Introduction: Clinical management of acute kidney injury (AKI) in ICU patients is variable despite guidelines. There are limited studies describing real-world practice patterns. This study used AI-based clustering to identify treatment variation. Methods: This retrospective study of adult patients from MIMIC-IV database with Stage I AKI based on Kidney Disease: Improving Global Outcomes (KDIGO) guidelines. Variables extracted were fluid administration, vasopressor use, creatinine checks, and comorbidities Data pre-processing with normalization and handling of missing values. Skewness and kurtosis tests were used to assess the shape of data distribution. Patients were clustered based on treatment and clinical features using K-Means clustering with an unsupervised machine learning algorithm. Clusters were compared using chi-square and Mann-Whitney U tests to evaluate differences in ICU mortality, length of stay, and dialysis initiation. Levene’s test assessed variance. Results: Two different patient clusters were identified: Cluster 1 (C1-pts, n = 5,955) and Cluster 2 (C2-pts, n = 17,922). C1-pts had higher ICU mortality (30.3% vs 11.3%, p< 0.001), longer ICU length of stay (median 5.3 days IQR: 2.8-10.8 vs 3.1 days IQR: 1.8-6.2, p< 0.001), and dialysis initiation (22.0% VS. 7.5%, p < 0.001) compared to C2-pts. In management, C1-pts received larger crystalloid volumes (4118.5 ml vs. 1275.7 ml, p < 0.0001); more frequent Norepinephrine (88% vs. 10%); and had more mechanical ventilation interventions (87% vs. 45%) compared to C2-pts. C1-pts had higher rates of sepsis (26% vs. 13%) and liver disease (25% vs. 15%), but fewer cases of diabetes (32% vs. 38%) and hypertension (63% vs.74%). Additional analysis within the C2-pts showed heterogeneity in total crystalloid volumes (p < 0.001), vasopressor timing (p < 0.001), mechanical ventilation (p = 0.02), fluid boluses (p < 0.001), and creatinine monitoring (p < 0.001) between survivors and non-survivors. Conclusions: Our study uncovered significant differences in AKI management in ICU patients, which could reflect the severity of underlying illness. Among homogeneous groups with low-risk, treatment variation was noticed. Applying AI-based clustering helped us to understand real-world practice and optimize a personalized treatment approach.
Telfah et al. (Sun,) studied this question.