Enhancing Causal Estimation through Unlabeled Offline Data

IEEE Access, 2019

The management of patient well-being can be performed by monitoring continuous time-series vital-sign data via lowcost wearable devices. Automated algorithms may then be used with the resulting data to provide early warning of deterioration of the health of an individual. Such algorithms are typically trained for a large population without considering the time-variability and inter-subject variability of the data being collected. In the case where limited numbers of subjects are available, it is difficult to create a generalised population model from a small sample size. Furthermore, some "normal" patients may exhibit different physiological patterns when compared to other "normal" patients, forming multiple "normal" clusters/subgroups. This also makes inferring a population model difficult. It is therefore preferable to develop patient/subgroup-specific time-series models to overcome these challenges. We propose using Bayesian Hierarchical Gaussian processes to infer the hidden latent structure of the vital sign's trajectory for each individual patient or group of patients who share similar patterns. We further demonstrate the feasibility of such model in novelty detection, using the symmetric Kullback-Leibler divergence. This allows us to identify any patterns that correspond to "normal" or "abnormal" physiology, and further classifying abnormal patterns from a model of normal latent trajectories. We tested our approach using two real datasets for different monitoring scenarios. Our model was compared to the performance of state-of-the-art unsupervised clustering algorithms, demonstrating at least 10% improvement in accuracy. We further benchmarked against two one-class classifiers, and showed at least 5% accuracy improvement when using the proposed metric in identifying abnormal physiological episodes.

Statistics in Medicine, 2002

Nowadays physicians are confronted with high-dimensional data generated by clinical information systems. The proper extraction and interpretation of the information contained in such massive data sets, which are often observed with high sampling frequencies, can hardly be done by experience only. This yields new perspectives of data recording and also sets a new challenge for statistical methodology. Recently graphical models have been developed for analysing the partial correlations between the components of multivariate time series. We apply this technique to the haemodynamic system of critically ill patients monitored in intensive care. In this way we can appraise the practical value of the new procedure by re-identifying known associations within the haemodynamic system. From separate analyses for di erent pathophysiological states we can even conclude that distinct clinical states are characterized by distinct partial correlation structures. Hence, this technique seems useful for automatic statistical analysis of high-dimensional physiological time series and it can provide new insights into physiological mechanisms. Moreover, we can use it to achieve an adequate dimension reduction of the variables needed for online monitoring at the bedside.

IEEE Journal of Biomedical and Health Informatics, 2014

Cardiovascular variables such as heart rate (HR) and 6 blood pressure (BP) are regulated by an underlying control system, 7 and therefore, the time series of these vital signs exhibit rich dynam-8 ical patterns of interaction in response to external perturbations 9 (e.g., drug administration), as well as pathological states (e.g., onset 10 of sepsis and hypotension). A question of interest is whether "sim-11 ilar" dynamical patterns can be identified across a heterogeneous 12 patient cohort, and be used for prognosis of patients' health and 13 progress. In this paper, we used a switching vector autoregressive 14 framework to systematically learn and identify a collection of vital 15 sign time series dynamics, which are possibly recurrent within the 16 same patient and may be shared across the entire cohort. We show 17 that these dynamical behaviors can be used to characterize the 18 physiological "state" of a patient. We validate our technique us-19 ing simulated time series of the cardiovascular system, and human 20 recordings of HR and BP time series from an orthostatic stress 21 study with known postural states. Using the HR and BP dynamics 22 of an intensive care unit (ICU) cohort of over 450 patients from the 23 MIMIC II database, we demonstrate that the discovered cardiovas-24 cular dynamics are significantly associated with hospital mortality 25 (dynamic modes 3 and 9, p = 0.001, p = 0.006 from logistic re-26 gression after adjusting for the APACHE scores). Combining the 27 dynamics of BP time series and SAPS-I or APACHE-III provided a 28 more accurate assessment of patient survival/mortality in the hos-29 pital than using SAPS-I and APACHE-III alone (p = 0.005 and 30 p = 0.045). Our results suggest that the discovered dynamics of 31 vital sign time series may contain additional prognostic value be-Q1 32 yond that of the baseline acuity measures, and can potentially be 33 used as an independent predictor of outcomes in the ICU. 34 Index Terms-Intensive care unit, physiological control systems, 35 switching linear dynamical systems. haviors or modes, and segmentation of the time series in terms 91 of the most likely dynamic describing the time series evolution 92 at any given point in time. The proposed framework enables 93 characterization of patients in terms of the dynamical modes 94 (e.g., the average time spent within the different modes), and 95 can potentially be used to capture changes in the underlying 96 cardiovascular control systems of human subjects in response 97 to internal (such as onset of infection) and external perturba-98 tions (such as postural changes). Furthermore, we hypothesize 99 that when applied to vital sign time series of patients in a criti-100 cal care setting, the proposed technique can be used to discover 101 dynamical modes with prognostic values for predicting clinical 102 outcomes of interests. 103 A preliminary version of this study was presented at the 34th 104 Annual International Conference of the IEEE Engineering in 105 Medicine and Biology Society (EMBC '12) [14]. Here, we ex-106 tend on our previous work to include a series of validation 107 studies, and a more comprehensive assessment of the utility of 108 the time series dynamics within the ICU.

Proceedings / AMIA ... Annual Symposium. AMIA Symposium

In intensive care physiological variables of the critically ill are measured and recorded in short time intervals. The proper extraction and interpretation of the information contained in this flood of information can hardly be done by experience alone. Intelligent alarm systems are needed to provide suitable bedside decision support. So far there is no commonly accepted standard for detecting the actual clinical state from the patient record. We use the statistical methodology of graphical models based on partial correlations for detecting time-varying relationships between physiological variables. Graphical models provide information on the relationships among physiological variables that is helpful e.g. for variable selection. Separate analyses for different pathophysiological states show that distinct clinical states are characterized by distinct partial correlation structures. Hence, this technique can provide new insights into physiological mechanisms.

Proceedings Amia Annual Symposium Amia Symposium, 2001

In intensive care physiological variables ofthe critically ill are measured and recorded in short time intervals. The existing alarm systems based on fixed thresholds produce a large number of false alarms. Usually the change of a variable over time is more informative than one pathological value at a particular time point. Intelligent alarm systems which detect important changes within a physiological time series are needed for suitable bedside decision support. There are various approaches to modeling time-dependent data and also several methodologies for pattern detection in time series. We compare several methodologies designed for online detection of measurement artifacts, level changes, and trends for a proper classification of the patient's state by means of a comparative casestudy.

International Journal of Contents, 2014

Physiological signals provide important clues in the diagnosis and prediction of disease. Analyzing these signals is important in health and medicine. In particular, data preprocessing for physiological signal analysis is a vital issue because missing values, noise, and outliers may degrade the analysis performance. In this paper, we propose PhysioCover, a system that can recover missing values of physiological signals that were monitored in real time. PhysioCover integrates a gradual method and EM-based Principle Component Analysis (PCA). This approach can (1) more readily recover long-and short-term missing data than existing methods, such as traditional EM-based PCA, linear interpolation, 5-average and Missing Value Singular Value Decomposition (MSVD), (2) more effectively detect hidden variables than PCA and Independent component analysis (ICA), and (3) offer fast computation time through real-time processing. Experimental results with the physiological data of an intensive care unit show that the proposed method assigns more accurate missing values than previous methods.

Studies in fuzziness and soft computing, 2002

Clinical information systems can record numerous variables describing the patient's state at high sampling frequencies. Intelligent alarm systems and suitable bedside decision support are needed to cope with this flood of information. A basic task here is the fast and correct detection of important patterns of change such as level shifts and trends in the data. We present approaches for automated pattern detection in online-monitoring data. Several methods based on curve fitting and statistical time series analysis are described. Median filtering can be used as a preliminary step to reduce the noise and to remove clinically irrelevant short term fluctuations. Our special focus is the potential of these methods for online-monitoring in intensive care. The strengths and weaknesses of the methods are discussed in this special context. The best approach may well be a suitable combination of the methods for achieving reliable results. Further investigations are needed to further improve the methods and their performance should be compared extensively in simulation studies and applications to real data.

Computer Methods and Programs in Biomedicine, 2010

Lumped parameter approaches for modelling the cardiovascular system typically have many parameters of which a significant percentage are often not identifiable from limited data sets. Hence, significant parts of the model are required to be simulated with little overall effect on the accuracy of data fitting, as well as dramatically increasing the complexity of parameter identification. This separates sub-structures of more complex cardiovascular system models to create uniquely identifiable simplified models that are one to one with the measurements. In addition, a new concept of parameter identification is presented where the changes in the parameters are treated as an actuation force into a feed back control system, and the reference output is taken to be steady state values of measured volume and pressure. The major advantage of the method is that when it converges, it must be at the global minimum so that the solution that best fits the data is always found.

Proceedings of the ... AAAI Conference on Artificial Intelligence, 2016

ICU mortality risk prediction may help clinicians take effective interventions to improve patient outcome. Existing machine learning approaches often face challenges in integrating a comprehensive panel of physiologic variables and presenting to clinicians interpretable models. We aim to improve both accuracy and interpretability of prediction models by introducing Subgraph Augmented Non-negative Matrix Factorization (SANMF) on ICU physiologic time series. SANMF converts time series into a graph representation and applies frequent subgraph mining to automatically extract temporal trends. We then apply non-negative matrix factorization to group trends in a way that approximates patient pathophysiologic states. Trend groups are then used as features in training a logistic regression model for mortality risk prediction, and are also ranked according to their contribution to mortality risk. We evaluated SANMF against four empirical models on the task of predicting mortality or survival 30 days after discharge from ICU using the observed physiologic measurements between 12 and 24 hours after admission. SANMF outperforms all comparison models, and in particular, demonstrates an improvement in AUC (0.848 vs. 0.827, p<0.002) compared to a state-of-the-art machine learning method that uses manual feature engineering. Feature analysis was performed to illuminate insights and benefits of subgraph groups in mortality risk prediction.

Intensive Care Medicine, 1998

Today most of our bedside decisions are based on subjective judgment and experience, rather than on hard data analysis. Most of the changes of a variable over time are more important than one pathological value at the time of observation. Over the past three decades mathematical methods have been developed that allow the assessment of single or multiple variables over time.