Artur Dubrawski

Gleaning Knowledge from Data in the Intensive Care Unit

It is often difficult to accurately predict when, why, and which patients develop shock, because signs of shock often occur late, once organ injury is already present. Three levels of aggregation of information can be used to aid the bedside clinician in this task: analysis of derived parameters of existing measured physiologic variables using simple bedside calculations (functional hemodynamic monitoring); prior physiologic data of similar subjects during periods of stability and disease to define quantitative metrics of level of severity; and libraries of responses across large and comprehensive collections of records of diverse subjects whose diagnosis, therapies, and course is already known to predict not only disease severity, but also the subsequent behavior of the subject if left untreated or treated with one of the many therapeutic options. The problem is in defining the minimal monitoring data set needed to initially identify those patients across all possible processes, and then specifically monitor their responses to targeted therapies known to improve outcome. To address these issues, multivariable models using machine learning data-driven classification techniques can be used to parsimoniously predict cardiorespiratory insufficiency. We briefly describe how these machine learning approaches are presently applied to address earlier identification of cardiorespiratory insufficiency and direct focused, patient-specific management.

Using Supervised Machine Learning to Classify Real Alerts and Artifact in Online Multisignal Vital Sign Monitoring Data.

Objective: The use of machine-learning algorithms to classify alerts as real or artifacts in online noninvasive vital sign data streams to reduce alarm fatigue and missed true instability. Design: Observational cohort study. Setting: Twenty-four–bed trauma step-down unit. Patients: Two thousand one hundred fifty-three patients. Intervention: Noninvasive vital sign monitoring data (heart rate, respiratory rate, peripheral oximetry) recorded on all admissions at 1/20 Hz, and noninvasive blood pressure less frequently, and partitioned data into training/validation (294 admissions; 22,980 monitoring hours) and test sets (2,057 admissions; 156,177 monitoring hours). Alerts were vital sign deviations beyond stability thresholds. A four-member expert committee annotated a subset of alerts (576 in training/validation set, 397 in test set) as real or artifact selected by active learning, upon which we trained machine-learning algorithms. The best model was evaluated on test set alerts to enact online alert classification over time. Measurements and Main Results: The Random Forest model discriminated between real and artifact as the alerts evolved online in the test set with area under the curve performance of 0.79 (95% CI, 0.67–0.93) for peripheral oximetry at the instant the vital sign first crossed threshold and increased to 0.87 (95% CI, 0.71–0.95) at 3 minutes into the alerting period. Blood pressure area under the curve started at 0.77 (95% CI, 0.64–0.95) and increased to 0.87 (95% CI, 0.71–0.98), whereas respiratory rate area under the curve started at 0.85 (95% CI, 0.77–0.95) and increased to 0.97 (95% CI, 0.94–1.00). Heart rate alerts were too few for model development. Conclusions: Machine-learning models can discern clinically relevant peripheral oximetry, blood pressure, and respiratory rate alerts from artifacts in an online monitoring dataset (area under the curve > 0.87).

Leveraging Publicly Available Data to Discern Patterns of Human-Trafficking Activity

We present a few data analysis methods that can be used to process advertisements for escort services available in public areas of the Internet. These data provide a readily available proxy evidence for modeling and discerning human-trafficking activity. We show how it can be used to identify advertisements that likely involve such activity. We demonstrate its utility in identifying and tracking entities in the Web-advertisement data even if strongly identifiable features are sparse. We also show a few possible ways to perform community- and population-level analyses including behavioral summaries stratified by various types of activity and detection of emerging trends and patterns.

Detection of Events In Multiple Streams of Surveillance Data

Simultaneous monitoring of multiple streams of data that carry corroborating evidence can be beneficial in many event detection applications. This chapter reviews analytic approaches that can be employed in such scenarios. We cover established statistical algorithms of multivariate time series baseline estimation and forecasting. They are relevant when multiple streams of data can be modeled jointly. We then present more recent methods which do not have to rely on such an assumption. We separately address techniques that deal with data in a specific form of a record of transactions annotated with multiple descriptors, often encountered in the practice of health surveillance. Future event detection algorithms will benefit from incorporation of machine learning methodology. This will enable adaptability, utilization of human feedback, and building reliable detectors using some examples of events of interest. That will lead to highly scalable and economical multi-stream event detection systems.

Stochastic validation for automated tuning of neural network's hyper-parameters

Abstract In this paper we describe a new method for automated tuning of hyper-parameters of supervised learning systems. It uses memory-based learning principles, follows certain ideas of experimental design and employs an alternative approach to resampling called stochastic validation. The described method allows not only for an efficient search through a decision space, but also for a corcurrent validation of the learning algorithm performance on a given data. Potential usefulness of the proposed approach is illustrated with the Fuzzy-ARTMAP neural network application to learning a qualitative positioning of an indoor mobile robot equipped with ultrasonic range sensors. Automatically selected neural network setpoints reach a comparable performance to those achieved by human experts in two-dimensional parameter optimization cases. Migration of the proposed method to higher-order optimization domains bears a big promise and requires further research.

Affordable System for Rapid Detection and Mitigation of Emerging Diseases

South and South-East Asian countries are currently in the midst of a new epidemic of Dengue Fever. This paper presents disease surveillance systems in Sri Lanka and India, monitoring a handful of communicable diseases termed as notifiable. These systems typically require 15-30 days to communicate field data to the central Epidemiology Units, to be then manually processed Prashant & Waidyanatha, 2010. Currently used analyses rely on aggregating counts of notifiable disease cases by district, disease, and week. The Real-Time Biosurveillance Program RTBP, a multi-partner initiative, aims at addressing those challenges by developing affordable paradigm-changing Information Communication Technology ICT, implementing and field-testing them in India and Sri Lanka. Key components of the proposed solution include real-time digitization of clinical information at hospitals and clinics with the mHealthSurvey mobile phone software Kannan et al., 2010, detecting anomalies in large multivariate biosurveillance data using the T-Cube Web Interface spatio-temporal statistical analysis tool Ray et al., 2008, and disseminating critical information pertaining to the adverse events to healthcare workers using the Sahana Alerting Module Sampath et al., 2010. This paper provides an overview of the applications and discusses utility of the technologies for real-time detection and mitigation of emerging threats to public health.

Learning temporal rules to forecast instability in continuously monitored patients

Inductive machine learning, and in particular extraction of association rules from data, has been successfully used in multiple application domains, such as market basket analysis, disease prognosis, fraud detection, and protein sequencing. The appeal of rule extraction techniques stems from their ability to handle intricate problems yet produce models based on rules that can be comprehended by humans, and are therefore more transparent. Human comprehension is a factor that may improve adoption and use of data-driven decision support systems clinically via face validity. In this work, we explore whether we can reliably and informatively forecast cardiorespiratory instability (CRI) in step-down unit (SDU) patients utilizing data from continuous monitoring of physiologic vital sign (VS) measurements. We use a temporal association rule extraction technique in conjunction with a rule fusion protocol to learn how to forecast CRI in continuously monitored patients. We detail our approach and present and discuss encouraging empirical results obtained using continuous multivariate VS data from the bedside monitors of 297 SDU patients spanning 29 346 hours (3.35 patient-years) of observation. We present example rules that have been learned from data to illustrate potential benefits of comprehensibility of the extracted models, and we analyze the empirical utility of each VS as a potential leading indicator of an impending CRI event.

Detection of radioactive sources in urban scenes using Bayesian Aggregation of data from mobile spectrometers

Mobile radiation detector systems aim to help identify dangerous sources of radiation while minimizing frequency of false alarms caused by non-threatening nuisance sources prevalent in cluttered urban scenes. We develop methods for spatially aggregating evidence from multiple spectral observations to simultaneously detect and infer properties of threatening radiation sources.Our Bayesian Aggregation (BA) framework allows sensor fusion across multiple measurements to boost detection capability of a radioactive point source, providing several key innovations previously unexplored in the literature. Our method learns the expected Signal-to-Noise Ratio (SNR) trend as a function of source exposure using Bayesian nonparametrics to enable robust detection. The method scales well in spatial search by leveraging conditional independence and locality in Bayesian updates. The framework also allows modeling of source parameters such as intensity or type to enable property characterization of detected sources. Approaches for incorporating modeling information into BA are compared and benchmarked with respect to other data fusion techniques. HighlightsBayesian Aggregation fuses multiple observations to detect radiation sources.Source intensity and type information can be incorporated in detection.Properties of detected sources can be inferred for appropriate response.

Trade-offs between Agility and Reliability of Predictions in Dynamic Social Networks Used to Model Risk of Microbial Contamination of Food

This paper evaluates trade-offs between agility and reliability of predictions arising due to sparseness of data modeled with dynamic social networks. We use real field data from food safety domain to illustrate the discussion. We model food production facilities as one type of entities in a social network evolving in time. Another type of entities denotes various specific strains of Salmonella. Two entities are linked in the graph if a microbial test of food sample conducted at the specific food facility over specific period of time turns out positive for the particular pathogen. We use a computationally efficient latent space model to predict future occurrences of pathogens in individual facilities. Empirical results indicate predictive utility of the proposed representation. However, sparseness of data limits the attainable agility of predictions. We identify exploiting recency of data and using the known patterns in it, such as seasonality, as plausible means of battling the challenge of sparseness.

A study into detection of bio-events in multiple streams of surveillance data

This paper reviews the results of a study into combining evidence from multiple streams of surveillance data in order to improve timeliness and specificity of detection of bio-events. In the experiments we used three streams of real food- and agriculture-safety related data that is being routinely collected at slaughter houses across the nation, and which carry mutually complementary information about potential outbreaks of bio-events. The results indicate that: (1) Non-specific aggregation of p-values produced by event detectors set on individual streams of data can lead to superior detection power over that of the individual detectors, and (2) Design of multi-stream detectors tailored to the particular characteristics of the events of interest can further improve timeliness and specificity of detection. In a practical setup, we recommend combining a set of specific multi-stream detectors focused on individual types of predictable and definable scenarios of interest, with nonspecific multi-stream detectors, to account for both anticipated and emerging types of bio-events.