Search for: All records

Sleep staging has a very important role in diagnosing patients with sleep disorders. In general, this task is very time-consuming for physicians to perform. Deep learning shows great potential to automate this process and remove physician bias from decision making. In this study, we aim to identify recent trends on performance improvement and the causes for these trends. Recent papers on sleep stage classification and interpretability are investigated to explore different modeling and data manipulation techniques, their efficiency, and recent advances. We identify an improvement in performance up to 12% on standard datasets over the last 5 years. The improvements in performance do not appear to be necessarily correlated to the size of the models, but instead seem to be caused by incorporating new architectural components, such as the use of transformers and contrastive learning.

 

Physiological and kinematic signals from humans are often used for monitoring health. Several processes of interest (e.g., cardiac and respiratory processes, and locomotion) demonstrate periodicity. Training models for inference on these signals (e.g., detection of anomalies, and extraction of biomarkers) require large amounts of data to capture their variability, which are not readily available. This hinders the performance of complex inference models. In this work, we introduce a methodology for improving inference on such signals by incorporating phase-based interpretability and other inference tasks into a multi-task framework applied to a generative model. For this purpose, we utilize phase information as a regularization term and as an input to the model and introduce an interpretable unit in a neural network, which imposes an interpretable structure on the model. This imposition helps us in the smooth generation of periodic signals that can aid in data augmentation tasks. We demonstrate the impact of our framework on improving the overall inference performance on ECG signals and inertial signals from gait locomotion. 

Body-rocking is an undesired stereotypical motor movement performed by some individuals, and its detection is essential for self-awareness and habit change. We envision a pipeline that includes inertial wearable sensors and a real-time detection system for notifying the user so that they are aware of their body-rocking behavior. For this task, similarities of body rocking to other non-related repetitive activities may cause false detections which prevent continuous engagement, leading to alarm fatigue. We present a pipeline using Bayesian Neural Networks with uncertainty quantification for jointly reducing false positives and providing accurate detection. We show that increasing model capacity does not consistently yield higher performance by itself, while pairing it with the Bayesian approach does yield significant improvements. Disparities in uncertainty quantification are better quantified by calibrating them using deep neural networks. We show that the calibrated probabilities are effective quality indicators of reliable predictions. Altogether, we show that our approach provides additional insights on the role of Bayesian techniques in deep learning as well as aids in accurate body-rocking detection, improving our prior work on this subject. 

Cough detection can provide an important marker to monitor chronic respiratory conditions. However, manual techniques which require human expertise to count coughs are both expensive and time-consuming. Recent Automatic Cough Detection Algorithms (ACDAs) have shown promise to meet clinical monitoring requirements, but only in recent years they have made their way to non-clinical settings due to the required portability of sensing technologies and the extended duration of data recording. More precisely, these ACDAs operate at high sampling frequencies, which leads to high power consumption and computing requirements, making these difficult to implement on a wearable device. Additionally, reproducibility of their performance is essential. Unfortunately, as the majority of ACDAs were developed using private clinical data, it is difficult to reproduce their results. We, hereby, present an ACDA that meets clinical monitoring requirements and reliably operates at a low sampling frequency. This ACDA is implemented using a convolutional neural network (CNN), and publicly available data. It achieves a sensitivity of 92.7%, a specificity of 92.3%, and an accuracy of 92.5% using a sampling frequency of just 750 Hz. We also show that a low sampling frequency allows us to preserve patients’ privacy by obfuscating their speech, and we analyze the trade-off between speech obfuscation for privacy and cough detection accuracy. Clinical relevance—This paper presents a new cough detection technique and preliminary analysis on the trade-off between detection accuracy and obfuscation of speech for privacy. These findings indicate that, using a publicly available dataset, we can sample signals at 750 Hz while still maintaining a sensitivity above 90%, suggested to be sufficient for clinical monitoring [1]. 

Longitudinal fetal health monitoring is essential for high-risk pregnancies. Heart rate and heart rate variability are prime indicators of fetal health. In this work, we implemented two neural network architectures for heartbeat detection on a set of fetal phonocardiogram signals captured using fetal Doppler and a digital stethoscope. We test the efficacy of these networks using the raw signals and the hand-crafted energy from the signal. The results show a Convolutional Neural Network is the most efficient at identifying the S1 waveforms in a heartbeat, and its performance is improved when using the energy of the Doppler signals. We further discuss issues, such as low Signal-to-Noise Ratios (SNR), present in the training of a model based on the stethoscope signals. Finally, we show that we can improve the SNR, and subsequently the performance of the stethoscope, by matching the energy from the stethoscope to that of the Doppler signal. 

Fossil single-celled marine organisms known as foraminifera are widely used in oceanographic research. The identification of species is one of the most common tasks when analyzing ocean samples. One of the primary criteria for species identification is their morphology. Automatic segmentation of images of foraminifera would aid on the identification task as well as on other morphological studies. We pose this problem as an edge detection task for which capturing the correct topological structure is essential. Due to the presence of soft edges and even unclosed segments, state-of-the-art techniques have problems capturing the correct edge structure. Standard pixel-based loss functions are also sensitive to small deformations and shifts of the edges penalizing location more heavily than actual structure. Hence, we propose a homology-based detector of local structural difference between two edge maps with a tolerable deformation. This detector is employed as a new criterion for the training and design of data-driven approaches that focus on enhancing these structural differences. Our approaches demonstrate significant improvement on morphological segmentation of foraminifera when considering region-based and topology-based metrics. Human ranking of the quality of the results by marine researchers also supports these findings.