Standard electrocardiography (ECG) allows to record the electrical activity of the heart from different angles called leads. The QT interval, which corresponds to the time elapsed between the onset of ventricular contraction and the end of ventricular relaxation, is an ECG biomarker of drug cardiotoxicity. Deep neural networks (DNNs) have achieved state-of-the-art performance in QT interval measurement but are missing uncertainty quantification, which is necessary for safer decision making. Uncertainty is usually encoded in DNNs through probability distributions over model weights. In this paper, we combine this approach with notions of multisensory integration whereby neural systems account for uncertainty by optimally integrating all available sensory inputs. We thus approximate the posterior predictive distribution of the QT interval given a multi-lead ECG as a weighted average across leads (lead integration) and models (deep ensembling) and derive 100(1 − α)% Bayesian prediction intervals (PIs). We apply this method to QT-based cardiac drug safety monitoring and compare it to an adapted version of conformal prediction. The Bayesian and conformal approaches yield comparable empirical coverage (77%-82% for mean PI widths of ∼28 milliseconds, α = 0.1). The former is more straightforward and shows better error-based calibration. Data and code implementation are available at https://github.com/mouslyddiaw/qt-uncertainty.

Continuous-time models such as Neural ODEs and Neural Flows have shown promising results in analyzing irregularly sampled time series frequently encountered in electronic health records. Based on these models, time series are typically processed with a hybrid of an initial value problem (IVP) solver and a recurrent neural network within the variational autoencoder architecture. Sequentially solving IVPs makes such models computationally less efficient. In this paper, we propose to model time series purely with continuous processes whose state evolution can be approximated directly by IVPs. This eliminates the need for recurrent computation and enables multiple states to evolve in parallel. We further fuse the encoder and decoder with one IVP-solver based on its invertibility, which leads to fewer parameters and faster convergence. Experiments on two EHR datasets show that the proposed approach achieves comparable classification performance while gaining more than 10x speedup over other continuous-time counterparts.

With a growing interest in sleep monitoring at home, sound-based sleep staging with deep learning has emerged as a potential solution. However, collecting labeled data is restrictive in the home environments due to the inconvenience of installing medical equipment at home. To handle this, we propose novel training approaches using accessible real-world sleep sound data. Our key contributions include a new semi-supervised learning technique called sequential consistency loss that considers the time-series nature of sleep sound and a semi-supervised contrastive learning method which handles out-of-distribution data in unlabeled home recordings. Our model was evaluated on various datasets including a labeled home sleep sound dataset and the public PSG-Audio dataset, demonstrating the robustness and generalizability of our model across real-world scenarios.

Survival analysis is a crucial semi-supervised task in machine learning with numerous real-world applications, particularly in healthcare. Currently, the most common approach to survival analysis is based on Cox's partial likelihood, which can be interpreted as a ranking model optimized on a lower bound of the concordance index. However, this interpretation of Cox's partial likelihood considers only pairwise comparisons. Recent work has developed differentiable sorting methods which relax this pairwise independence assumption, enabling the ranking of sets of samples. However, current differentiable sorting methods can not account for censoring, a key factor in many real-world datasets. To address this limitation, we propose a novel method called \emph{diffsurv}. We extend differentiable sorting methods to handle censored survival analysis tasks by predicting matrices of possible permutations that take into account the uncertainty introduced by censored samples. We contrast this approach with methods derived from partial likelihood and ranking losses. Our experiments show that diffsurv outperforms established baselines in various simulated and real-world risk prediction scenarios. Additionally, we demonstrate the benefits of the algorithmic supervision enabled by diffsurv by presenting a novel method for top-k risk prediction that outperforms current methods.

Analyzing and grouping typical patient trajectories is crucial to understanding their health state, estimating their prognosis, and determining optimal treatment. The increasing availability of electronic health records (EHRs) opens the opportunity to support clinicians in their decisions with machine learning solutions. We propose the Multi-scale Health-state Variational Auto-Encoder (MHealthVAE) to learn medically informative patient representations and allow meaningful subgroup detection from sparse EHRs. We derive a novel training objective to better capture health information and temporal trends into patient embeddings and introduce new performance metrics to evaluate the clinical relevance of patient clustering results.

Missing data is a major challenge in clinical research. In electronic medical records, often a large fraction of the values in laboratory tests and vital signs are missing. The missingness can lead to biased estimates and limit our ability to draw conclusions from the data. Additionally, many machine learning algorithms can only be applied to complete datasets. A common solution is data imputation, theprocess of filling-in the missing values. However, some of the popular imputation approaches perform poorly on clinical data. We developed a simple new approach, Time-Dependent Iterative imputation (TDI), which offers a practical solution for imputing time-series data. It addresses both multivariate and longitudinal data, by integrating forward-filling and Iterative Imputer. The integration employs a patient, variable, and observation-specific dynamic weighting strategy, based on the clinical patterns of the data, including missing rates and measurement frequency. We tested TDI on randomly masked clinical datasets. When applied to a cohort consisting of more than 500,000 patient observations from MIMIC III, our approach outperformed state-of-the-art imputation methods for 25 out of 30 clinical variables, with an overall root-mean-squared-error of 0.63, compared to 0.85 for SoftImpute, the second best method. MIMIC III and COVID-19 inpatient datasets were used to perform prediction tasks. Importantly, these tests demonstrated that TDI imputation can lead to improved risk prediction.

Explainable artificial intelligence (AI) techniques are increasingly being explored to provide insights into why AI and machine learning (ML) models provide a certain outcome in various applications. However, there has been limited exploration of explainable AI techniques on time-series data, especially in the healthcare context. In this paper, we describe a threshold-based method that utilizes a weakly supervised model and a gradient-based explainable AI technique (i.e. saliency map) and explore its feasibility to identify salient frames of time-series data. Using the dataset from 15 post-stroke survivors performing three upper-limb exercises and labels on whether a compensatory motion is observed or not, we implemented a feed-forward neural network model and utilized gradients of each input on model outcomes to identify salient frames that involve compensatory motions. According to the evaluation using frame-level annotations, our approach achieved a recall of 0.96 and an F2-score of 0.91. Our results demonstrated the potential of a gradient-based explainable AI technique (e.g. saliency map) for time-series data, such as highlighting the frames of a video that therapists should focus on reviewing and reducing the efforts on frame-level labeling for model training.

Human electroencephalography (EEG) is a brain monitoring modality that senses cortical neuroelectrophysiological activity in high-temporal resolution. One of the greatest challenges posed in applications of EEG is the unstable signal quality susceptible to inevitable artifacts during recordings. To date, most existing techniques for EEG artifact removal and reconstruction are applicable to offline analysis solely, or require individualized training data to facilitate online reconstruction. We have proposed CLEEGN, a light-weight convolutional neural network for plug-and-play automatic EEG reconstruction. CLEEGN is based on a subject-independent pre-trained model using existing data and can operate on a new user without any further calibration. The results of simulated online validation suggest that, even without any calibration, CLEEGN can largely preserve inherent brain activity and outperforms leading online/offline artifact removal methods in the decoding accuracy of reconstructed EEG data.

Continuous-time models such as Neural ODEs and Neural Flows have shown promising results in analyzing irregularly sampled time series frequently encountered in electronic health records. Based on these models, time series are typically processed with a hybrid of an initial value problem (IVP) solver and a recurrent neural network within the variational autoencoder architecture. Sequentially solving IVPs makes such models computationally less efficient. In this paper, we propose to model time series purely with continuous processes whose state evolution can be approximated directly by IVPs. This eliminates the need for recurrent computation and enables multiple states to evolve in parallel. We further fuse the encoder and decoder with one IVP-solver based on its invertibility, which leads to fewer parameters and faster convergence. Experiments on two EHR datasets show that the proposed approach achieves comparable classification performance while gaining more than 10x speedup over other continuous-time counterparts.

Irregularly sampled time series are becoming increasingly prevalent in various domains, especially in medical applications. Although different highly-customized methods have been proposed to tackle irregularity, how to effectively model their complicated dynamics and high sparsity is still an open problem. This paper studies the problem from a whole new perspective: transforming irregularly sampled time series into line graph images and adapting powerful vision transformers to perform time series classification in the same way as image classification. Our approach largely simplifies algorithm designs without assuming prior knowledge and can be potentially extended as a general-purpose framework. Despite its simplicity, we show that it substantially outperforms state-of-the-art specialized algorithms on several popular healthcare and human activity datasets. Our code and data are anonymously available at \url{https://anonymous.4open.science/r/ViTST-TSRL4H-ICLR2023}.

Wearable health devices have revolutionized our ability to continuously analyze human behavior and build longitudinal statistical models around illness by measuring physiological indicators like heart rate over several months of an individual's life. Shifts in these indicators have been correlated with the onset of illnesses such as COVID-19, leading to the development of Wearable-Based Illness Detection (W-BID) models that aim to detect the onset of illness. While W-BID models accurately detect illness, they often over-predict illness during healthy time periods due to variance caused by seemingly random human choices. However, it is because W-BID models treat each input window as independent and identically distributed samples that we are unable to account for the weekly structure of variance that causes false positives. Towards preventing this, we propose a system for identifying structural variance in wearable signals and measuring the effect they have on W-BID models. We demonstrate how a simple statistical model that does not account for weekly structure is strongly biased by weekly structure, with a Pearson correlation coefficient of 0.899.

Multivariate biosignals are prevalent in many medical domains. Modeling multivariate biosignals is challenging due to (1) long-range temporal dependencies and (2) complex spatial correlations between electrodes. To address these challenges, we propose representing multivariate biosignals as time-dependent graphs and introduce GraphS4mer, a general graph neural network (GNN) architecture that models spatiotemporal dependencies in multivariate biosignals. Specifically, (1) we extend the Structured State Spaces architecture, a state-of-the-art deep sequence model, to capture long-range temporal dependencies in graph data and (2) we propose a graph structure learning layer to learn dynamically evolving graph structures in the data. We evaluate our proposed model on three distinct tasks and show that GraphS4mer consistently improves over existing models, including (1) seizure detection from electroencephalography signals, outperforming a previous GNN with self-supervised pre-training by 3.1 points in AUROC; (2) sleep staging from polysomnography signals, a 4.1 points improvement in macro-F1 score over existing sleep staging models; and (3) electrocardiogram classification, outperforming previous state-of-the-art models by 2.7 points in macro-F1 score.

Electroencephalograms (EEGs) are useful for analyzing brain activity, and spatiotemporal patterns in the EEG signal have clinical value, serving for example as biomarkers of diseases such as epilepsy. EEGs are a combination of components from multiple sources within the brain, the electrical activity of muscles, including the heart, and artifacts due to movement and external signals (e.g, line noise). Separating and classifying the sources of these components is important for analyzing the brain patterns in the EEG data. We propose \textit{bag-of-waves} (BoWav), a new feature for the classification of EEG independent components (ICs). BoWav represents the IC time series through the distribution of counts of waveforms from a learned shift-invariant dictionary based reconstruction. We found that BoWav has a promising predictive performance, outperforming the state-of-the-art method for IC classification, ICLabel, in two of three classes of interest.

The recent availability of Electronic Health Records (EHR) has allowed for the development of algorithms predicting inpatient risk of deterioration and trajectory evolution. However, prediction of disease progression with EHR is challenging since these data are sparse, heterogeneous, multi-dimensional, and multimodal time-series. Clustering is an alternative approach to identify similar groups within thepatient cohort that can be leveraged to predict patient status evolution. In particular, the identification of phenotypically separable clusters has proven very useful in improving healthcare delivery. Some clustering models have been proposed to identify phenotypically separable clusters; however, these have struggled in clinical settings characterised by several, highly imbalanced event classes. Tothat end, we propose a generative model to cluster EHR data based on identifying clinically meaningful phenotypes with regard to patient outcome prediction and physiological trajectory. We introduce a novel probabilistic method that is capable of simultaneously a) generating observation data, b) modelling temporal cluster assignments, and c) predicting admission outcomes. Our results show performancesimilar to state-of-the-art methods, with increased clustering separability, and capability to generate observation data.

Understanding how different regional networks of the brain get activated and how those activations change over time can help in identifying the onset of various neurodegenerative diseases, studying the efficacy of different treatment regimens for those illnesses, and developing brain-computer interfaces for patients with different types of disabilities. To explain dynamic brain networks, an RNN-VAE model named DyNeMo has recently been proposed. This model can take into account the whole recorded history of brain states while modeling their dynamics and is able to better capture the complexities in larger datasets than previous works. In this paper, we show that the latent representations learned by DyNeMo through unsupervised training are not sufficient for downstream classification tasks and propose a new semi-supervised model named DyNeMoC that overcomes this shortcoming. The downstream task we study is the classification of visual stimuli from MEG recordings. We show that both of our proposed variants of DyNeMoC --- DyNeMoC-RNN and DyNeMoC-Transformer --- lead to more useful latent representations for stimuli classification with the transformer variant outperforming the RNN one. Learning representations that are directly linked to a downstream task in this manner could ultimately be used to improve the monitoring and treatment of certain neurodegenerative diseases and building better brain-computer interfaces.

With the proliferation of time-series analysis in the healthcare sector, each new time series requires a new model to be fit and trained on those specific data. The vast majority of time-series forecasting approaches require a training dataset. There is very recent work on zero-shot forecasting---pretraining on one series and evaluating on another---yet its performance is inconsistent depending on the training dataset. In this work, we take a different approach and devise ForecastPFN, the first universal zero-shot model, pretrained purely on synthetic data. Drawing inspiration from TabPFN, a recent breakthrough in tabular data, ForecastPFN is the first forecasting model to approximate Bayesian inference. To accomplish this, we design a synthetic time-series distribution with local and global trends, and noise. Through experiments on multiple datasets, we show that ForecastPFN achieves competitive performance without ever seeing the training datasets, compared to popular methods that were fully trained on the training dataset.

Pretraining deep learning models on a large corpus of unlabeled data using selfsupervised learning approaches can be an efficient a mitigation strategy to dealwith the lack of annotated data. We proposed to use a siamese framework forthe pretraining of a convolutional neural network on the Computing in Cardiology2021 dataset therefore making it invariant to ECG lead configuration changes. The obtained representation was then trained and tested on a heartbeat classification task on the MIT BIH Arrhythmia database, and on an external independent set, namely the INCART database. The proposed model reached a median F1 score of 0.89 on the MIT BIH Arrhythmia database comparable to the 0.90 F1 score obtained without pretraining. However, the pretrained model obtained a median F1 score of 0.74 on average over the different leads, compared to 0.53 the model without pretraining. The proposed pretraining approach, leveraging the availability of relatively large database of un-(or weakly)annotated ECG data, allows for the training of more generalizable, lead-agnostic, heartbeat classification models. Such an approach would ensure avoiding overfitting complex deep learning models on the small MIT-BIH arrhythmia database.

Univariate high-frequency time series are dominant data sources for many medical, economic and environmental applications. In many of these domains, the time series are tied to real-time changes in state. In the intensive care unit, for example, changes and intracranial pressure waveforms can indicate whether a patient is developing decreased blood perfusion to the brain during a stroke, for example. However, most representation learning to resolve states is conducted in an offline, batch-dependent manner. In high frequency time-series, high intra-state and inter-sample variability makes offline, batch-dependent learning a relatively difficult task. Hence, we propose Spatial Resolved Temporal Networks (SpaRTeN), a novel composite deep learning model for online, unsupervised representation learning through a spatially constrained latent space. SpaRTeN maps waveforms to states, and learns time-dependent representations of each state. Our key contribution is that we generate clinically relevant representations of each state for intracranial pressure waveforms.

Wearable health devices have revolutionized our ability to continuously analyze human behavior and build longitudinal statistical models around illness by measuring physiological indicators like heart rate over several months of an individual's life. Shifts in these indicators have been correlated with the onset of illnesses such as COVID-19, leading to the development of Wearable-Based Illness Detection (W-BID) models that aim to detect the onset of illness. While W-BID models accurately detect illness, they often over-predict illness during healthy time periods due to variance caused by seemingly random human choices. However, it is because W-BID models treat each input window as independent and identically distributed samples that we are unable to account for the weekly structure of variance that causes false positives. Towards preventing this, we propose a system for identifying structural variance in wearable signals and measuring the effect they have on W-BID models. We demonstrate how a simple statistical model that does not account for weekly structure is strongly biased by weekly structure, with a Pearson correlation coefficient of 0.899.

Multivariate biosignals are prevalent in many medical domains. Modeling multivariate biosignals is challenging due to (1) long-range temporal dependencies and (2) complex spatial correlations between electrodes. To address these challenges, we propose representing multivariate biosignals as time-dependent graphs and introduce GraphS4mer, a general graph neural network (GNN) architecture that models spatiotemporal dependencies in multivariate biosignals. Specifically, (1) we extend the Structured State Spaces architecture, a state-of-the-art deep sequence model, to capture long-range temporal dependencies in graph data and (2) we propose a graph structure learning layer to learn dynamically evolving graph structures in the data. We evaluate our proposed model on three distinct tasks and show that GraphS4mer consistently improves over existing models, including (1) seizure detection from electroencephalography signals, outperforming a previous GNN with self-supervised pre-training by 3.1 points in AUROC; (2) sleep staging from polysomnography signals, a 4.1 points improvement in macro-F1 score over existing sleep staging models; and (3) electrocardiogram classification, outperforming previous state-of-the-art models by 2.7 points in macro-F1 score.

Irregularly sampled time series are becoming increasingly prevalent in various domains, especially in medical applications. Although different highly-customized methods have been proposed to tackle irregularity, how to effectively model their complicated dynamics and high sparsity is still an open problem. This paper studies the problem from a whole new perspective: transforming irregularly sampled time series into line graph images and adapting powerful vision transformers to perform time series classification in the same way as image classification. Our approach largely simplifies algorithm designs without assuming prior knowledge and can be potentially extended as a general-purpose framework. Despite its simplicity, we show that it substantially outperforms state-of-the-art specialized algorithms on several popular healthcare and human activity datasets. Our code and data are anonymously available at \url{https://anonymous.4open.science/r/ViTST-TSRL4H-ICLR2023}.