Sepsis, a life-threatening condition, claims approximately 270,000 lives annually across the United States. This unpredictable medical emergency can escalate quickly, causing dangerous drops in blood pressure, extensive tissue damage, multiple organ failure, and ultimately death.
While timely medical interventions can save lives, certain sepsis treatments may inadvertently worsen a patient's condition, making the selection of optimal therapy a challenging decision for healthcare providers. For example, excessive intravenous fluid administration during early severe sepsis can significantly increase mortality risk.
To assist clinicians in avoiding potentially harmful treatments, researchers from MIT and collaborating institutions have engineered an innovative artificial intelligence system designed to identify therapies that carry higher risks than alternatives. Their sophisticated machine learning model can also alert physicians when septic patients are approaching a critical "medical dead end"—the point where survival becomes unlikely regardless of treatment—enabling earlier intervention.
When tested on intensive care unit data from sepsis patients, the researchers' AI system revealed that approximately 12% of treatments administered to non-surviving patients were actively harmful. The study also uncovered that about 3% of deceased patients had entered a medical dead end up to 48 hours prior to death.
"Our AI detection system provides warnings nearly eight hours before clinicians typically recognize patient deterioration. This advance notification is incredibly valuable in these critical situations where every minute matters. Understanding how a patient's condition is evolving and the risks associated with specific treatments at any given moment is crucial," explains Taylor Killian, a graduate student in the Healthy ML group at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL).
Killian's collaborators include his advisor, Assistant Professor Marzyeh Ghassemi, who leads the Healthy ML group; lead author Mehdi Fatemi, a senior researcher at Microsoft Research; and Jayakumar Subramanian, a senior research scientist at Adobe India. Their groundbreaking research was presented at the recent Conference on Neural Information Processing Systems.
Data Scarcity Challenge
This research initiative was inspired by Fatemi's 2019 paper examining reinforcement learning applications in environments where exploring arbitrary actions could be dangerous—making it difficult to generate sufficient data for effective algorithm training. These scenarios, where additional data cannot be actively collected, are known as "offline" settings.
In traditional reinforcement learning, algorithms learn through trial and error to take actions that maximize rewards. However, in healthcare contexts, generating adequate data for these models to learn optimal treatments is nearly impossible, as experimenting with various treatment strategies raises ethical concerns.
The researchers addressed this challenge by reinventing the reinforcement learning approach. They utilized limited ICU data to train a model focused on identifying treatments to avoid, with the primary goal of preventing patients from reaching medical dead ends.
"Learning what to avoid represents a more statistically efficient methodology requiring less data," Killian explains. "When considering dead ends in driving, we might think of them as road endings, but actually, every foot along that road toward the dead end could be classified as part of that dead end. Similarly, once a patient's treatment path leads to a point where any decision results in progression toward death, they've entered a medical dead end."
Fatemi adds, "One central concept involves reducing the probability of selecting each treatment in proportion to its likelihood of forcing the patient into a medical dead end—a property we call treatment security. This presents a challenging problem since the data doesn't directly provide such insights. Our theoretical breakthrough allowed us to reframe this concept as a reinforcement learning problem."
To develop their approach, called Dead-end Discovery (DeD), the team created two neural networks working in tandem. The first network focuses exclusively on negative outcomes (patient deaths), while the second concentrates solely on positive outcomes (patient survivals). This dual-network approach enabled researchers to identify risky treatments with one network and verify them with the other.
Each neural network receives patient health statistics and proposed treatments, then outputs treatment value estimates and evaluates the probability of a patient entering a medical dead end. The researchers compare these estimates against predetermined thresholds to identify warning situations.
A yellow flag indicates a patient entering an area of concern, while a red flag signals a situation where recovery is highly unlikely.
Treatment Impact Analysis
The researchers validated their model using data from sepsis patients at Beth Israel Deaconess Medical Center's intensive care unit. This dataset contained approximately 19,300 patient admissions, with observations from a 72-hour period centered around the first manifestation of sepsis symptoms. Their analysis confirmed that some patients in the dataset did indeed encounter medical dead ends.
The study revealed that 20-40% of non-surviving patients triggered at least one yellow flag before death, with many raising flags up to 48 hours prior. When comparing treatment trends between survivors and non-survivors, researchers noted a significant divergence in treatment value following the first flag, indicating a critical decision-making window.
"These findings confirm that treatment choices significantly impact outcomes. We observed that after the first flag, treatment approaches diverged markedly between those who survived and those who didn't. Our analysis suggests that over 11% of suboptimal treatments could potentially have been avoided, as superior alternatives were available to physicians at those moments. This represents a substantial improvement opportunity when considering the global volume of sepsis patients in hospitals at any given time," Killian notes.
Ghassemi emphasizes that their AI system is designed to support, not replace, human clinicians. "Medical decisions should remain with human healthcare providers. Our system offers guidance on treatments to avoid without changing this fundamental principle. We can identify risks and establish appropriate safeguards based on outcomes from 19,000 patient treatments—equivalent to a single clinician seeing more than 50 septic patient outcomes daily for an entire year," she explains.
Looking ahead, the research team aims to establish causal relationships between treatment decisions and patient health progression. They plan to enhance their model to provide uncertainty estimates around treatment values, supporting more informed clinical decisions. Additional validation through testing with data from other hospitals represents another important future direction.
This research received partial support from Microsoft Research, a Canadian Institute for Advanced Research Azrieli Global Scholar Chair, a Canada Research Council Chair, and a Natural Sciences and Engineering Research Council of Canada Discovery Grant.
