Critical care units are dynamic, complex and resource intense environments, where humans directly interface with technology. Decisions are time-sensitive and often occur in circumstances where clinicians need to quickly aggregate and integrate data from multiple sources. In turn there can be variability in practice and uncertainty in management, whether it be related to the patient disease and diagnosis, the patient physiologic state and unpredictable responses to management, and to the ability of clinical teams to interpret data correctly. In addition, there are numerous competing pressures inherent to any ICU environment around work flow, communication, distraction and resource utilization.
The continuous physiologic data streaming from devices and monitors are time-series data, data in motion. They are characterized by a number of features: huge in volume and velocity, signals are variable in frequency and subject to considerable artifacts. No question that these data are
essential for supporting decision at the bedside, but when it comes to modelling for the purpose of real-time (bedside) predictive analyses, the ability to capture, label and integrate these data is limited. The data is messy; hard to manage, store and retrieve. The problem of volume has placed constraints around data storage and retrieval, and there are bottlenecks as the data is input/output (I/O) bound. In our department, we have built and deployed a bespoke data management platform to facilitate collection, file indexing, compression and decompression (see www.laussenlabs.ca). As an indication of the scale of continuous physiologic data that is measured and can be collected, and depending on the complexity of disease and treatment, there are routinely 500+ signals per hour and between 70-150 million physiologic data
points per day generated in the 42 beds of our paediatric intensive care unit. Over 700,000 hours and 2 trillion data points from over 4000 patients is currently stored in the database.
Data Analysis
There are broad categories for using these data in critical care. Some of these include: 1) Describing the physiologic phenotype and individualized physiology according to age, disease, treatment and time, 2.) Understand the physiologic state, such as risk for low oxygen delivery, hemodynamicinstability, effectiveness of mechanical ventilation, risk for neurologic injury, and metabolic state, 3.) Develop early warning systems and the risk for an event within a physiologic state through recognizing patterns in the data, such as risk for sepsis or cardiac arrest, 4.) Track the trajectory of a patient in response to treatment protocols, and direct specific interventions according to a change in the expected trajectory, 5.) Develop decision support and business analytic tools to ensure the efficiencies of care against validated outcome metrics, such as length of stay and risk for readmission, 6.) Enhanced signal processing and waveform analysis such as to diagnose changes in heart rhythm and uncover previously hidden signals that may be embedded within composite waveforms, and 7.) Develop new insights into the underlying physiology, tease out sub-populations of patients that may respond to a particular treatment, and develop prognostic and predictive enrichment strategies that can lead to individualized and precise critical care management. The promise and problem with physiologic data.
There is a disconnect between how we make decisions and the trust we might otherwise have in a model or algorithm. Models can be opaque, containing weighted features, components and simplifications with blind spots related to the inputs and the priorities and judgement of their 11 creators. They are mathematical outputs that may not take into consideration of conditions and behaviours. The risk therefore may be incorrect assumptions and spurious correlations, reinforced and contaminated by bias. Models need feedback of mistakes and of the results and outcomes; they need to be explainable, scalable and context sensitive.
It is nevertheless possible to utilize the data generated by continuous physiologic signals at the bedside to help us understand physiologic states and phenotypes in critical care. At the same time, it is important to also understand that using big physiologic data to determine these states will not replace the clinician at the bedside, rather it will augment our decision making, improve communication and information transfer.