Optimising brain perfusion in patients with brain injury is a key goal of intensive care. To do so requires use of multi-modality monitoring that can provide indices on brain physiology useful for targeting therapy for raised intracranial pressure (ICP), reduced cerebral perfusion pressure (CPP), Brain tissue oxygenation (PbrO2), Near Infrared Spectroscopy (NIRS) or other surrogate measures of cerebral perfusion.
As we are dealing with complex, time-dependent interacting physiological systems under both normal and pathological states, the BrainIT PERFUSE-IT research programme approach is to integrate three areas of research and development:
Device Assessment Research: Developing and assessing new sensors for increasing the types of time-series data streams potentially useful for clinical decision-making. As an example, view the group projects tabe to see the Glasgow pilot project described on assessing the SPECK (wireless accelerometer technology) as a means of detecting patient movement and following changes in hydrostatic pressure gradients (HPG). In this research arm, we will also be calling for projects addressing questions such as “Which form (or combination) of brain monitoring (PbrO2, NIRS, SjO2, others?) should be used as a measure(s) of Brain Perfusion?”.
Perfusion Optimisation Research: Recently, studies have shown that indices derived from inherent variations in mean arterial pressure (MAP) and ICP such as the “PrX” not only relate to clinical outcome (GOS) but can also be useful for identifying the optimal CPP (CPPopt) in patients. There is considerable interest in using these index’s for guiding therapy for raised ICP, however, there remain a number of unanswered questions in this field such as: “Which index of “cerebral autoregulation” (CAR) should be used and over which “averaging time window “ should data be analysed? See the group project tab for a description of a new project proposing comparing a number of autoregulation models over different time-scales.
Data Science Research: Increasingly we are analysing high resolution raw data streams captured directly from bed-side monitoring devices. For example, in the calculation of indices of cerebral autoregulation, and depending on the model and approach used, can often require ECG, BP and ICP data-streams sampled at > 150 Hz and processed in real-time. Hospital based computing services generally do not have the resources to handle this quantity of “Big-Data” in a time-frame capable of supporting real-world clinical decision making. As an example of an approach to address this issue, go to the Group Projects Tab where we describe a pilot Glasgow project (recently funded) called “CHART-ADAPT” (Connecting healthcare and research through a data analysis provisioning technology). This project, although not an official BrainIT project, if successful in this pilot phase, may be expanded to include BrainIT centres in the future.
The PERFUSE-IT project (approved by the Steering Group), is a long term, multi-centre research programme. It seems clear that we need to consider all three domains simultaneously (sensor development, pathophysiology research and analysis infrastructure development) in our plans for assessing approaches for improving the acute management of patients within intensive care. Such an approach is inevitably a long term one as sensor technology, data science technology and our understanding of cerebral and cerebrovascular pathophysiology are continuously developing domains.
We envisage that centres interested in participating in this research programme can do so by joining one or more of the three research arms described above. One or more steering group members will be responsible for coordinating the sub-projects within each research arm and proposals for project ideas can be submitted via this new BrainIT web-site.
Avert-IT was an EU-funded project to develop a mechanism, for use within intensive and high-dependency care units, which will have the ability to monitor and predict the likelihood of arterial hypotension (low blood pressure) adverse events. The full project title is “Advanced Arterial Hypotension Adverse Event prediction through a Novel Bayesian Neural Network” and was completed in 2012.
Intensive Care patients can experience adverse events associated with sudden episodes of low blood pressure. These adverse events may impact all of the main organs resulting in longer lengths of stay, increased care costs and reducing quality of outcomes. Existing technologies enable clinicians to know when these events have occurred and treat the effects. Medical therapies and management for treating adverse events such as low blood pressure exist but clinicians don’t have a reliable way to predict the occurrence, so there’s no opportunity for early intervention.
Research indicates average lengths of stay in intensive care could be significantly reduced if these adverse events can be avoided through prediction and earlier intervention. By reducing the cost of intensive care intensive care services (typically > 1000 Euro's per day) could lead potentially to savings across the EU in the billions of euros, annually.
As arterial hypotension is a common form of adverse event, a model for predicting these offers potential for improving outcomes across a wide range of conditions and or illnesses.
The main scientific objective of the project is the determination of the weighted association between multiple patient parameters and subsequent arterial hypotension. The association will then be used to define the novel Bayesian neural network, which will be trained against the existing BrainIT dataset (collected from 20 centres across Europe), before undertaking a clinical trial to demonstrate the Avert-IT project concept.
The main technological objective will be the development of an IT-based decision support system (“HypoPredict&rdquot;) appropriate for deployment within intensive and high dependency care units. The system will be capable of:
Having successfully trained the BANN on existing cleaned data acquired from the BrainIT group (www.brainit.org), the final stage of the project was focused on collecting data from 60 patients in an observational study to test that the AVERT-IT technology meets the clinician’s minimum requirements for sensitivity (>30%) and false positive rate (FPR) (< 10%) for prediction of arterial hypotension (low blood pressure). This work was successfully completed and the calculated sensitivity and specificity from using the BANN system in a live clinical environment were found to be 40.09% and 92.57% respectively.
In the figure above - the ROC curve from the research phase is shown by the black line. The point estimates for mean sensitivity and false positive rate at the threshold values 0.1 to 0.9 are shown on the solid red curve. The prospective clinical study results (red line) are very close to the research results particularly in the range of interest with a threshold setting of between 0.3 and 0.4 giving a specificity range of 46.03% and 40.09% and a specificity range of 87.13% and 92.57% respectively.
The BANN hypopredict technology is showing promise. We are optimistic that this research will provide the basis for future research assessing the clinical utility of this type of medical decision support upon reducing the length of stay for patients managed in the intensive care environment.
We are currently seeking funding to further develop the AVERT-IT model towards improving it's prediction sensitivity and exploiting it's technology for use within the NHS and health care sector.
One of the findings from the AvertIT project was that we estimate that approximately 30% of potential arterial hypotension events are not quantifiable due to either missing data or artifact from blood sampling, patient handling or other clinical interventions. It is felt that before implementing models such as the AvertIT BANN, we should first investigate approaches to automatically detecting and cleaning blood pressure artifact from the raw data. Towards that end, Glasgow has a project funded from the Scottish Chief Scientists Office (CSO) collaborating with (Chris Williams) Machine Learning group at Edinburgh University. They are developing two models for detection of artifact in blood pressure time-series data: a) a Factorial Switching Linear Dynamical Systems (FSLDS) model and b) a Discrimitive Switching Linear Dynamical Systems (DSLDS) model. The figure below shows examples of the two models performance for detecting blood sample artifact and arterial pressure damped trace artifact. If this project is successful in showing proof of concept for detecting and cleaning major artifact from the Blood Pressure time-series signal in real time in a live clinical environment, this will make if much more feasible to run predictive models that use the Blood Pressure channel in the intensive care setting. This project will be completed May 2015. If successful, we plan to seek further funding to assess the effectiveness of these models on a multi-centre basis.
Figure: Example of DSLDS and FSLDS inferences for a damped trace event (top) and a blood sample event (bottom). Note the Ground Zero Truth (expert manual validation) is shown in a gold colour at the bottom bar).
Following directly on from the First EU Project, This 4 year EU funded project achieved the following aims:
We have developed new software methods including several new tools for collection of the BrainIT core dataset.
We have recruited new centres into the BrainIT network acquiring patient data from 22 Neuro intensive care centres from 11 EU countries.
We have also successfully completed a prospective data collection using these new tools and the data is validated and accessible from our SQL database.
Publications arising from this project include:
Neuman J, Chambers I, Citerio G, Enblad P, Gregson B, Howells T, Mattern J, Nilsson P, Piper I, Ragauskas A, Sahuquillo J, Yau H, Kiening K on Behalf of the BrainIT Group. The use of hyperventilation therapy after Brain Injury in Europe: An analysis of the BrainIT database. Intensive Care Medicine 2008; S00134-008-1123-7
Chambers I, Gregson B, Citerio G, Enblad E, Howells T, Kiening K, Mattern J, Nilsson P & Piper I, Ragauskas A, Sahuquillo J, Yau YH on behalf of the BrainIT Group. BrainIT collaborative network: analyses from a high time-resolution dataset of head injured patients. Acta Neurochir Suppl (2008) 102: 223–227
Shaw M, Piper I Chambers I, Citerio G, Enblad P, Gregson B, Howells T, Kiening K, Mattern J, Nilsson P, Ragauskas A, Sauquillo J, Yau YH on behalf of the BrainIT Group (www.brainit.org). The brain monitoring with Information Technology (BrainIT) collaborative network: data validation results. Acta Neurochir Suppl (2008) 102: 217–221
Chambers IR, Barnes J, Piper I, Citerio G, Enblad P, Howells T et al. BrainIT - a transnational head injury monitoring research network. In: Hoff J, Keep R, Xi G, Hua Y editors. Brain Edema XIII; 2006; Ann Arbor, Michigan: Acta Neurochirugica Sup 96:, Springer Verlag; 2006
Nilsson P, Piper I, Citerio G, Chambers I, Contant C, Enblad P, Fiddes H, Howells T, Kiening K and Yau YH for the BrainIT Group. The BrainIT Group: concept and current status 2004 Acta Neurochir (2005) [Suppl] 95: 33–37
Kiening K, Schoening W, Unterberg A, Stover J, Citerio G, Enblad P, Nilsson P and the Brain-IT Group. Assessment of the relationship between age and continuous intracranial compliance Acta Neurochir (2005) [Suppl] 95: 293–297
Barnes J, Chambers I, Piper I, Citerio G, Contant C, Enblad P, Fiddes H, Howells T, Kiening K, Nilsson P, and Yau for the BrainIT Group. Accurate data collection for head injury monitoring studies: a data validation methodology. Acta Neurochir (2005) [Suppl] 95: 39–41
Photo 1: PDA Device for Entering BrainIT Coredata.
Photo 2: Web Client Software for Entering BrainIT Coredata.
Photo 3: Odin Software (Tim Howells-Uppsala) for Browsing BrainIT Coredata.
Photo 4: Odin Software (Tim Howells-Uppsala) for Showing Burden of ICP Insults.
Photo 5: BrainIT Data Disc (Latest Version on FTP Site).
This one year EC funded project allowed us to expand the group from the 5 members interested in compliance measurement to 22 centres capable of collectng research data. We defined a core-dataset standard for the collection of high resolution intensive care patient data. We conducted a paper based pilot data collection exercise to determine the feasibility of collecting the core dataset in all centres. Four group meetings enabled us to discuss group projects, a number of which have now completed. A technical sub-committee was formed which designed the interface protocols required in each centre to collect the BrainIT data. The technical sub-committee also discussed and designed a flexible database format to hold the BrainIT group data. Building upon this base, the group was successful in obtaining further EC research and infrastructure support (BrainIT-2: QLG3-CT-2002-01160) to build IT tools in order to quantify the feasibility for collection of the core-dataset in a live clinical environment from across the 22 participatng BrainIT centres.
A BrainIT core dataset definition was published: Piper I, Citerio C, Chambers I et al. The BrainIT Group: Concept and Core Dataset Definition. Acta Neurochir 145:615-629 2003.
Photo 1: Some of the BrainIT Group During one of the First EC Funded Project Meetings
In addition to a PDF based BrainIT Core Dataset Definition Document, an XML Schema definition for the core dataset has also been defined.
Figure 1: A "Snippet" from the BrainIT XML Data Schema
BrainIT now has a "Generic Research Programme Framework"" - See the "PERFUSE-IT Project Tab". Group Members can upload their own project ideas fitting in with this theme, or join existing projects run by other members. Stand-alone database analyses on existing data are also managed through this Group Project Section.
|Chris Hawthorne, Martin Shaw||Use of Cluster Analysis on the BrainIT Database.||(Database Analysis) We plan to use cluster analysis, a form of unsupervised learning, to identify similar groups of patients within the BrainIT database using: A) demographic and admission physiological variables and B) physiological variables recorded every minute during their NICU admission.|
|Milly Lo||Paediatric BrainIT (kidsBrainIT)||(Perfusion Research Arm) We aim to test the hypotheses that (1) optimal CPP varies in time and between patients in paediatric TBI in an age dependant manner; and that (2) maintaining CPP above optimal CPP improves childhood brain trauma outcome through a multi-centre study which will form a paediatric brain monitoring with information technology group (kidsBrainIT group) within the existing BrainIt group.|
|Bart Depreitere, Fabian Güiza, Geert Meyfroidt (Leuven)||Visualizing secondary CPP insults||(Database Analysis) Using exactly the same methodology as in the recent ICP insult visualization paper (Visualizing the pressure and time burden of intracranial hypertension in adult and paediatric traumatic brain injury), we will look at thresholds and time burden of both low ànd high CPP insults. Moreover, the plots will be stratified for age, pressure active/passive events and +/- decompressive craniectomy. This project is a natural sequence to the ICP insult work.|
|Matthias Hueser, M. Hueser, M. Jaggi, V. De Luca||Forecasting of intracranial hypertension using waveform features and machine learning||(Database Analysis) Background: Intracranial hypertension is an important risk factor for secondary brain insult after traumatic brain injury. In current clinical practice intracranial pressure (ICP) is continuously monitored and episodes of elevated ICP are treated after they occur. Instead a more pro-active approach would be desirable, giving doctors more time to intervene. Contributions: In this thesis we deliver as proof of this concept forecasting models for the ICP mean in >= 10 minutes, short-term trend of ICP mean and occurence of rapid increases of ICP in the near future. Our framework combines several types of features extracted from a multi-scale history of the ICP waveform: summary statistics, coefficients from spectral analysis and computer vision features on spectrograms. We then apply various machine learning models to the multitude of constructed features, among them Stochastic Gradient Descent and Decision tree ensembles. Validation: To validate our method we integrate and compare against recently proposed morphological waveform features on two independent data-sets from the MIMIC-II and Brain-IT databases, respectively. Our preliminary results show that our SGD mean forecasting model incurs a mean absolute loss of 1.05 which represents a 4-fold reduction in comparison to a naive mean regressor. Furthermore we have compared the predictivity of various subsets of the features for the forecasting targets to uncover potential precursors of elevated ICP. Discussion: Finally we discuss how our prediction system could be deployed in clinical practice and deliver a software tool that integrates the prediction component with an online analysis of the ICP signal.|
|Ian Piper, Anthony Stell, Laura Moss, Ian Piper||Evaluating clinical variation in the management data of ICP and CPP in brain-injured patients||Best practice dictates that if clinical guidelines exist, patients should be treated according to published guidelines. Survey’s indicate that when asked, clinicians believe they treat according to guidelines  and yet when data gathered on treatments are analysed, guideline adherence is less than expected . The Brain Trauma foundation clinical guidelines for the acute management of traumatic brain injury (TBI) suggest the following approach to management of ICP and CPP. This research project aims to test the following hypotheses: a) Treatment processes for ICP and CPP management in TBI can be expressed by a work-flow data structure, comprised of “primitive” objects (a simple point value and time stamp) and “complex” objects (many values with interacting sub-structures). b) The treatment processes that are extracted are clinically meaningful and accurately reflect clinical treatment in a neurological ICU environment and a tool can be developed that can extract and compare treatment processes against clinical guidelines. This project will use the BrainIT database as a test-set for assessing this methodology for automatic extraction of treatment processes and comparison to clinical guidelines. The figures below summarise this approach.|