Herein, we developed a scoring system to predict ARC onset in critically ill adults in Japan (the JPNARC score) and validated its superior predictive performance. This new prediction score, comprising five risk factors—age, sex, serum creatinine, and diagnosis at ICU admission (trauma and CNS disease)—outperformed the existing scores in terms of availability and predictive accuracy. Upon comparing the predictive performance using optimal cutoff values for each score, we found that the JPNARC score was more sensitive than the other scores. This difference may be attributed to variations in the clinical settings and detailed thresholding of the score factors. The populations used to develop the ARC and ARCTIC scores consisted exclusively of patients with sepsis and/or trauma, with median ages of 42 and 48 years, respectively [12,13,14]. Although a previous study that externally validated the ARCTIC score included both medical and surgical patients, the median patient age was 58 years [15]. Conversely, our cohort consisted of mixed ICU patients with a median age of 71 years, and the frequency of ARC remained substantial even in patients aged > 65 years (21% [87/409]). Additionally, the large sample size of this study allowed the establishment of more detailed age thresholds. Age is a key risk factor for ARC, and these differences likely contribute to the higher sensitivity of the JPNARC score. Given that the therapeutic concentrations of drugs excreted by the kidneys are reduced in patients with ARC, sensitivity, rather than specificity, is more critical in predicting ARC in a critically ill population requiring high therapeutic drug levels. In this regard, the JPNARC score, while potentially overestimating the risk in a population with high predictive probability, offers the advantage of earlier detection of at-risk patients when compared with other scores.

The JPNARC score factors were more similar to the ARCTIC scores than the ARC scores (Table 3). The ARC score uniquely incorporated sepsis as a factor; however, even in our validation set, where sepsis was more frequent, its predictive performance was inadequate. Although several previous studies have reported sepsis as a risk factor for ARC, it was not selected as a critical factor in the meta-analysis [10]. Similarly, neither sepsis nor SOFA score contributed significantly to ARC development in our patient population. Severe sepsis is associated with a high incidence of renal dysfunction [5]. In this older mixed ICU population, sepsis was not significantly associated with the onset of ARC. In contrast, the ARCTIC score showed sufficient predictive ability in our cohort, indicating a robust association between ARC and younger age, male sex, and trauma, even in critically ill Japanese patients. However, the serum creatinine threshold for the JPNARC score was unexpectedly higher than anticipated. Racial differences in creatinine-based renal function estimation equations, particularly between Asian and non-Asian populations, have been reported [16, 17], primarily due to differences in muscle mass. We expected the serum creatinine threshold in the JPNARC score to be lower than that in the ARCTIC score, which was developed for a Western population; however, the opposite was found to be true. This discrepancy may also be attributed to the older baseline age of the study population. Further evaluation is needed to explore this hypothesis using baseline characteristics consistent with those of previous ARC studies.

The JPNARC score offers superior predictive performance and advantages in terms of labor and cost. Although ARC has been recognized as a phenomenon associated with drug treatment resistance in critical care settings, CrCl measurements in patients without renal impairment are rarely utilized. Nonetheless, creatinine- and cystatin C-based renal function estimation equations are not recommended because they tend to underestimate the measured CrCl levels in patients with ARC [18, 30, 31]. Similarly in the present study, creatinine-based estimated CrCl underestimated measured CrCl in patients with ARC and was not useful. Thus, using a simple JPNARC score would reduce the labor and costs associated with unnecessary CrCl measurements. As ARC is strongly linked to suboptimal pharmacokinetic outcomes [1,2,3,4], our model could facilitate the early detection of ARC and improve patient prognosis through timely intervention.

This study has several limitations. First, the results were not prospectively validated. Although we obtained a statistically adequate number of critically ill patients in the ARC prediction study, the high heterogeneity of the target population warrants a cautious interpretation of the findings. Second, as this was a single-center study, validation may have favored the JPNARC score. However, our study population was considerably larger than those used to develop existing scores, and external validation was performed. In addition, the validation set had a higher rate of sepsis, but the results were similar when these patients were excluded (data not shown). Further validation in larger populations is required to eliminate potential group bias. Third, the cohort study did not include TDM data on therapeutic drug monitoring. The goal of ARC studies is to prevent a reduction in therapeutic drug concentrations. However, it may be necessary to consider that many commonly used beta-lactams have not been monitored. Fourth, the study did not account for the time course of critical illness. The JPNARC score uses data at ICU admission to predict subsequent ARC onset; however, the time course of ARC varies across patients. Critically ill patients often experience notable muscle mass loss [32], which can reduce serum creatinine release over time. Therefore, caution is needed regarding the timing of the score evaluation. Accordingly, in-hospital emergency patients with prolonged histories of hospitalization were excluded from the analysis. Finally, treatment-related events were excluded from the analysis. Some patients in our study experienced a recurrence after an initial ARC pause, which could have affected the pharmacokinetics of subsequent renally excreted drugs. While drug doses are typically adjusted upward in patients with ARC, excessive drug exposure can worsen neurological outcomes [33]. Additionally, ARC duration varies widely among patients [18]. As it is difficult to predict the end of ARC, clinical decisions should be based on the patient’s day-to-day condition. Therefore, careful consideration of each patient’s treatment intensity and tolerability, particularly during the maintenance phase of drug therapy, is crucial.