A 24-Hour Warning Before Kidney Function Deteriorates: Multicenter AI Traces the Sources of Patient Risk

Acute kidney injury often develops quietly during hospitalization. By the time creatinine levels have risen significantly, the window for preventing harm may already have narrowed. A multicenter study sought to provide earlier warning: not only determining whether a patient might develop the condition within 24 hours, but also breaking down the alert into specific sources of risk so healthcare professionals know what to examine next.

The research team analyzed a total of 140,637 hospital admissions across four hospitals of different levels and in different regions of China, with an overall acute kidney injury incidence of 4.8%. The system was built on the Qwen2.5-7B architecture, with two specialized models performing separate tasks: AKI-PM reads electronic medical record data updated every six hours and predicts risk over the next 24 hours; AKI-RAM distinguishes between non-modifiable and actionable factors, such as age, pre-existing diseases, fluid depletion, or newly developed infections, and then generates corresponding recommendations.

In an internal validation involving 17,074 admissions at the same hospital, the prediction model achieved an area under the curve (AUC) of 0.95 and a positive predictive value of 0.68. When applied directly to three external hospitals, its AUC ranged from 0.88 to 0.91. After the researchers performed few-shot adaptation using 1,000 randomly selected records from each hospital, the AUC increased to 0.92–0.96 and the positive predictive value reached 0.69–0.74. This indicates that the model still needs to absorb some site-specific patient composition and data patterns to achieve optimal performance.

Compared with traditional prediction tools that merely list variable importance, the risk attribution model attempts to translate alerts into clinical language. Six nephrologists with at least five years of experience reviewed 200 cases from internal and external data. Across eight dimensions—including case comprehension, guideline consistency, clinical reasoning, acceptability of recommendations, and risk of harm—the model received mean scores of 4.18–4.88 on a five-point scale; inter-rater agreement was moderate to good.

However, physicians finding the explanations reasonable does not mean that these recommendations have already been shown to improve patient outcomes. The study relied entirely on retrospective validation using existing medical records and has not yet tested in routine clinical workflows whether the alerts change management, reduce the incidence of acute kidney injury, or create new alert fatigue. The model’s performance was also less stable among intensive care unit patients, and it currently can describe only the temporal associations of actionable factors, without numerically quantifying each factor’s contribution to an individual’s risk.

Therefore, the system is currently closer to a promising clinical decision-support prototype than a diagnostic and treatment tool capable of making judgments independently. The next hurdle is prospective clinical validation, preferably using a randomized design, along with assessments of healthcare systems across countries, differences in data quality, privacy governance, and ongoing monitoring requirements. Only when the interventions prompted by early warnings are shown to be genuinely beneficial will strong discrimination metrics translate into patient benefit.