Thank you. So next slide. So disclosures, none other than federal funding from the NIH, DOD, and AHRQ. So I'd like to introduce the topic of sepsis prediction, utilizing machine learning, artificial intelligence methods. Artificial intelligence and machine learning have been widely used across a number of domain which also includes critical care. And much of the critical care literature, particularly with respect to AIML, has been actually associated with sepsis prediction, just because of the importance that it exhibits. One of the elements of that has been a significant reliance on data that's been captured through electronic medical records, and much of this has been retrospectively collected. There are, however, a number of emerging alternatives to EMR-based models, and some of these include bedside point-of-care tools to rapidly assay by a specimen and generate some risk profile of each of these patients. Broadly, some of the general classes that sepsis detection or prediction models fall into have been across these three broad areas. The first one, I call them the EMR sniffers. They're utilizing data that's been entered as part of clinical workflow to identify trends in that data. These may be subtle changes in vital signs, or subtle changes in labs, or an interaction across multiple variables. And utilizing retrospective data, these models often will learn that when there are certain thresholds or certain interactions between that data, it's associated with increased risk for developing sepsis. And therefore, this has been a major tool for developing machine learning models. There are, however, some caveats to that, and we'll discuss them in the next few slides. The second area in which there's been a number of models that's been contributed has been the use of these host response biomarkers. Oftentimes these are either proteomic or transcriptomic markers that are assayed at the bedside and used to generate some risk profile of sepsis. And an emerging area of interest has been the use of these pathogen-based biomarkers, such as through a high-throughput microbiome-based methods, or metagenomic-based methods to sample for traces of RNA or DNA in the blood that may be associated with microbial colonies, and how those may influence the risk of sepsis for each of these patients. I'd like to highlight a recent paper that was published in the—or perspective that was published in The Lancet, led by Caterpill and colleagues. What you're seeing over here is an illustration of how sepsis manifests possibly in the floor or in the ICU or when patients arrive to the emergency room. The risk of sepsis continuously increases. There are some thresholds that arbitrarily can be applied, and those thresholds will oftentimes be met depending on the levels in which they're set. But as the risks increase, there are very specific processes that activate. So for instance, there may be an alert that gets displayed to the EMR, such as, what would you like to do? And there are certain options that get prompted. And otherwise, there may be certain actionable clinical decisions that are prompted based on time or acuity. And these may be decisions such as, should we possibly escalate this patient to the ICU? And this workflow assumes that data becomes continuously available in the background for machine learning models to actively consume and use as part of their pipeline to generate those relative risks. And so you can see a possible trend. However, there are very specific cases in which some of these trends may not look so uniform. They may, in fact, go up and down. These are challenges with the noise and the relative data quality issues that oftentimes are faced. For instance, if a patient's weight gets accidentally inputted as 200 kilograms or whatnot, obviously all of these things make a difference in terms of the risks and the alerts that are generated. Recently, there was a landmark study that was published in Nature Medicine, where Adams and colleagues presented a machine learning algorithm for sepsis detection that was applied in a multi-site setting, prospectively, where a number of these patients were, a number of these alerts were then adjudicated clinically and determined that, for the first time in a large-scale study, that earlier interventions that were preceded by a sepsis alert using machine learning methods actually resulted in improved outcomes. And these were quantified and demonstrated tangibly within the study. The underlying mechanisms by which the algorithm generated these risks were using clinical data and vital science data from the EMR. Many of these include things like Glasgow-Como scale, platelets, Bunt-Kratni ratio, arterial pH, and so on. And essentially, these models were looking for what were the interactions that were going on between all of these variables, and how was that predictive in terms of the onset of sepsis. Other methods have utilized other sources of continuous data. The prior work that I showed you utilized data that was obtained through clinical workflows, so actively entered by nurses or physicians. But in this method, what you're seeing here is the use of continuous data that's generated by bedside monitoring or other sensor-based tools to generate risk profiles of patients. In the figure on the right, you're seeing the increasing risk profile across different modalities of these vital signs, and how that actually contributes to quantifiable risk of patients who go on to develop sepsis. In terms of these point-of-care DAMP-based methods, you're seeing here is one example that's recently been shown and developed as a product by Safrika and colleagues, which show that the use of this bedside transcriptomic-based, mRNA-based markers were able to identify the risk of bacterial and viral infection with significant AUC. So with bacterial AUC, the performance here was shown to be around 0.94, and in terms of looking at viral infections, they are able to distinguish those with 0.89 AUCs, again, significantly improved over alternatives, such as the use of procalcitonin, and in terms of viral one, there's no alternatives here. So all of these indicate that the use of these alternative tools or these markers, when combined with machine-learning-based methods, are able to accurately identify the risk profile fairly robustly. Another point-of-care tool that has been shown within the sepsis detection space is the use of monocyte distribution width. This single variable itself was shown to be fairly robust with an AUC of around 0.78 or 0.79, and again, showing significant robustness when compared to controls. And this indicates the utility of using these individual markers within the scope of machine-learning-based tools to generate robust predictions. In the next class of large machine-learning-based tools, there is an emerging interest in utilizing machine-learning tools not only for predictive purposes, but also possibly for decision support. And therefore, a class of algorithms called reinforcement learning algorithms have been used within the context of sepsis management to determine the extent of or the means by which patients should be resuscitated, the amount of fluids they should be given, when they should be given these fluids, and in what order. And therefore, these emerging studies have shown that when taken into context, the reinforcement learning algorithms can be a viable aid or decision support that not only improves mortality or reduces mortality, but also can adjunct clinical decision-making. And a similar study, which again showed that the use of these reinforcement learning algorithms, and they're candidly called artificial intelligence clinicians, have been fairly robust in following the patient's trajectory to identify when they would best benefit in terms of the sorts of interventions that are applied. And in this model, they utilize, they allow the AI clinician to recommend both fluid therapy as well as vasopressors, and show here that when the clinicians agreed or followed the decisions that were generated by the AI clinicians, mortality was actually reduced, which again is a fairly interesting finding. So that covers really the bulk of the methods that have been used in terms of sepsis prediction and the use of machine learning. It's been divided across two broad classes, predictive models to identify when a patient may turn septic, and then secondly, classes of models that try to recommend certain decisions and certain interventions and when they should be performed in order to optimize outcomes, and if that be mortality or length of stay or so on. Of course, there are challenges with all of these methods, some of which I've described as the use of good data, and then others are the translational efforts that are required to take these models from retrospective modeling to actually prospective implementation. Thank you so much for your attention.

sepsis prediction

machine learning

artificial intelligence

electronic medical records

bedside point-of-care tools