AI Prediction of Acute Kidney Injury

Acute kidney injury is a sudden drop in kidney function. Clinicians define it with the KDIGO criteria: serum creatinine rises by at least 0.3 mg/dL within 48 hours, or rises to 1.5 times baseline within seven days, or urine output stays below 0.5 mL/kg per hour for six hours. The catch is that creatinine is a lagging marker. By the time it rises, the injury may be a day or more old. Prediction models try to get ahead of that lag. They learn from many past admissions which patterns came before AKI: small creatinine drifts, rising blood urea nitrogen, low blood pressure, sepsis, recent surgery, chronic kidney disease, and exposure to drugs that can harm the kidneys, such as vancomycin, aminoglycosides, NSAIDs or IV contrast. The best-known study came from DeepMind and the US Department of Veterans Affairs and was published in Nature in 2019. A recurrent neural network trained on VA records predicted a little over half of inpatient AKI episodes up to 48 hours ahead. It produced roughly two false alerts for every true one. It also performed worse for women, who made up only a small share of VA patients. That shows how the training population shapes who a model serves well. A common misconception is that DeepMind's Streams app at the Royal Free Hospital in London was AI. It wasn't. Streams sent clinicians' phones alerts from the NHS England AKI algorithm, which is a rule-based comparison of creatinine values. Streams became better known for a 2017 finding by the UK Information Commissioner's Office that the hospital had not complied with data protection law when it shared patient records with DeepMind. The bigger lesson is that prediction is not the same as benefit. Randomized trials at Yale, led by F. Perry Wilson, tested electronic alerts that fired once AKI was present. The alerts did not reduce outcomes such as dialysis or death overall. An alert helps only if someone acts on it and an effective action exists for that patient.