We are very grateful and excited to see the results of our months-long collaboration with the University of Cincinnati published in the peer-reviewed Journal of Surgical Research.
Why We Undertook This Initiative
Implementing an AI tool in the highly specialized process of caring for patients with traumatic injuries involves multiple steps and presents complex technical, clinical, and operational challenges. A key foundational part of this process is developing and validating AI models using local data, which includes patient demographics, clinical practices, and typical injury types. This aspect was well articulated in a recent paper in the NEJM AI Journal (https://ai.nejm.org/doi/full/10.1056/AIp2400223).
What We Accomplished
Through this collaboration, we developed two de novo models to predict mortality in polytrauma patients. The first model was used for external validation of the variables from the previously published Parkland Trauma Index of Mortality.
The second model was developed and validated through a complete model development lifecycle. We considered a much wider set of clinical variables that are collected or documented in the care of polytrauma patients (approximately 3500) and used the process of clinical review and statistical feature selection to get to the final set of variables for the model.
This served two purposes. First, it allowed us to evaluate if there was a significant variation in mortality predictors from one center to another without being biased by prior work (there wasn’t – in fact, there was a significant overlap). Second, it enabled us to develop a model that addressed a key workflow limitation of prior work – its inability to make predictions within the first 12 hours of patient admission.
Enhancing the Current State
One of the more important questions that needs to be asked during the implementation process of any AI product is – “as compared to what?”, that is, what is the current best practice or baseline. Given that base excess and lactate are known measures of clinical and physiological instability and are widely used by providers as such in clinical decision making, we made a reasonable assumption to use them as a method to measure provider ability to predict mortality. When comparing baseline predictors (base excess and lactate) with our machine learning model at similar levels of recall (sensitivity), our machine learning model had significantly better precision (Positive Predictive Value), meaning it produced fewer false positives. In a clinical setting, this improvement translates to a more efficient allocation of resources and potentially better patient outcomes.
Our Next Steps
We will continue making progress towards the next steps: clinical implementation, prospective evaluation, monitoring, and improvement! In parallel, our team is also collaborating with other Level 1 trauma centers to expand this work and validate this work on additional data sets.
We have always believed that collaboration and transparency are key to making AI more useful and effective in healthcare. We are grateful that Parkland, PCCI, Dr. Adam Starr, Dr. Caroline Park, Dr. Michael Cripps have previously shared and published their work in this domain. And now, we are fortunate to have collaborated with Dr. Ellen Becker, Dr. Adam Price, Dr. Claude Sagi, and Dr. Michael Goodman at the University of Cincinnati, who share the same spirit of collaboration and transparency.