Development and Validation of a Predictive Model of Major Adverse Kidney Events within 30 Days in Patients with Sepsis-Associated Acute Kidney Injury： Retrospective Cohort Study

In the intensive care unit, the body is often fighting a war on two fronts. One enemy is sepsis, a life-threatening reaction to an infection that sends the immune system into overdrive, damaging organs and tissues. The other is acute kidney injury, a sudden failure of the kidneys to filter waste from the blood. When these two conditions strike together, the situation becomes particularly dangerous. The kidneys are among the first organs to suffer when the body is overwhelmed by sepsis, and when they fail, the risk of death or long-term disability rises sharply. For decades, doctors have relied on general guidelines to diagnose and manage these conditions, but these rules are designed for broad populations, not for the unique, complex reality of an individual patient lying in a hospital bed. The challenge has been to move from general rules to precise predictions: to know, early on, which specific patients are most likely to suffer severe kidney outcomes within a month, so that care can be tailored to save them.

A team of researchers set out to build a tool that could make this prediction with greater accuracy. They turned to a vast collection of digital medical records from thousands of patients who had been admitted to intensive care units in the United States. Specifically, they focused on a group of 10,906 adults who had developed acute kidney injury within two days of being diagnosed with sepsis. The researchers wanted to see if they could identify a pattern in the data that would signal a high risk of "major adverse kidney events" within 30 days. This outcome is a serious combination of three things: dying in the hospital, needing a machine to filter the blood (dialysis), or having kidney function that remains permanently damaged after discharge. To find the answer, the team did not rely on a single formula or a simple checklist. Instead, they used a sophisticated type of computer learning, often called machine learning, which allows a computer to find complex, hidden relationships in massive amounts of data that human eyes might miss.

The researchers fed the computer a wide array of information about each patient, collected during the first 24 hours of their ICU stay. This included basic details like age and gender, a history of other illnesses such as heart disease or diabetes, and a snapshot of their current physical state. The computer looked at vital signs like heart rate and breathing speed, as well as detailed blood test results measuring how well the liver and kidneys were working, how much oxygen was in the blood, and how the body was handling infection. They also noted whether patients had been given medications that could be hard on the kidneys or drugs to support blood pressure. The team tested eight different types of machine learning algorithms, each using a slightly different mathematical approach to learn from the data. They trained these models on a large group of patients and then tested them on a separate group to see if the models could correctly predict the outcome for people they had never seen before.

After running the tests, one model stood out as the most reliable. It was a system known as CatBoost, which proved better at distinguishing between patients who would have a good outcome and those who would face a major kidney event than any of the other seven models it was compared against. When the researchers applied this best-performing model to a completely different set of patients from another database, it maintained its accuracy, suggesting that the patterns it learned were real and not just a fluke of the first group of data. The model achieved a level of accuracy that is considered strong in medical prediction, correctly identifying the risk for the vast majority of patients. This means that for a new patient with sepsis and kidney injury, the model could look at their initial data and provide a clear estimate of their risk of suffering a major kidney event within a month.

To ensure that doctors could trust and understand the model's decisions, the researchers did not treat it as a "black box" that simply gives an answer without explanation. They used a method to break down the prediction and show exactly which factors were pushing the risk up or down for each individual. The analysis revealed that the most powerful predictor of a bad outcome was simply the severity of the kidney injury itself at the time of admission. Patients with more advanced stages of kidney damage were far more likely to have a poor outcome. Beyond the kidneys, the model highlighted the importance of the patient's overall health history, the severity of their organ failure, and their breathing rate. It also showed that exposure to certain medications that can harm the kidneys increased the risk. By showing how these factors combined to create a specific risk profile for a single person, the model moves beyond a generic statistic to offer a personalized view of the patient's situation.

The study confirms that it is possible to use data from the first day of an ICU stay to forecast serious kidney complications in patients with sepsis. The findings suggest that by combining a patient's immediate vital signs, their blood work, and their medical history, a computer model can provide a transparent and accurate assessment of their future risk. This approach does not replace the judgment of a doctor but offers a powerful tool to support it. It allows medical teams to identify the patients who are most vulnerable early on, potentially enabling them to intervene sooner or monitor more closely. While the model was tested on data from American hospitals and may need further validation in other parts of the world, the core discovery is that the signals for a poor kidney outcome are present and measurable from the very beginning of a critical illness. This work represents a step toward a future where care in the intensive care unit is guided by precise, individualized predictions rather than broad averages, helping to ensure that the most at-risk patients receive the attention they need.