"Decisions are made at every moment in the ICU. What if the decision-maker had access to all relevant data for every such decision? The relevant data include everything from this patient's current physiological readings to the results of drug trials done years before. Machine learning can bring all these data to bear in a usable form. Of the many tools in the machine learning toolbox, model-based methods hold the greatest promise. In the not-too-distant future they may be able to support accurate real-time diagnosis and the synthesis of complex intervention strategies. On the other hand, no matter how much data we have, uncertainty -- and with it the need for continuous observation -- will be unavoidable and must be embraced by technology."

"Predicting tissue function based upon an individual's unique cells requires a multiscale Systems Biology approach to understand the coupling of intracellular signaling with spatiotemporal gradients of extracellular biochemicals controlled by convective-diffusive transport. During thrombotic or hemostatic episodes, platelets bind collagen and release ADP and thromboxane A2 (TXA2) to facilitate the recruitment of additional platelets to a growing deposit that distorts the flow field. Using high throughput experimentation, we obtained a large set of platelet calcium responses to combinatorial activators in order to train a neural network (NN) model of platelet activation for several individuals. Each NN model was then embedded into a kinetic Monte Carlo/finite element/lattice Boltzmann simulation of stochastic platelet deposition under flow. Simulations predicted the unique clot buildup dynamics for each donor and responses to various pharmacological inhibitors (measured in microfluidic assays). Consistent with measurement and simulation, one donor displayed a gain of function phenotype, while another donor was distinguished by combined aspirin-resistance and U46619-insensitivity, consistent with a thromboxane receptor mutation. In silico representations of an individual's platelet phenotype allows prediction of blood function, essential to prioritizing patient-specific cardiovascular risk and drug response or to identify unsuspected gene mutations."

Introduction: The disease state of the critically ill is described by complex data from multiple sources. For the clinician the challenge is to distinguish clinically relevant changes from noise and artifacts. In this contribution we will give an overview and present some examples of robust statistical signal extraction methods.
Requirements and methods: Algorithms for use in clinical real-time environments have to be robust against artifacts and missing values as well as efficient and fast. Online use demands instantaneous updates with every new incoming value. Many different univariate and multivariate algorithms have been proposed for use in such situations, including median filters, Kalman filters, different robust filters, graphical models, multivariate regression as well as approaches from artificial intelligence. While some of these approaches showed promising results, none has gained universal acceptance or even commercial implementation.
Examples from vital signs monitoring: Robust regression techniques represent one approach to signal extraction. Repeated median regression seems to be the best choice for intensive care monitoring because of the quality of signal extraction and the favorable computational demands. We expanded this approach to multivariate time series using an adaptive online trimmed repeated median least squares filter. In intensive care monitoring data the univariate filters can eliminate a major percentage of false positive alarms without compromising alarm sensitivity in a clinically relevant manner.
Conclusions: Recognizing true failure in the critically ill requires fast and reliable interpretation of complex data with high sensitivity and specificity. Currently commercially available methods are often inadequate for this task. Statistical research has made new univariate and multivariate alarm algorithms available that can provide better and more robust alarm detection.

The expansion of electronic medical records (EMR) provides an unprecedented opportunity to use syndrome surveillance technology for the development of "smart alarms" or sniffers that improve the safety of critically ill hospitalized patients. Preliminary studies which use rule-based alerts and/or computer-generated messages directed at clinicians demonstrate a decrease in patient length of hospital stay and decreased response time before the physician institutes appropriate treatment. However the timeliness of an intervention depends critically not only on syndrome pattern recognition but vigilant follow up of a system's delivery of care. Unfortunately, failure-to-recognize and failure-to-rescue errors are a common and extremely important cause of morbidity and mortality in the ICU. Monitoring for errors of this type is a task that requires a degree of vigilance in addition to pattern detection. The ICU electronic environment contains many markers of human agents' actions and these may be used to detect errors of omission. During this presentation, speaker will address "Failure to rescue" concept in clinical diagnostic alerts and also will outline the problem of information overload from unnecessary multiple clinical alerts while presenting some key potential steps to address this issue.

This presentation will describe the INtelligent VENTilator project (INVENT), the goal of which is to build, evaluate and integrate into clinical practice, a model-based decision support system for control of mechanical ventilation. Models included in INVENT will be presented including those of pulmonary gas exchange, and acid-base status of blood along with decision-theoretic penalty functions for balancing the competing goals of mechanical ventilation. The level of model complexity to enable patient specific tuning will be discussed and examples will be given of patients cases to illustrate the potential of model-based decision support in this field. In addition, examples will be shown of how models often lead to new ideas, scientific projects and commercial applications. These examples will include a system for measuring pulmonary gas exchange and a system for mathematical arterialisation of the acid-base and oxygen status of peripheral venous blood; removing the need for painful arterial punctures.

There are both potential rewards and pitfalls to attempts at formal descriptions of biological systems. The rewards are obvious: as a science almost entirely grounded on empiricism, biological knowledge does not lend itself to the scalable analysis that is a primary benefit of theory. Each biological system becomes a subject of essentially unique study, only very weakly linked to other "similar" systems through poorly characterized analogy. The establishment of a formal description of biological systems would greatly enhance the ability to generate and utilize biological knowledge, and usher in a modern age of biological investigation. However, there are significant limitations to existent attempts at creating biological formalisms, such as Rosen's M-R systems and topological analyses. While effective in terms of description, they are lack the ability to instantiate the essential dynamics of biology, and therefore do not provide a means of translating insight gained through their use to overcome the relative empirical uniqueness of biological systems. They lack in utility to map back onto the individual instances of the biological systems they purport to represent in a way that obviates such a high degree of empirical description that essentially renders them superfluous. Attempts of formal descriptions of biology need to remain, counter intuitively, grounded in the physical substance of biology, and the primary generative forces associated with biology, namely evolution and its relationship with the Second Law of Thermodynamics.

Like many biological processes, inflammation and its various manifestations in disease are multi-dimensional. The advent of multiplexed platforms for gathering biological data, while providing an unprecedented level of detailed information about complex biological systems such as the inflammatory response, has paradoxically also flooded investigators with data they are unable to use in the reductionist, linear paradigm of hypothesis generation and testing. Herein, we discuss the use of dimensionality reduction techniques as standalone methods with both mechanistic and translational applications in trauma/hemorrhage, traumatic brain injury, and sepsis. We also describe the incorporation of these techniques into mechanistic simulations of the acute inflammatory response, and the use of such simulations within a paradigm of Translational Systems Biology.

Objectives: In this work, we present a neuroadaptive control architecture to control lung volume and minute ventilation with input pressure constraints that also accounts for spontaneous breathing by the patient.
Methods: Specifically, we develop a pressure- and work-limited neuroadaptive controller for mechanical ventilation based on a nonlinear multi-compartmental lung model. The control framework does not rely on any averaged data and is designed to automatically adjust the input pressure to the patient's physiological characteristics capturing lung resistance and compliance modeling uncertainty. Moreover, the controller accounts for input pressure constraints as well as work of breathing constraints. Finally, the effect of spontaneous breathing is incorporated within the lung model and the control framework.
Results: It is shown that the delivered air volume using nominal controllers significantly exceeds the desired values in the absence of adaptation, whereas satisfactory tracking of the desired air volume is achieved with adaptation. Failure to adequately regulate the mode and parameters of ventilatory support can result in failure to oxygenate, failure to achieve adequate lung expansion, or overexpansion of the lung resulting in lung tissue rupture. These problems oftentimes occur when open-loop volume-control or pressure control is employed, or when averaged respiratory data is used to choose the parameters for a closed-loop ventilation control algorithm. In contrast, the proposed neuroadaptive control algorithm avoids reliance on average respiratory data and achieves system performance without excessive reliance on system model parameters.
Conclusions: Acute respiratory failure due to infection, trauma, and major surgery is one of the most common problems encountered in intensive care units and mechanical ventilation is the mainstay of supportive therapy for such patients. In particular, mechanical ventilation of a patient with respiratory failure is a critical life-saving procedure performed in the intensive care unit. Failure to adequately regulate the mode and parameters of ventilatory support can result in failure to oxygenate, failure to achieve adequate lung expansion, or overexpansion of the lung resulting in lung tissue rupture. In this research, we developed a neuroadaptive control algorithm for mechanical ventilation to control lung volume and minute ventilation. The adaptive controller accounts for input pressure constraints as well as work of breathing constraints in the face of lung resistance and compliance model uncertainty.

This talk will address three important uses of prediction models in the Intensive Care Unit (ICU), with focus on mortality prediction. I will review our work on using prediction models for benchmarking ICUs, in which model predictions of an ICU¹s outcomes are compared to the ICU¹s actual outcomes. I will demonstrate how changes over time in the performance of these models impact the assessment of quality of care of ICUs. Next I will explain how discovered frequent patterns in organ functioning status can be used to provide daily predictions to help make treatment decisions for individual patients. I will show how these predictions compare to those made by ICU nurses and physicians. Finally I will present applications of subgroup discovery algorithms for identifying ³interesting² patient subgroups that behave markedly differently than the rest. I will reveal a condition in which a simple classification tree algorithm performs better than an algorithm specifically designed for sub-group discovery. The talk will provide promising future research directions in prognostic modeling.

For each patient, critical care physicians are confronted with hundreds of disjoint variables collected by patient monitors and general clinical information systems. To cope with this data overload, clinicians need assistance to: 1) gather and pool together patient data from disparate devices and clinical information systems; 2) automatically analyze the data in a timely manner to extract clinically relevant information; and 3) efficiently interface with the data and the analytics to make quick accurate clinical decisions. Our research addresses these problems with the design of the Online Healthcare Analytics infrastructure (a.k.a. Artemis), and its extensions for exploration. OHA is a framework for real-time analysis in intensive care leveraging IBM InfoSphere Streams (Streams), a state-of-the-art stream computing platform that we developed at IBM Research.
From a data integration perspective, OHA is designed to interface with an open set of devices and clinical information management systems through well defined adaptation mechanisms. A tight integration with the Excel Medical BedMasterEX system is enabled with appropriate adapters to acquire physiologic data from patient monitors.
From an analytical perspective, OHA has been initially designed to operate in an open loop configuration where analysts specify explicitly real-time analytics for the detection of specific patterns in patient data. Physicians and researchers at SickKids, Toronto and at the University of Ontario Institute of Technology are leveraging OHA in this configuration to collect data, mine them externally and test clinical hypotheses to come up with new ways to detect nosocomial infections in neonatal ICUs. Recently, we have been extending OHA with exploration capabilities and have designed an Exploration Platform for Intensive Care (EPIC). EPIC is a closed loop extension of OHA. It provides assistance to analysts in the exploration of patient data, to accelerate the discovery of interesting patterns in streaming data. While working with the Columbia Medical Center Neuro-ICU, we are developing EPIC for application to the early detection of complications like delayed cerebral ischemia.
Users interact with EPIC and OHA at different levels of abstraction. At the lowest level, analysts program real-time analytics in the Stream Processing Language (SPL), a generic programming language for the specification of streaming analytics. At higher levels of abstraction analytics can be composed from pre-defined analytical SPL building blocks, shielding domain experts from SPL programming and allowing them to quickly compose and deploy streaming analytics.

Critical care units internationally boast state of the art medical equipment that constantly monitor vital organs. However, they are at a critical crossroad where their ability to gather this information has outpaced their ability to aggregate and interpret in a clinically meaningful way. Recent medical research has reported that conditions such as late onset neonatal sepsis, pneumothorax, intraventricular haemorrhage, and periventricular leukomalacia, appear to exhibit early pathophysiological indicators in physiological data. However research in this domain has been limited to either patient centric, diagnosis centric or physiological stream centric approaches with a heavy emphasis on retrospective analysis and minimal translation to real-time monitoring. This research presents Artemis, a framework for concurrent multi-patient, multi- diagnosis and multi-stream (ie multidimensional) temporal analysis in real time for clinical management and historically for clinical research. The real-time component utilizes new stream processing approaches while the clinical research utilizes new approaches to data mining more suited to the analysis of physiological stream behaviours and within the healthcare context. Critical care patients in rural, remote and some urban healthcare facilities do not have the same level of access to intensivist support as patients in higher care level urban critical care units (CCUs). The provision of clinical decision support tools, in a cost effective way to all CCUs has the potential to reduce mortality and morbidity rates, reduce critical care patient transportation between CCUs and in so doing reduce healthcare costs. This research also presents Artemis Cloud, a cloud computing based Software-as-a-Service and Data-as-a-Service approach for the provision of remote real-time patient monitoring and support for clinical research. This research is demonstrated using a neonatal intensive care unit case study.