image: Meng Li, associate professor of statistics at Rice University.
Credit: Rice University.
When patients undergo heart surgery, their kidneys face significant stress. Acute kidney injury (AKI) is one of the most common complications after cardiac procedures, often resulting in longer hospital stays, higher costs and an increased risk of death. Now, a new collaboration between Rice University and Baylor College of Medicine (BCM) aims to change that, using artificial intelligence to alert clinicians to early signs of kidney trouble giving them critical time to intervene before lasting damage occurs.
Funded by a nearly $2.5 million grant from the National Institutes of Health, the project pairs Rice’s strengths in statistics and machine learning with BCM’s clinical expertise and rich, real-world data from thousands of cardiac surgery patients. Rice will lead the development of machine learning and statistical modeling, and the AI component will be co-led by the data science team at BCM over four years.
“Early prediction would enable targeted interventions that can improve outcomes, but prior risk tools are static and have limited value in the dynamic post-operative environment,” said Meng Li, associate professor of statistics at Rice and the site principal investigator for the project. Li is also an affiliated faculty member of the recently launched Digital Health Institute and a member of the Ken Kennedy Institute.
AKI after heart surgery affects about 1 in 5 patients and can increase mortality fivefold and hospital costs threefold, the researchers said. The challenge is that clinicians typically diagnose AKI using drops in urine output or rises in serum creatinine — signals that often appear after the best window for treatment has passed.
“Presently, AKI is identified via clinical parameters, but these represent late findings often manifesting after the ideal treatment window,” said Dr. Ravi Ghanta, the principal investigator for the project and cardiothoracic surgeon at BCM.
Catching AKI earlier could prompt adjustments to fluids, blood pressure–supporting medications and avoidance of kidney-stressing drugs — simple steps that can prevent or lessen the injury.
Every heart-surgery patient generates a trove of data — vital signs, lab results, medication doses and fluid intake and output — updated minute by minute in their electronic medical record (EMR). To make sense of this complex information, the Rice–Baylor team will train ensemble machine-learning models using a uniquely comprehensive BCM dataset of more than 9,000 patients and roughly 68 million data points. The goal is threefold: to detect AKI earlier, potentially up to 24 hours before conventional signs appear; to recommend personalized, data-driven interventions that can reduce an individual patient’s risk; and to validate the system prospectively inside cardiovascular ICUs, measuring both its accuracy and how closely clinicians’ actions align with the tool’s recommendations.
“Our central hypothesis is that dynamic machine-learning models can accurately predict AKI in real time from routinely collected EMR data and augment clinical decision-making by quantifying risk reduction of therapeutic interventions,” Li said.
A major barrier to clinical AI is trust. This project emphasizes interpretability, showing which factors drive each prediction and how much specific actions might lower risk. The researchers will combine robust feature engineering with symbolic regression to create human-readable “digital biomarkers” and a simple bedside scoring system clinicians can understand and critique.
“Interpretability is key for high-stakes decisions,” Ghanta said. “It’s necessary for physician acceptance and adoption in clinical practice.”
“This collaboration is exemplary of what we might call ‘AI4Medicine,’ where Rice and the Texas Medical Center are uniquely positioned,” Li added. “We are working with complex, real-world clinical data in a high-stakes setting, and that, in turn, gives rise to what I call ‘Medicine2AI’: a powerful dynamic where clinical challenges render existing data science tools inadequate and inspire new advances in statistics and machine learning.”
Li emphasized that the data they work with is far from clean, textbook examples — it’s real-world, incomplete data, full of missing values, hidden patterns and dauntingly high dimensionality.
“It takes a full team from both BCM and Rice to preprocess and interpret them,” Li said. “But this complexity is a gift. It is seeding the development of new statistical machine learning methods and theory, from high-dimensional time series and tabular data to interpretable machine learning, uncertainty quantification, and of course, predictive modeling that is often the starting point of a responsible AI pipeline. I’m especially excited that these innovations are driven by concrete, impactful clinical questions and will be broadly applicable across scientific and biomedical applications.”
Many medical AI tools look promising in the lab but fail at the bedside. This project, however, is engineered for real-world deployment. A secure clinical deployment environment will stream electronic medical record data every 15 minutes, allowing the models to generate rolling risk profiles and recommend potential actions in real time. The team will also track “concordance” — how often clinicians independently choose the same actions the model would suggest — and then link those results to actual rates of acute kidney injury, providing a clear measure of the tool’s real-world impact.
“The electronic medical record provides complex, multidimensional data that clinicians incorporate into decision-making, but which remains underutilized in clinical decision support,” Ghanta said. “We aim to change that.”
This project is also a fertile ground for growing the next generation of AI talent.
“Our teams are deeply committed to working with and training motivated prospective researchers, including statistical PhD students, postdoctoral fellows and clinical research fellows, who can become fluent in both the language of data science and the language of biomedicine,” Li said. “Trainees often enter with a native discipline, but as a team, we advocate a bilingual environment. Not everyone needs to master both, but exposure to both areas has significantly shaped individual growth. We actively support and celebrate the value of being interdisciplinary. This educational dimension is just as critical as the algorithms we develop.”
By the end of the four-year grant, the team expects to deliver a system capable of earlier and more accurate detection of AKI in real time, along with actionable, personalized recommendations proven to lower patient risk. They will also complete prospective, real-world validation of a machine learning–enabled clinical decision support tool and create a generalizable blueprint for translating trustworthy AI into bedside care — applicable well beyond heart surgery or kidney injury.