When hospitalized patients’ health unexpectedly takes a turn for the worse, they are transferred to the intensive care unit (ICU), where they are more likely to die or remain for a considerable time. But our research could provide life-saving information to doctors before the patients’ health deteriorates. This is thanks to a mathematical model commonly used in Artificial Intelligence (AI) whose use could bring down the mortality rate in ICUs by 20%.
What if hospital doctors had a reliable way to identify the patients whose health was most likely to take a turn for the worse, and then proactively send those patients to the ICU? With almost 6 million patients admitted annually to ICUs in the United States, the question is anything but anodyne. Our research is based on nearly 300,000 hospitalizations in the Kaiser Permanente Northern California. Kaiser is recognized as one of America’s top hospitals in treatment for illnesses like leukemia and heart attacks.
Its data indicated that by proactively transferring patients to ICUs, hospitals reduce the mortality risk and length of stays. But there is a risk of going too far. Indeed, other research indicates that if doctors transfer too many patients to these units they may become congested and the survival rate becomes negatively impacted. Should the ICUs be filled to capacity, this could mean that some patients who need ICU care are not able to obtain it.
Our research suggests that for a proactive ICU transfer policy to work, three policies should be instigated: arrival rates must be recalibrated; the number of nurses in the ICU should be reviewed; and decisions about the transfer of patients must be gauged according to their recovery rate. If these metrics are not aligned, doctors might not make the right transfer decisions.
Creation of a Simulation Model for Hospitals
One of our key collaborators for this research, Gabriel Escobar, served as the regional director for hospital operations research at Kaiser Permanente Northern California. Kaiser provided us with unprecedented and anonymized hospitalization data on patients from 21 Kaiser Permanente facilities. Thanks to this information, we built a simulation model which mimics how an actual hospital works. This includes generating arrival and departure rates, the evolution of the patients’ condition, and every interaction they have with the system. With such micro-modeling, we can track the simulated patient as if (s)he were a real hospitalized patient. This enabled us to test different scenarios of arrivals and transfer policies.
To build our simulation model we used the mathematical framework called Markov Decision Process (MDP), a common practice in AI. This is a model for sequential decisions over time, allowing users to inspect a sequence of decisions, and analyze how one choice influences the next one. The sequence is influenced only by earlier decisions, not by future ones. We thus designed an optimization method, based upon a machine learning model, to estimate the impact of various transfer policies.
When we ran the model, we discovered that relatively small adjustments can have an impact on the mortality of the overall patient population. Given a certain way of transferring patients, we saw the estimated mortality rate could fall by 20 %!
AI Won’t Replace Human Decision-making in Hospitals
The question remains: should humans still be involved in ICU transfers, or should we rely solely on algorithms to do it? We believe these two methods could be complementary. Humans must have the final word. But their decisions could usefully be assisted by the recommendations from an algorithm. Our research seeks to encourage the implementation of simple transfer decision rules based on common health metrics summarizing the health conditions of the patients and certain thresholds. This type of threshold policy is extremely simple to deploy and readily interpretable. Using micro-modeling to understand a complicated enterprise and develop algorithms to assist in decision making can – and should - lead to better outcomes.
