An interpretable machine learning model for early sepsis prediction in patients with heart failure and type 2 diabetes: a retrospective cohort study

In the intensive care unit, where the stakes are highest and time is the most precious resource, doctors constantly weigh the delicate balance between a patient's chronic conditions and new, acute threats. Two of the most common and demanding chronic conditions are heart failure, where the heart struggles to pump blood effectively, and type 2 diabetes, a condition where the body cannot properly manage sugar levels. When these two conditions exist together, they create a perfect storm of vulnerability. The body's ability to fight infection is already compromised by the strain on the heart and the metabolic chaos of diabetes, making these patients far more likely to develop sepsis. Sepsis is a life-threatening reaction to an infection that can rapidly shut down organs and lead to death. Because the window to treat sepsis is narrow, identifying it early is critical, yet current tools often fail to spot the warning signs in this specific group of patients before it is too late.

A team of researchers set out to build a new kind of safety net for these patients, one that could listen to the subtle signals hidden within a patient's medical data. They turned to machine learning, a branch of artificial intelligence where computers learn to recognize patterns by studying vast amounts of past examples, rather than following a rigid set of rules written by humans. The goal was to create a system that could look at the routine data collected when a patient with heart failure and diabetes arrives in the intensive care unit and predict, with reasonable accuracy, whether they are about to develop sepsis. Unlike many complex computer models that act as "black boxes," giving an answer without explaining why, this team wanted a system that could also show its work, pointing out exactly which factors tipped the scales toward danger.

The researchers began by gathering data from two massive, publicly available databases containing records from thousands of intensive care patients in the United States. They focused specifically on adults who had been admitted with both heart failure and type 2 diabetes. From a pool of over 9,000 potential candidates, they narrowed their focus to 7,806 patients from one database to build their model, and then tested it on a separate group of 633 patients from a different set of hospitals to see if it would hold up in the real world. They looked at a wide array of information recorded during the first 24 hours of a patient's stay, including vital signs like heart rate and temperature, blood test results, and whether the patient was on a breathing machine.

To find the most important clues, the team did not just guess which factors mattered. They used a rigorous process to sift through the data, comparing different statistical methods to see which variables consistently pointed toward sepsis. This process whittled down a long list of possibilities to just seven key indicators. These included whether the patient had acute kidney injury, if they were on a mechanical ventilator to help them breathe, their white blood cell count, their body temperature, their level of consciousness measured by a standard scale, a measure of acid balance in their blood, and the level of lactate, a substance that builds up when tissues are not getting enough oxygen.

The team then trained seven different computer algorithms to learn from these seven indicators. They tested each one to see which could best distinguish between patients who would develop sepsis and those who would not. One algorithm, known as XGBoost, emerged as the most reliable. In the initial testing phase, it correctly identified the risk in about 71 percent of cases. When the researchers took this same model and applied it to the separate group of patients from the other hospitals, it maintained a respectable performance, correctly identifying risk in about 66 percent of cases. While this is not perfect, it represents a meaningful step forward for a group of patients who previously had no specialized tool to help doctors spot the danger early.

What makes this work particularly valuable is that the researchers did not stop at the prediction. They used a technique called SHAP to open the "black box" and see exactly how the model was making its decisions. They found that the computer was not guessing randomly; it was prioritizing the same things a seasoned doctor would look for, but with a mathematical precision. The model determined that acute kidney injury, a lower level of consciousness, and high white blood cell counts were the strongest signals of impending sepsis. It also learned that body temperature mattered in a specific way: both a high fever and a dangerously low temperature were warning signs, while a normal temperature was reassuring. Similarly, it recognized that high levels of lactate and the need for a breathing machine significantly increased the risk.

The researchers were careful to note that while their model performed well, it is not yet ready to replace a doctor's judgment or to make treatment decisions on its own. The external testing showed that while the model is useful, it is not infallible, and it needs to be tested further in different hospitals and with different patient populations before it can be used in everyday clinical practice. The study serves as a proof of concept, demonstrating that it is possible to build an interpretable tool that uses routine, everyday medical data to flag high-risk patients with heart failure and diabetes. By highlighting the specific combination of kidney failure, breathing difficulties, and blood chemistry changes that signal trouble, this work offers a new way to help clinicians stay one step ahead of a condition that moves too fast to ignore.