. Associations are found between sleep type on the one hand, and variables that describe health history and habits on the other. These findings suggest that the discovered sleep types describe medically meaningful groups of sleep behaviors that may be useful in future sleep research.

Our goal in this research is to construct predictive models for clinical performance of pancreatic cancer patients. Current predictive model design in medical oncology literature is dominated by linear and logistic regression techniques. We seek to show that novel machine learning methods can perform as well or better than these traditional techniques.We construct these predictive models via a clinical database we have developed for the University of Massachusetts Memorial Hospitalin Worcester, Massachusetts, USA. The database contains retrospective records of 91 patient treatments for pancreatic tumors.Classification and regression prediction targets include patient survival time, ECOG quality of life scores, surgical outcomes,and tumor characteristics. The predictive accuracy of various data mining models is described, and specific models are presented.

We introduce a specialized association rule mining technique that can extract patterns from complex sleep data comprising polysomnographic recordings, clinical summaries, and sleep questionnaire responses. The rules mined can describe associations among temporally annotated events and questionnaire or summary data; e.g., the likelihood that an occurrence of a rapid eye movement (REM) sleep stage during the second 100 sleep epochs of the night is associated with moderate caffeine intake. We use chi2 analysis to ensure statistical significance of the mined rules at the level P < 0.05. Our results, obtained by mining sleep-related data from 242 human subjects, reveal clinically interesting associations among the polysomnographic and summary variables. Our experience suggests that association mining may also be useful for selection of variables prior to using logistic regression.

We introduce an association rule mining technique for complex datasets described by both static and time-dependent attributes, and apply this technique to find associations among sleep questionnaire responses, clinical summary information, and all-night polysomnographic recordings of sleeping human subjects. Questionnaire data and clinical summaries comprised a total of 63 variables including gender, age, body mass index, Epworth and depression scores. The Rechtschaffen and Kales (R&K) sleep staging information that is standard in sleep research was extracted from the polysomnographic data, and the polysomnographic signals were discretized. The resulting preprocessed polysomnographic data attributes consist of 6 time sequences: sleep stage, airway pressure, blood oxygen potential, heart rate, apneaic episodes and desaturation events, and the patient's body position. An extension of the Apriori association rule mining algorithm designed to deal with time-varying sequences using time windows was developed and employed to uncover statistically significant (P<0.01) and clinically meaningful associations among summary and polysomnographic time series variables.