Early identification means the chance to intensify monitoring before crisis forces the issue.
In the wake of a pandemic that exposed the fragility of human bodies and healthcare systems alike, researchers at Mount Sinai Health System turned the flood of clinical data into a form of foresight. Using machine learning trained on the first twelve hours of a patient's hospitalization, they built a model capable of predicting which COVID-19 patients would face kidney failure or death — not to replace clinical judgment, but to sharpen it. The work reflects a broader turn in medicine: the belief that patterns hidden in the ordinary data of illness can, if read early enough, change what happens next.
- During COVID-19's first surge, nephrologists at Mount Sinai faced a wave of kidney failure cases that overwhelmed their capacity to anticipate and respond in time.
- The stakes were stark — patients requiring dialysis faced sharply elevated risks of death, and clinicians had little early warning to act or prepare families for what was coming.
- A research team tested five machine learning algorithms against admission data from the first twelve hours, seeking the model that could most reliably predict dialysis need and critical illness across multiple points in a patient's hospital stay.
- XGBoost, driven by markers of kidney stress and metabolic strain, outperformed all rivals and was already deployed at Mount Sinai to flag the highest-risk incoming patients.
- The model still requires external validation before it can travel beyond Mount Sinai's walls, but its early deployment signals a shift toward prediction-driven care in crisis medicine.
When COVID-19 struck New York, nephrologists at Mount Sinai found themselves inside a crisis nested within the larger one: hospitalized patients were losing kidney function at alarming rates. Acute kidney injury had emerged as a serious and common complication of severe COVID-19, and when it progressed to dialysis, it often signaled that death was near. Clinicians understood that earlier identification of these patients could change everything — the monitoring, the treatment, and the conversations with families that no one wanted to have too late.
A Mount Sinai research team set out to build a machine learning model using only data available in the first twelve hours of admission — age, comorbidities, blood work, vital signs. They tested five algorithms, including logistic regression, random forest, and two versions of XGBoost, asking each to predict dialysis need and critical illness at one, three, five, and seven days into hospitalization.
XGBoost without imputation proved the strongest, outperforming the others across every time window and holding up when tested on patients the model had never encountered. Three variables drove its predictions most powerfully: red cell distribution width, creatinine, and blood urea nitrogen — all markers of kidney stress and metabolic strain. The model was already running at Mount Sinai, quietly flagging the patients most at risk.
The clinicians behind the work each named what it meant to them. One saw it as medicine finally making use of the data it had always collected. Another, who had lived through the surge of kidney failures firsthand, valued the chance to act and to speak honestly with families before crisis forced the conversation. A third pointed to machine learning's core gift: finding signal in complexity, fast enough to matter.
External validation across other hospital systems remained the necessary next step before broader deployment. But the model's existence pointed toward something larger — that in a disease which attacked the kidneys with particular force, the ability to see danger arriving within hours of admission could quietly reshape how medicine meets its most vulnerable patients.
During the first surge of the COVID-19 pandemic, nephrologists at Mount Sinai Health System in New York found themselves managing an unexpected crisis within the crisis: a flood of hospitalized patients whose kidneys were failing. Acute kidney injury, or AKI, emerged as a common and serious complication of severe COVID-19, and when it progressed, patients needed dialysis—a sign of critical illness and a predictor of death. The doctors knew that if they could identify which patients would develop this complication early, they could adjust their monitoring, change their treatment approach, and have more honest conversations with families about what might come next.
A team of researchers at Mount Sinai decided to build a machine learning model that could make that prediction. They started with data from the first twelve hours after a patient arrived at the hospital—basic information that clinicians already collected: age, existing health conditions, blood work, vital signs. They tested five different algorithmic approaches, each one a different way of teaching a computer to recognize patterns in that early data that might signal trouble ahead. The algorithms were logistic regression, LASSO, random forest, and two versions of XGBoost, one with a technique called imputation and one without. Then they asked each model to predict, on day one, day three, day five, and day seven of hospitalization, which patients would eventually need dialysis or face critical illness leading to death.
The XGBoost model without imputation won. It outperformed the others across every time window they tested, both when validated against the data it had learned from and when tested on new patients it had never seen before. The researchers found that three measurements in particular drove the model's predictions: red cell distribution width, creatinine, and blood urea nitrogen—all markers of kidney function and overall metabolic stress. The model was already being used at Mount Sinai Hospital to flag incoming COVID-19 patients at highest risk.
Dr. Girish Nadkarni, who directs the clinical intelligence center at Mount Sinai's medical school, framed the work as a natural consequence of modern medicine. Electronic health records had created vast repositories of data, he explained, and that data could be transformed into tools that directly improved patient care. Dr. Lili Chan, a nephrologist who had lived through the overwhelming surge of kidney failure cases, saw the practical value: early identification meant the chance to intensify monitoring and to have difficult conversations with patients and families before crisis forced the issue. Dr. Akhil Vaid, the postdoctoral fellow who helped develop the model, emphasized that machine learning's strength lay in its ability to find hidden patterns in complexity—to take thousands of data points and extract signal from noise, then do it fast enough to matter in real time.
The model was not yet ready for deployment across other hospital systems. It needed external validation, testing on patients from different hospitals and different populations to prove it worked beyond Mount Sinai's walls. But the principle was clear: in a pandemic that killed more than 103 million people worldwide, and in a disease that attacked the kidneys with particular ferocity, the ability to predict who would need dialysis or face death within hours of admission could reshape how clinicians cared for the sickest patients. It was a small tool, built from data and mathematics, but it pointed toward a future where prediction and prevention might arrive before crisis.
Citações Notáveis
Electronic health records have created vast data repositories that can be transformed into tools that directly improve patient care.— Dr. Girish Nadkarni, Associate Professor of Medicine (Nephrology) at Mount Sinai
Early identification of at-risk patients enables nephrologists to intensify monitoring and have informed conversations with patients and families about treatment goals.— Dr. Lili Chan, Assistant Professor of Medicine (Nephrology) at Mount Sinai
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter so much to identify these patients on day one, when you could just wait and see what happens?
Because by day one, you're already in the window where intervention changes outcomes. If a patient is going to need dialysis, you want to prepare—get the nephrologist involved, have the conversation about what that means, adjust your monitoring. You're not waiting for the crisis; you're meeting it halfway.
But the model only predicts risk. It doesn't actually prevent kidney failure, does it?
No, it doesn't. But it changes what the doctor does next. A patient flagged as high-risk gets closer watching, different medication choices, earlier discussions about goals of care. That's not nothing. And for families, it's the difference between shock and preparation.
Why XGBoost specifically? What makes it better than the other four approaches?
It's fast, it's accurate, and it's interpretable—you can actually understand why it made a prediction, not just that it did. In a hospital, that matters. A doctor won't trust a black box. They need to see the reasoning.
The model was already deployed at Mount Sinai while the study was still being published. Doesn't that skip some steps?
It does, but it also reflects the reality of the pandemic. You can't wait for perfect external validation when patients are dying now. They deployed it, they're watching it, and they're being honest that it needs testing elsewhere before it spreads. That's the responsible version of moving fast.
What happens to the model if a new variant emerges with different patterns of kidney injury?
That's the real question. The model learned from 2020 data. If the disease changes, the patterns change, and the model becomes less reliable. It's not a permanent solution—it's a tool for this moment, and it will need updating as the virus does.