Mount Sinai Taps Federated AI to Predict Postoperative Kidney Injury
Mount Sinai researchers are building AI models to identify which surgical patients face the highest risk of postoperative acute kidney injury (AKI), a complication that can lengthen hospitalization, increase costs, and set off a cascade of downstream harm. The work is backed by a 5-year, $4 million NIH grant awarded to the Icahn School of Medicine at Mount Sinai, with the aim of moving kidney injury prevention upstream, before the first incision is made.
The project centers on a problem that remains stubbornly common in perioperative medicine. An estimated 50 million noncardiac surgeries take place each year in the U.S., and past history suggests that about 12% of those patients develop postoperative AKI. Even when kidney function recovers, the episode can disrupt care, complicate recovery, and expose patients to further clinical instability. In severe cases, the injury is not an isolated event at all; it becomes the opening move in a broader postoperative decline.
AKI after surgery develops in a physiologic environment where the kidneys are vulnerable to swings in blood pressure, heart rate, and ventilation. Blood loss, hemodynamic shifts, and surgical stress can transiently reduce renal perfusion and impair filtration, especially in patients who already have kidney or heart disease. Supportive care remains the standard response once AKI occurs, but that reactive approach does little to identify who should have received extra protection before the injury began.
One sentence captures the clinical stakes: postoperative AKI rarely stays confined to the kidneys.
Dr. Ira Hofer, an anesthesiologist and clinical informaticist at Mount Sinai and the study’s principal investigator, said AKI prolongs hospital stays and is linked with other serious complications, including heart attack, stroke, delirium, and death. That risk profile explains why perioperative teams are increasingly focused on prevention rather than rescue. A kidney injury that begins as a laboratory abnormality can quickly reshape fluid management, medication choices, postoperative monitoring, and discharge timing.
The economic burden is also substantial. One published study found that the risk-adjusted average cost of care for surgical patients with any AKI was nearly $16,000 higher than for those without AKI. That kind of difference makes the case for prediction models that can separate patients who need intensive surveillance from those who may do well with standard care. In practice, the value of risk stratification lies not only in avoiding AKI, but in steering scarce resources toward the patients most likely to benefit.
Federated Learning Across 15 Hospitals
The modeling effort will draw on data from more than 320,000 patients treated across 15 hospitals, using federated learning rather than centralizing individual patient records in one location. Under this approach, institutions collaboratively train AI models while patient-level information stays within each health system. The method matters because perioperative data are highly sensitive, and large-scale collaboration often stalls when hospitals cannot easily share records.
This structure is especially suited to AKI prediction. Electronic health records already contain much of the information needed to estimate risk: baseline kidney function, cardiovascular comorbidity, medication exposure, procedure type, fluid balance, and perioperative vital signs. The challenge is not data scarcity so much as data fragmentation, with relevant signals distributed across institutions, systems, and time points. A federated framework allows those signals to be used together without asking hospitals to hand over raw records.
Mount Sinai is partnering with Vanderbilt University and the University of California, Los Angeles on the project, bringing together expertise in anesthesiology, nephrology, biomedical informatics, and health data science. That mix reflects the practical realities of perioperative kidney injury. An accurate model has to account for surgical physiology, renal vulnerability, and the operational details of care delivery, not just static baseline demographics.
The planned analysis goes beyond whether the model can predict AKI more accurately. Investigators also want to learn which preventive interventions work best for different patient groups. Patients identified as high risk may receive closer monitoring of kidney function, more careful management of blood pressure, and tighter control of fluid balance during and after surgery. Lower-risk patients may continue with standard monitoring, avoiding unnecessary intervention and preserving resources for the patients most likely to deteriorate.
Personalizing Prevention Before the Operating Room
The clinical logic is straightforward: if risk can be identified early enough, prevention can be individualized instead of broadly applied. AKI does not arise from a single mechanism in most surgical patients. It often reflects the convergence of vascular instability, reduced renal reserve, inflammatory stress, and perioperative exposure patterns that differ from one patient to the next. A model that can detect those differences in advance may help clinicians tailor hemodynamic targets, volume strategy, and postoperative surveillance before the kidney injury becomes established.
Hofer said the goal is to move beyond one-size-fits-all prevention and spot risk earlier so personalized strategies can be implemented. The team already has access to the data needed to create the models, and once the models are working, clinical trials will begin to test them in practice. That sequencing is important. Prediction alone does not improve outcomes unless it changes decisions at the bedside, and the next phase will have to show that the model’s output leads to better care rather than just better risk labeling.
There is also a broader systems implication here. Hospitals routinely invest in intensive postoperative monitoring for some patients while others receive standard observation, but the boundary between those groups is often imprecise. AI-driven risk prediction could make that allocation more exact, especially in a complication where early physiologic intervention may preserve organ function and prevent escalation. If the model proves reliable across settings, it could become a tool for anesthesiology, surgery, and nephrology teams trying to intervene before renal injury hardens into a more serious postoperative event.
Even so, the decisive test will come in clinical use. A model that performs well in development still has to prove that it can guide care consistently across hospitals, procedures, and patient populations. For now, the Mount Sinai grant marks a large-scale attempt to turn perioperative data into earlier action against one of surgery’s most common and costly complications.