Thank you so much, Dr. Stevens. I'd like to introduce the broad area of AI, machine learning, and critical care just briefly before we get into the specifics. Machine learning and artificial intelligence, as many of you know, have really provided an opportunity to not only profoundly improve the ways in which we perform day-to-day work, but also in how it impacts outcomes for patients. Machine learning methods would span the gamut of different mathematical constructs that allow us to capture really highly complex interactions that occur within the data that are not always evident and apparent just by human eye, and therefore require really robust methods and tools to extract that meaning. There have been a number of techniques and methods that exist within the broad space of AI ML. And I'm going to be covering two broad classes, specifically those that are meant to be denoted as predictive algorithms, and others that have broader applications in prognostic enrichment, as Dr. Hector Wong would have said. And these are classes of unsupervised machine learning methods. And I'm going to be covering some of the work that's been done in this space over the last few years. I think the key message in all of this presentation is that the appropriate selection of the tools and the appropriate selection of the validation methods are important when assessing AI ML as it pertains to utilization and technology adoption within critical care. So that's the general introduction. And I'd like to get into how critical illness specifically relates to AI ML and the interactions there. I'm first introducing a really nice figure that had been generated by Dr. Yoon, Dr. Pinsky, and Dr. Claremont out of UPMC in their recent review of AI and critical care medicine. This image is notable for many reasons. One, there are these denotations, these lines that characterize when patients cross between healthy state and distressed state, and how the patient progression would have occurred were there no earlier intervention or earlier identification. As you can see, they correctly characterize the deteriorating progression of patients in the line that's marked in red. If there was no intervention, if there was no recognition, the patient would continue to deteriorate. However, there is an opportunity where when machine learning methods, when predictive methods are used to identify these distress signals earlier than humanly apparent, and when those earlier identification is met, there is an opportunity for earlier intervention and arresting some of the potential deterioration and bringing that patient back to the healthy state. And this is an important attribute of what predictive algorithms can do. The first class and the most popular areas in which predictive algorithms have been used in the literature have been in this area of sepsis detection, identifying patients who meet sepsis criteria. And generally, the divisions of work have been around three broad classes, one of which the most popular is the use of electronic medical record data, or data that exists already in the EMR, and using that retrospective data to characterize patterns in the clinical elements in that EMR to use as predictors, such as labs, vital signs, other demographic factors, and to longitudinally assess for when those variables may change, and using that change to then generate alerts that then allow for clinicians to intervene earlier. The second class and more specialized class of sepsis predictive models have been those that are broadly described as host response biomarkers. These have largely been transcriptomic or proteomic based methods that use point of care tools to identify changes in either the expression of certain genes or cytokines, and use them for earlier identification of deterioration and possibly sepsis. There has been further work that's been done in this space to delineate further whether this is bacterial, viral, or otherwise. And that's work that's been led by Dr. Tim Sweeney and others within the informatics space. There's also been emerging interest in the use of pathogen-based biomarkers and to identify specific pathogen, whether by using robust methods like metagenomics and high throughput sequencing. And these pathogen-based biomarkers have been able to further identify not only when there is general class of deterioration, but also what specimen or what microbial colony is contributing to that, and therefore able to allow for more specificity about how to intervene. And this has been an emerging area that continues to grow over the last few years and continues to grow in the next. The most significant development in this space has been the publication of the first large scale clinical deployment of a machine learning algorithm within a real time scenario and the association of the deployment of this machine learning algorithm to see how that algorithm altered the course of clinical development and whether there was any benefit in the use of that. And so this paper that was led by Roy Adams and colleagues demonstrated that earlier predictions of sepsis resulted in improvement in the types of therapies that were performed and improvements in outcomes. And underlying the algorithm utilized a number of clinical markers from the EMR data set and identified where certain deterioration events were occurring and marked those and allowed for earlier predictions to occur just based on not only a single variable, but also multiple variables that were changing, allowing for time to intervene and possibly prevent that further arrest. The other class of broad work that's been done in this space has been in the broad area of acute kidney injury. And there's been a number of work that's been done. And I wanted to highlight some work that was led by Bagliano and colleagues, which were recently published in Clinical Kidney Journal. And here they looked at a critical review of machine learning algorithms that were applied to acute kidney injury. And you can see here that there's been a number of models that have been contributed In fact, they went through and they screened for over 3,000 papers that met criteria for the use of machine learning algorithms and selected 46 relevant papers that had all the necessary criteria and showed that while there was a wide array of algorithms and machine learning methods that were used to develop the predictive models in the space of AKI, the underlying problem has been that many of these methods have relied on the use of internal validation, which means the use of data that was derived, that used for derivation and validation remain the same data set. And that many of these models, in fact, relied on public data. And there were very few that validated these models on external data sets that were not publicly available. And this caused a number of biases to creep into machine learning algorithms, one of which notably is the sort of preprocessing and filtering and harmonizations that go into standard public data sets that may not always translate to real world data, such as in local hospital systems. So when you do take an algorithm that's been developed in the public domain and implement it in a localized environment, there are a number of barriers, and some of which are more to do with how the data was prepared and how the data was organized. That just does not facilitate that translation. They also point out that while there's been a steady growth in the contribution of machine learning algorithms, they've been largely dominated by some of the classic suspects, including random forests, gradient-boosted machines, support vector machines, recurrent neural networks, and so on. So it's been largely dominated by traditional learning algorithms, whereas there's been some deep learning algorithms which are a little bit more difficult to interpret that have been contributed in this space. One specific example of a model that does utilize deep learning algorithms has been the one that was recently published by Alfieri and colleagues, titled The Deep Learning Model for Continuously Predicting Severe AKI. And this model utilized just urine output as an input and found that they could achieve a robust AUC of 0.89 compared to others when utilizing this deep learning algorithms. And these also allowed for these investigators to stratify by the severity based on how the alterations were changing. But this is one component which I wanted to highlight that there may, while some of these results may look fairly convincing, the underlying data may be very difficult and unreliable. Urine output, as many of you know, may not be captured appropriately and therefore may not be a reliable signal in all cases for determining whether it can be a reliable predictor for a machine learning algorithm. So these are considerations that we see frequently in the machine learning literature where some of the underlying variables are just maybe not a good fit for the general use case. Finally, I wanted to, or next, I wanted to discuss some of the algorithms that have been contributed in the predictions of acute respiratory failure. And this one I wanted to highlight was one that was a review of the ARF literature by Dr. Wong and colleagues. And this suggests that there's been a number of work that's been done across the general deterioration space, but very limited use or implementation of these machine learning algorithms within that broader ARF, ARDS literature. And there's an opportunity to use some of those overlapping constructs here to develop more robust and clinically useful algorithms in the space. So Dr. Wong and colleagues have generated some of those predictions, the developing electronic medical record, and compared it to traditional clinical results. That includes the MUSE algorithms and shown that when machine learning algorithms are used on this complex set of data, there can be robust predictive performance that can be achieved. And this had been demonstrated both temporally and also in external hospital sets, but yet to be deployed prospectively. So this would be interesting to see how that performs. In further work led by Jabber and colleagues, they demonstrated that not only using the EMR data can be useful, but also utilizing some of the complex data sets that include imaging can be useful to identify some of the predictive features that may change over time. And in this case, what they've done is they've actually taken the x-ray images and using deep learning algorithms showed that the deep learning algorithm is able to correctly identify regions of opacities and others for different disease states that can be useful for predicting which patient is going to go into acute respiratory failure. Now, the second class, so the first class, as I mentioned, was mostly about predictive algorithms. The second class has really been driven by clustering and phenotyping and the derivation of these phenotypes. And the derivation of these phenotypes have been very robust, and they've used a number of different methods. The paper that's shown here that has been developed by Castella-Forte and colleagues show that cohort definitions and development of phenotypes can actually occur using multiple methods, some of which are traditional statistical methods like hierarchical clustering. Others are more centroid-based methods like k-means. Others are more density-based methods like deep learning methods and so on. And then the combination of some of these statistical and machine learning methods like dynamic time warping and such to learn some of that construct in data allow for a very robust characterization of phenotypes and may be useful for prognostic enrichment. Some of the most famous work in this space has been contributed by Dr. Kalfi and colleagues, and that is related to the subphenotyping of ARDS and the demonstration that ARDS may actually consist of two subphenotypes, the hyperinflammatory and the hypoinflammatory group, and there's differences in mortality, differences in ventilator-free days and such. And they had used latent class analysis as a clustering algorithm to derive these phenotypes. Others have also derived phenotypes, however, using non-structured data. They've used biological data such as gene expression, RNA sequencing data and shown that there is actually further enrichment and differences that are notable in terms of mortality, ventilator-free days and such, and specifically looking at the altered biological pathways that are contributing to this. And so they've used a different data set but arrived at fairly similar conclusions, which again shows a robustness of the methodology. And finally, kind of looking at the clustering methods, not in a static sense but also in a dynamical sense, and this is an area that's been emerging is the use of temporal or trajectory models to identify how trajectories change over time and whether certain patients fit certain trajectories. And this is work led by Dr. Bhavani and colleagues and suggests that, in fact, in the case of septic shock, there are possibly four phenotypes or temporal trajectories that exist and that distinguish those who go on to get septic shock versus those that do not and have been differentiated, delineated by different classes such as hypothermic slow resolvers, hyperthermic fast resolvers, normothermic and then hypothermic. So in conclusion, I'd like to just highlight that there's been an array of machine learning algorithms that have been contributed in the field. These contributions have largely been focused on the use of public data, the machine learning algorithms and the methods that they've covered have focused on two broad areas such as predictive and phenotyping tasks. However, there are challenges, particularly in how these models are validated and the continual use of public datasets to generate these models and the lack of external validation remains a major concern and remains a concern that limits some of the applicability and the adoption of these models in the real world. Thank you so much for your time and hand the pen.

Machine learning

Artificial intelligence

Critical care

Predictive algorithms

Personalized care