In this study, we proposed a BIA + PEW model for PEW diagnosis, which is suitable for Chinese maintenance dialysis patients. The model shows good discrimination and calibration in both internal and external validation, and has higher diagnostic accuracy than some existing diagnostic models. We find that BIA indicators can be used as good predictors of PEW, and the combination of BIA indexes (BCM, water ratio, VFA, phase angle) and single nutritional indicator from ISRNM (that is, cholesterol) has a high predictive value for PEW. These objective parameters included in the model are based on regular laboratory results, consequently cost-effective and easy to carry out.

Diagnosis of PEW is a challenging theme. Because there has been no single diagnostic marker or tool to perfectly determine whether a patient is PEW or not, clinical studies focusing on PEW inevitably require diagnostic definition of PEW by combining one or more of the nutrition-related surrogates to allocate patients into a binary variable pertaining to PEW. According to ISRNM, PEW diagnostic standard includes biochemical indicators, BMI, muscle mass, and diet. Optimally, each criterion should be documented on at least three occasions, preferably 2–4 weeks apart [1]. This diagnostic standard includes longitudinal data, such as changes in body weight and muscle mass over a period of time, which may require dynamic and multiple observations, causing inconvenience to the diagnosis of PEW. Thus, the practical application of the strict diagnostic standard in clinical practice is somewhat limited. Furthermore, the threshold for each of the four items is still controversial [2,3,4], and some indicators don’t fully reflect nutritional status. For example, a decrease in albumin may be a result of worsening liver function, while a decrease in muscle mass may be attributed to a natural process of aging [1]. Kovesdy et al. summarized the drawbacks of ISRNM critera [28]. In fact, each nutritional method should be adjusted depending on racial, ethnic and social backgrounds. However, there’s still a lack of PEW diagnostic standard targeted for large population of Chinese dialysis patients.

Several nutrition-related tests have been proposed to assess nutritional status. The 3-point scaled Subjective Global Assessment (SGA-3) [27] scores patients as A (well nourished), B (moderately malnourished) or C (severely malnourished) (Table supplementary 1). Although this test was validated in dialysis patients [9, 29], its semi-quantitative character and the fact that it does not adequately detect the degree of malnutrition [9] led to modifications like the 7-point scaled SGA (SGA-7) [9, 29] and the Malnutrition Inflammation Score (MIS) [30,31,32]. Other clinical nutritional scores or parameters that have been related to mortality in dialysis patients include the geriatric nutritional risk index (GNRI) [33,34,35,36], dialysis malnutrition score (DMS), and composite score of protein-energy nutritional status (cPENS) [37, 38]. It is currently unknown which test should be used to assess PEW most adequately [39]. In addition to above nutritional assessment means, Moreau-Gaudry et al. mentioned a “PEW score” tool, including 4 indicators of cross section (Table supplementary 4). The model has been proved to be able to predict the survival of dialysis patients [17]. As’habi et al. assessed PEW score with a high sensitivity of 100% but a low specificity of 28.6%, which may overestimate the risk of PEW [40]. Yamada et al. proposed modified PEW score, which was modified from the original simple PEW score by adjusting the cutoff values of those parameters suitable for Japanese patients receiving hemodialysis [24] (Table supplementary 5). Ruperto et al. proposed a model combining 3 nutrition-related indexes (serum albumin, percentage of mid-arm muscle circumference, standard body weight) to predict PEW risk, with a high AUC of 0.86 [26] (Table supplementary 6). However, the above tools solely use readily available clinical and biological values at bedside, without considering other components like appetite, dietary intake and physical examination.

In recent years, electrical bioimpedance has become the most useful, simple, and reproducible method for the study of body composition. According to the principle of Omron’s law, when the current passes through human tissues, it generates resistance and reactance. The resistance is related to the hydration state, while reactance is related to the capacitance. The composition of human body components can be derived by using the impedance value of current conduction in different tissues [41]. In a multi-frequency BIA machine, current frequency of 5 ~ 1000 kHz can be selected. At very low frequencies, virtually no conduction occurs because of high cell membrane capacitance, thus allowing for the quantification of ECW. At very high frequencies, total conduction through the cell membrane occurs, thus allowing for the quantification of TBW [42]. BIA is a practical method mainly used nowadays to assess dry weight, and it has been proven to be as accurate as the reference methods considered as the gold standard [43]. In this study, we find that there’s a certain correlation between BIA indexes and PEW. Water ratio is an independent risk factor, while BCM, VFA and phase angle are independent protective factors of PEW. Zhou et al. mentioned that increased volume load was an independent risk factor for PEW [44]. Dekker et al. also found that the higher the volume load was, the worse the nutritional status was, which is partially consistent with the results of this study [45]. Rymarz et al. found that the BCM level of hemodialysis patients was positively correlated with creatinine and handgrip strength, which are indicators of muscle mass, and negatively correlated with interleukin 6. By monitoring changes of BCM, the composition of muscle tissue can be observed at an early stage [12]. Valente et al. found that BCM was an independent factor for PEW, which excludes ECW, avoiding a possible masking of the nutritional status [46]. Ruperto et al. confirmed that PA < 4 ° is an independent risk factor for PEW [26], which is similar to our results. Bansal et al. demonstrated that phase angle was significantly associated with mortality in patients with CKD and hemodialysis [47]. By evaluating and observing the changes of the above indicators, it is helpful to identify PEW at early stages and take measures to reduce the incidence of PEW. Also, we find that compared with a single indicator from ISRNM to diagnose PEW, the combination of BIA indexes and single ISRNM indicator has a better predictive ability for PEW. This observation is acceptable because each marker provides only partial information on nutritional status. The combination of multiple surrogates enables us to assess nutritional status in a multifaceted way and offers a better prediction than a single surrogate. Currently, models have been developed for screening and diagnosing PEW in dialysis patients by using BIA. Wieskotten et al. proposed a decision tree model, which divided participants into adequate nutritional status, nutrition monitoring needed and insufficient nutritional status based on BIA measurement results [25] (Figure supplementary 1). Arias-Guillen et al. confirmed that the decision tree for nutritional status assessment was a practical tool for classifying patients, and by using this method, ‘insufficient nutritional status’ was an independent diagnostic factor of PEW. Combined with other nutritional assessment methods, this decision flowchart can provide additional value for selecting patients who need to focus on nutritional intervention in clinical practice [48].

Similarly, our BIA + PEW model also combined BIA indexes and ISRNM nutritional indicators. The model has an area under the curve of 0.843 and shows good discrimination and calibration in both internal and external validation. In the diagnostic test evaluation, we divided participants from the internal validation set into negative and positive groups using different PEW diagnostic methods. Compared to previous models, our BIA + PEW model has the highest C-index and NRI. This can be explained as follows. For SGA, its semi-quantitative character leads to difference of results from different observers. Only 27.5% of the patients with PEW were identified by SGA in our research, indicating the unreliability of SGA results. The results of PEW score and modified PEW score model are presented as 4 levels (severe waste, moderate waste, slight waste, normal nutritional status), the exact diagnostic bivariate thresholds for PEW of which have not been established. Therefore, we selected the optimal cutoff value as 2 by the principle of “Jordan Index maximum”, the score lower than which was diagnosed as PEW. In addition, the performance of 3-index model is slightly inferior to our model though it was established using the same logistic regression method as our BIA + PEW model. The decision tree model shows high specificity (88.3%) but low sensitivity (26.8%) in our research. These can be explained that these models originated from France, Japan, Spain and Germany, respectively, and there are slight differences in indicators from different races and populations, resulting in poor recognition of PEW in Chinese dialysis patients.

The present study has as main strengths the total number of patients studied, adequate internal and external validation. But some weaknesses and limitations of this study should be considered. On the one hand, although BIA was also shown to be a valid method for assessing body fluids in persons with varying hydration status in some study [49], most experts believe that it’s still not valid in subjects with altered hydration status [50,51,52,53]. Ho et al. evaluated the accuracy of BIA against multiple dilution (gold standard to detect TBW) to measure TBW in individuals pre- and post-dialysis, which showed no statistically difference between them in terms of TBW average and reasonably better agreement between the two methods at post-dialysis moments than at pre-dialysis moments [54]. So in our study, BIA index measurement time point is limited to 15 min after the end of dialysis, which to the greatest extent limits the imprecision caused by the unstable volume load, though it does not rule out the measurement error caused by insufficient or excessive dialysis completely. Thus, bioelectrical impedance vector analysis (BIVA), proposed by Piccoli et al. in 1994 [53], which is reported to be an alternative method that has been validated and used for hydration status and body composition assessment in different populations, may help to further expand the validity of this study. Chamney model, proposed by Chamney et al. has been used in some BIA devices, which can distinguish muscle mass from the fluid overload and differentiate excess fluid from normally hydrated tissue, thus providing meaningful estimates of nutrition assessment for dialysis patients [22]. Moreover, our data only includes baseline levels of nutritional markers instead of repeated measures. Furthermore, as an observational study from single center, it is difficult to account unmeasured or residual confounding factors, which can lead to bias. However, though cross-sectional nature of the study may limit accuracy partially, the proposed diagnostic method can diagnose PEW quickly, conveniently, and economically, which fits the purpose of our research well and may have great value in clinical application. What’s more, though the model shows high AUC of 0.829 in external validation, the sample size of external validation is too small, resulting in wide confidence interval and calibration with slightly higher deviation. It is unclear whether our BIA + PEW model is a good predictor of clinical outcomes such as death, cardiovascular disease events, bone fracture, or hospitalization. Therefore, further studies of larger samples are necessary to determine the usefulness and validity of the model developed in our study.

In conclusion, it is hard to assess PEW in maintenance dialysis patients in daily clinical practice. Based on the recommendations of ISRNM, we suggest a new combination of parameters, which are readily available and strongly associated with other nutritional parameters. A single index of ISRNM combined with BIA indexes can also well diagnose PEW and evaluate its risk when it is impossible to obtain all the PEW diagnostic criteria.