A novel approach to dry weight adjustments for dialysis patients using machine learning

Hae Ri Kim; Hong Jin Bae; Jae Wan Jeon; Young Rok Ham; Ki Ryang Na; Kang Wook Lee; Yun Kyong Hyon; Dae Eun Choi

doi:10.1371/journal.pone.0250467

. 2021 Apr 23;16(4):e0250467. doi: 10.1371/journal.pone.0250467

A novel approach to dry weight adjustments for dialysis patients using machine learning

Hae Ri Kim ^1,^2,^#, Hong Jin Bae ^3,^#, Jae Wan Jeon ¹, Young Rok Ham ^2,⁴, Ki Ryang Na ^2,⁴, Kang Wook Lee ^2,⁴, Yun Kyong Hyon ^5,^‡,^*, Dae Eun Choi ^2,^4,^‡,^*

Editor: Bhagwan Dass⁶

PMCID: PMC8064601 PMID: 33891656

Abstract

Background and aims

Knowledge of the proper dry weight plays a critical role in the efficiency of dialysis and the survival of hemodialysis patients. Recently, bioimpedance spectroscopy(BIS) has been widely used for set dry weight in hemodialysis patients. However, BIS is often misrepresented in clinical healthy weight. In this study, we tried to predict the clinically proper dry weight (DW_CP) using machine learning for patient’s clinical information including BIS. We then analyze the factors that influence the prediction of the clinical dry weight.

Methods

As a retrospective, single center study, data of 1672 hemodialysis patients were reviewed. DW_CP data were collected when the dry weight was measured using the BIS (DW_BIS). The gap between the two (Gap_DW) was calculated and then grouped and analyzed based on gaps of 1 kg and 2 kg.

Results

Based on the gap between DW_BIS and DW_CP, 972, 303, and 384 patients were placed in groups with gaps of <1 kg, ≧1kg and <2 kg, and ≧2 kg, respectively. For less than 1 kg and 2 kg of GapDW, It can be seen that the average accuracies for the two groups are 83% and 72%, respectively, in usign XGBoost machine learning. As Gap_DW increases, it is more difficult to predict the target property. As Gap_DW increase, the mean values of hemoglobin, total protein, serum albumin, creatinine, phosphorus, potassium, and the fat tissue index tended to decrease. However, the height, total body water, extracellular water (ECW), and ECW to intracellular water ratio tended to increase.

Conclusions

Machine learning made it slightly easier to predict DW_CP based on DW_BIS under limited conditions and gave better insights into predicting DW_CP. Malnutrition-related factors and ECW were important in reflecting the differences between DW_BIS and DW_CP.

Introduction

Accurately establishing dry weight is important in patients with end stage renal disease (ESRD) on hemodialysis. It is well known that identifying the proper dry weight is associated with the adequate efficiency of dialysis, minimizes the burden on the cardiovascular system, and improves the survival rate of dialysis patients [1–3].

In general, dry weight is(DW) defined as the lowest tolerated postdialysis weight achieved via gradual change in postdialysis weight at which there are minimal signs or symptoms of hypovolemia or hypervolemia [4].

It has been reported that DW overestimation is a major risk factor in the development of hypertension, left ventricular hypertrophy, and cardiovascular disease, therefore affecting the mortality risk of patients on hemodialysis [5–8]. Underestimation of DW can result in hypovolemia and can induce hypotension, cramps, and dizziness [9,10]. Hypovolemia can also result in a reduced blood supply to vital organs, causing ischemia and consequently contributing to the loss of residual renal function [8]. As a consequence, it is crucial to implement an efficient and precise method to determine DW in patients with ESRD on hemodialysis.

DW assessment via clinical signs, such as measurements of blood pressure, jugular venous pressure, and the presence or absence of edema, is the most commonly used method. However, this method does not reflect the patient’s underlying illnesses and decreases in muscle mass [11].

Over the years, paraclinical tools have been developed to help nephrologists assess DW. Measurements of the cardiothoracic ratio on chest X-rays has long been used but is imprecise due to its dependency on underlying cardiovascular conditions, such as cardiac hypertrophy [12]. Ultrasonic measurements of the inferior vena cava diameter and lung can also be useful tools for measuring DW but are limited because they only provide information concerning overhydration [13,14].

Recently, a method to determine the amount of body water using the bioimpedance spectroscopy (BIS) method has been developed [15]. Even though DW measurements using BIS are useful and reliable when applied to actual clinical patients, there have been reports that the measured DW was not optimal for some patients depending on the prescription of blood pressure medications and the presence of cardiovascular disease [7,16].

Recently, machine-learning-based artificial intelligence has been successfully applied in the medical field and has shown promising results in predicting complications [17–19]. Machine-learning approaches can be classified into two main types: unsupervised learning for unlabeled data and supervised learning for labeled data. The latter is more appropriate for predicting complications; decision tree-based machine-learning algorithms are especially appropriate for table-type data with labels, i.e., target factors [20–22]. However, there are few studies that use machine-learning approaches to predict DW in hemodialysis patients.

In this study, machine learning was applied to accurately determine DW. Using this approach, various clinical factors influencing patients with a large gap between the DW measured using BIS (DW_BIS) and the clinically stable dry weight were identified and analyzed. Additionally, we evaluated whether a more accurate clinical dry weight could be presented using machine learning based on the various clinical factors and BIS data of the patient.

Materials and methods

Study design

The present study was approved by the institutional review board of Chungnam National University Hospital (IRB number: CNUH 2015-12-025-003). This study was entirely retrospective (using existing data and medical records available as of the date of submission of the IRB), and was permitted a consent waiver through the IRB. The study was conducted as a single center and retrospective study on dialysis patients who were treated at the Chungnam National University Hospital between January 2011 and September 2015. All data were collected from the medical records of the Chungnam National University Hospital. Information on mortality was obtained from the database of the National Health Insurance Service.

Patient selection

All patients were diagnosed with ESRD and started maintenance hemodialysis between June 2011 and December 2015. Maintenance hemodialysis is defined here as hemodialysis performed three times a week.

Dialysis was maintained for at least 3 months after initiation. Patients who received temporarily hemodialysis due to acute renal injury were excluded. In addition, we excluded patients with unstable clinical conditions (i.e., acute infections or intensive care unit admission). Patients for whom we did not have information concerning their hydration status or who’s DW could not be confirmed using the BIS method were excluded. Patients whose antihypertensive drug prescriptions changed between the 2 weeks before and after the time their body compositions were measured were also excluded.

Body composition measurements

Body composition, including hydration status (overhydration; OH), was assessed using a portable BIS device (BCM; Fresenius Medical Care, Bad Homburg, Germany). Patient measurements were obtained on the day of dialysis prior to dialysis. All data measured using the BIS method, including extracellular water (ECW), intracellular water (ICW), total body water (TBW), lean tissue index (LTI), fat tissue index (FTI), ECW/ICW (E/I) ratio, body mass index (BMI), lean tissue mass (LTM), fat mass (FAT), adipose tissue mass (ATM), and body cell mass (BCM), were included. These fluid volumes were then used to determine the fluid overload, expressed as the OH value. OH was measured as a negative or positive value. ECW, ICW, TBW, and OH are all expressed in units of liters.

Definitions

In general, DW is defined as the lowest tolerated postdialysis weight achieved via gradual change in postdialysis weight at which there are minimal signs or symptoms of hypovolemia or hypervolemia [4]. In this study, the concept of clinically proper DW (DW_CP) was used to distinguish it from DW_BIS. DW_CP was defined as the lowest post-dialysis weight a patient could tolerate without signs or symptoms of overhydration or dehdyration during or after dialysis. DW was defined by combining judgments on the clinical situation with reference to DW predicted by the BIS method. The clinical judgment of the appropriate dry weight used the patient’s indications and imaging studies. At the center where the study was conducted, the medical staff checks and records events from the previous dialysis to the day on the dialysis day. For example, when determining overhydration, it was included based on cases where peripheral edema, generalized edema or chest discomfort, or pleural effusion or pulmonary edema was not observed in chest PA. We checked whether musle spasm or dizziness occurred through medical records. The OH value measured using BIS is denoted OH_BIS, and the DW assessed using BIS (DW_BIS) was calculated from the body weight (Weight just before dialysis) minus the OH_BIS value (kg). The gap between DW_CP and the body weight (Weight just before dialysis) is denoted OH_CP. The measurement gap between DW_BIS and DW_CP is the DW gap (Gap_DW), i.e., it is defined as the absolute value of the body weight minus DW_BIS from DW_CP in units of kg.

Measurements of clinical parameters

The baseline patient characteristics, its mean and standard deviation (SD) analyzed included the age on the start day of dialysis (days), sex, height (cm), initial body weight (kg), predialysis systolic/diastolic blood pressure (mmHg), and comorbidities (presence of diabetes or hypertension). Factors predicted to affect the patient’s DW were collected from the blood test results. Laboratory tests were performed within three days before and after the DW_BIS measurement, and hemoglobin, blood urea nitrogen, creatinine (Cr), albumin, total calcium (TCa), phosphorus (P), sodium (Na), potassium (K), and chloride (Cl) measurements were included.

Machine-learning adjustment

The machine-learning methodology, which is currently attracting attention, was applied to predict OH (proper). Since many external sources on machine learning exist, detailed information on the introduction and methodology of machine learning is not included. Briefly, machine learning is a very effective method of learning a given data to predict a specific aim attribute (OHcp in this paper), and using it to predict the corresponding aim attribute when new patients data are inputted. A lot of research is being conducted, and it is widely used in artificial intelligence technology. In this study, we used a decision tree-based methodology suitable for application to table-type medical data, the light gradient boosting method (LightGBM). the extreme gradient boosting tree (XGBoost), and random forest methods [20–22]. These tree-based methods are more efficient for table type data structure and relatively small data set, especially, RF is better than others for given small data set. Even though deep learning is more famous in machine learning applications, it requires bigger data set for learning model and getting certain prediction accuracy. Before learning the prediction model from the collected data, outliers were first removed by pre-processing the data. To verify the prediction result, 20 random samplings were performed. In each random sample, 80% of the collected data were sampled and used for training and the remaining 20% were used for testing. The suggested values obtained from the predicted results are presented after taking the average of the 20 samplings.

Results

Baseline characteristics

The average age of the patients was 65.1 years. There were 672 (40%) men, and diabetes and hypertension were found in 1,002 (50%) and 1,328 (79%) of the patients, respectively. The average duration of dialysis was approximately 806 days.

To confirm the predicted level of machine learning according to the gap between DW_BIS and DW_CP, the analysis was divided into groups with differences of less than 1 kg and of more than 2 kg. In patients with a Gap_DW value of more than 2 kg, the proportion of women was high, the age was high, and there were many patients with diabetes (Table 1).

Table 1. Baseline characteristics.

Group	Total N = 1,672		GAP (<1 kg) N = 972		GAP (≧1kg, <2 kg) N = 303		GAP (>2 kg) N = 384
Feature	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Age	65.01	12.23	64.91	12.44	65.56	11.58	64.84	12.30
Gender	672 (40%)	0.49	408 (42%)	0.49	119 (39%)	0.49	142 (37%)	0.48
Height (cm)	161.46	8.51	160.98	8.25	161.14	8.62	162.69	8.92
Weight (kg)	60.50	11.30	60.24	11.24	60.34	10.39	61.15	12.07
Vintage of HD (day)	806.79	826.20	786.98	809.59	837.01	808.27	827.89	873.59
DM	1,002 (50%)	0.49	561 (58%)	0.49	185 (61%)	0.49	247 (64%)	0.48
HTN	1,328 (79%)	0.40	797 (82%)	0.38	240 (79%)	0.41	279 (73%)	0.45
Hb (g/dL)	10.13	1.55	10.25	1.50	10.15	1.59	9.83	1.61
Total protein (g/dL)	6.34	0.70	6.37	0.66	6.43	0.64	6.19	0.81
Albumin (g/dL)	3.38	0.57	3.46	0.55	3.39	0.49	3.17	0.65
BUN (mg/dL)	56.05	23.65	57.72	23.21	53.94	23.29	52.95	24.07
Cr (mg/dL)	7.68	3.37	7.98	3.26	7.73	3.35	6.89	3.55
Tca (mg/dL)	8.36	0.79	8.36	0.77	8.33	0.73	8.37	0.88
P (mg/dL)	4.35	1.57	4.43	1.55	4.41	1.66	4.10	1.48
Na (mEq/L)	137.06	3.97	137.18	3.95	136.70	3.71	137.03	4.15
K (mEq/L)	4.74	0.92	4.77	0.90	4.81	0.96	4.60	0.96
Cl (mEq/L)	101.34	5.16	101.59	5.16	100.69	5.06	101.21	5.13
TBW (L)	32.00	7.11	31.62	6.90	32.08	6.81	32.85	7.79
ECW (L)	15.54	3.51	15.19	3.32	15.59	3.23	16.37	3.99
ICW (L)	16.45	4.20	16.41	4.03	16.48	4.08	16.51	4.71
E/I	0.97	0.19	0.94	0.16	0.97	0.18	1.03	0.24
BMI (kg/m²)	23.16	3.80	23.18	3.78	23.20	3.52	23.09	4.06
LTI (kg/m²)	13.05	3.42	13.07	3.24	13.18	3.32	12.92	3.92
FTI (kg/m²)	9.11	4.61	9.23	4.56	8.96	4.52	8.88	4.83
LTM (kg/m²)	34.39	10.74	34.31	10.34	34.59	10.60	34.41	11.85
FAT (kg)	18.70	9.67	18.74	9.48	18.29	9.04	18.71	10.52
ATM (kg)	21.85	10.79	22.08	10.56	21.34	10.50	21.57	11.64
BCM (kg)	18.93	7.28	18.92	6.99	19.13	7.18	18.82	8.08
DW_BIS (kg)	58.41	11.19	58.44	10.88	58.12	10.32	58.35	12.54
OH_BIS (kg)	2.11	2.40	1.77	1.91	2.17	2.07	2.94	3.32
DW_CP (kg)	58.63	10.94	58.44	10.86	58.44	10.16	59.29	11.50
Gap_DW (kg)	0.22	3.07	0.00	0.48	0.32	1.50	0.94	4.79
OH_CP (kg)	1.87	3.17	1.80	2.15	1.91	2.07	1.87	4.00

Open in a new tab

Abbreviations: SD, standard deviation; HD, Hemodialysis; DM, diabetes mellitus; HTN, hypertension; Hb, hemoglobin; BUN, blood urea nitrogen; Cr, serum creatinine; TCa, serum total calcium; P, serum phosphorus; Na, sodium; K, serum potassium; Cl, serum chloride; TBW, total body water; ECW, extracellular water; ICW, intracellular water; E/I, extracellular water to intracellular water ratio; BMI, body mass index; LTI, lean tissue index; FTI, fat tissue index; lean tissue mass, LTM; FAT, fat mass; ATM, adipose tissue mass; BCM, body cell mass; DW, dry weight; OH, overhydration.

Prediction of clinical DW using machine learning

The accuracy of the results when applying machine learning to all the collected data is shown in Table 2. As can be seen from the results, the prediction accuracy of machine learning for OH_CP was very low, less than 40%, and the maximum error for OH_CP was very high at 1.5 kg. It was difficult to obtain a valid machine-learning-based predictive model, which is required for OH_CP prediction.

Table 2. Prediction using machine learning.

		LightGBM Test Accuracy (%) (OH-CP(MAE))	XGBoost Test Accuracy (%) (OH-CP(MAE))	Random Forest Test Accuracy (%) (OH-CP(MAE))
Gap_DW < 1 kg	Mean	81.25% (547.0 g)	82.89% (515.5 g)	79.52% (570.3 g)
	Max	84.57% (499.1 g)	86.33% (459.9 g)	83.11% (530.1 g)
	Min	73.85% (622.1 g)	75.21% (596.1 g)	70.44% (656.2 g)
Gap_DW < 2 kg	Mean	69.71% (770.6 g)	72.02% (734.6 g)	71.00% (747.3 g)
	Max	78.33% (693.9 g)	80.06% (674.3 g)	79.03% (682.2 g)
	Min	53.76% (876.9 g)	55.84% (831.0 g)	54.18% (857.0 g)
Total	Mean	23.58% (1,351.9 g)	28.54% (1,287.5 g)	30.82% (1,250.5 g)
	Max	35.85% (1,224.6 g)	38.83% (1,178.8 g)	39.36% (1,139.4 g)
	Min	11.19% (1,448.7 g)	21.01% (1,404.2 g)	21.43% (1,377.9 g)

Open in a new tab

All collected data (1,672 patients) were included. CP(MAE) indicates the clinically proper mean absolute error.

We classified and extracted two groups based on Gap_DW (|DWCP−DW_BIS|) from the collected data. One group contained data where Gap_DW is less than 1 kg, and the other contained data where Gap_DW is less than 2 kg. The results of applying machine learning to these two groups are presented in Table 2. It can be seen that the average accuracies for the two groups are 83% and 72%, respectively, in the case of XGBoost; in particular, the accuracy for the Gap_DW < 1 kg patients was dramatically improved compared to the prediction result for the entire dataset. It is obvious from this result that, when data with large Gap_DW are included and as Gap_DW increases, it is increasingly difficult to predict the target property (OH_CP) as a BIS measurement property.

In terms of machine learning, the difference between the two groups can also be confirmed via the feature importance of the machine-learning methodology. Comparing the feature importance of the two groups (Group 1: Gap_DW < 1 kg and Group 2: Gap_DW < 2 kg) reveals very different patterns, as shown in Fig 1. In the feature importance of Group 1, E/I and ECW (L) play important roles in learning the machine-learning predictive model and predicting the target feature; the other features, especially TBW (L), appear to play a secondary role. Conversely, Group 2, which includes patients with Gap_DW between 1 kg and 2 kg, shows a further enhancement in the importance of E/I and ECW (L). In addition, the roles of features of low importance in Group 1 are slightly increased in Group 2 and distributed downward. Therefore, the data (1 kg ≦ Gap_DW < 2 kg) newly included in Group 2 changes the properties of the dataset and appears to play a role in lowering the accuracy of the prediction.

The factors affecting Gap_DW

To compare the clinical differences from the point of view of the data distribution, box plots were used to compare the clinical factors in groups with Gap_DW less than 1 kg, between 1 kg and 2 kg, and greater than 2 kg (Fig 2). In box plot, the mean values of hemoglobin, total protein, serum albumin, serum creatinine, phosphorus, serum sodium, serum potassium, and FTI tended to decrease, as the Gap_DW increased. The height, TBW, ECW, and E/I ratio tended to increase as the Gap_DW increased. LTI and BMI showed no change in their trends for the three groups.

Fig 2 — Distribution of the data according to GapDW: (A) clinical data; (B) blood chemistry; and (C) BIS parameters.

Discussion

In this study, we derived a more clinically accurate DW using machine learning based on volume status information measured using BIS as well as clinical information concerning the patients.

In addition, it was confirmed that patient’s blood chemistry, including hemoglobin, total protein, albumin, creatinine, phosphorus, and potassium, was a factor that caused a gap between DW_BIS and DW_CP.

It is important to maintain adequate hydration during the treatment of patients with ESRD on hemodialysis. Determining the amount and rate of water removal by targeting the appropriate patient DW during dialysis treatment is a major part of the dialysis regimen. In general, the DW evaluation is determined by a clinical decision. DW is defined as the lowest weight that is tolerable to a patient without symptoms or the occurrence of hypotension [23].

Several methods have been proposed to measure the volume status, and blood tests including atrial or brain natriuretic peptide levels, inferior vena cava diameter measurements using ultrasound, and blood pressure measurements have been used. However, DW results using these tools are not accurate [24,25]. BIS or multifrequency bioimpedance analysis was introduced to establish DW in dialysis patients [15]. Using a method to distinguish between ECW and ICW over multiple frequencies, it is possible to more accurately predict DW by measuring the volume status of a dialysis patient [26]. However, sometimes, dialysis targeting DW using the BIS method can result in a hypotensive status or peripheral and pulmonary edema in dialysis patients [16].

In this study, we tried to correct the gap between DW_BIS and DW_CP using various clinical data gathered from patients. The application of machine learning to the BIS-based data and clinical parameters resulted in accuracy compatible with that of predicting OH_CP using DW_BIS alone; however, the mean absolute error (MAE) for OH_CP was improved.

In particular, when only data with a Gap_DW value of less than 1 kg were analyzed as a dataset using machine learning, a mean accuracy of 83.26% and a maximum accuracy of 86.65% were found using the XGBoost method and the predicted difference from OH_CP was up to 458 ml. For OH_CP, the average measurement range was 731 ml and it was more predictable for DW_CP than for BIS-based data alone. When using all of the data, the method was very inaccurate when predicting OH_CP; however, because BIS-based data play an important role in the prediction, when there was too much of a difference with respect to the clinical DW, or when data that accurately reflect the volume status were used, the machine-learning prediction was thought to be inaccurate overall.

In view of the Gap_DW groups, the box plots indicated that the Gap_DW difference and older ages did not show a tendency but that the importance of age increased in the group with a Gap_DW difference of more than 2 kg. In a previous study, when comparing the hydration status of elderly dialysis patients and young dialysis patients, it was reported that ICW is low, ECW/TBW tends to be low, and a larger hydration status is shown. In the distribution of the actual data, an age trend was not observed but it was confirmed that the importance of the features analyzed by machine learning is useful for applications to actual clinical data.

Box plots were used to reveal the distributions of the BIS-based data and the clinical data according to the gap differences based on Gap_DW. In the box plots, the mean value of ECW showed a tendency to increase as Gap_DW increased. Conversely, the overall distribution of hemoglobin, serum total protein, serum albumin, serum creatinine, phosphorus, and serum potassium had lower values as Gap_DW increased.

The ECW value reflects overhydration. As Gap_DW increases, ECW tends to increase. As the degree of overhydration increases, the prediction accuracy of DW_BIS is expected to decrease.

Previous studies have shown that hypoalbuminemia (low serum albumin concentration) is associated with overhydration in patients undergoing maintenance dialysis. ECW increased in patients with low serum albumin concentrations, and similar results were shown regardless of the dialysis method. In addition, when hypoalbuminemia is present, overall nutritional weight, BMI, BUN, serum creatinine and potassium levels, and dietary protein intake are normal [27–29]. Hypophosphatemia is significantly associated with overhydration. In addition, it has been reported to be related to higher age and lower serum albumin, creatinine, hemoglobin, and serum calcium levels [30].

In previous studies, a relationship between overhydration and low hemoglobin concentration was reported and anemia tended to worsen when overhydration was severe. In addition to chronic kidney disease, more than half of patients with advanced heart failure who could show volume overload showed anemia and it was reported that this anemia was caused by a dilution effect due to fluid overload [31,32]. There are limitations to directly using anemia as a measure of overhydration; however, it can be considered as a factor affecting the increase in Gap_DW.

In several previous studies, serum creatinine was used as a biomarker for muscle metabolism when assessing muscle mass in patients with ESRD on dialysis. Early studies estimating muscle mass based on creatinine kinetics demonstrated good correlation with other estimates of muscle mass [33,34]. In addition, serum creatinine has been significantly correlated with LTM [35]. Based on these findings, lower predialysis serum creatinine levels are predicted to affect the Gap_DW increase.

Hypokalemia is used as an indicator of malnutrition, and its association with malnutrition via albumin and total protein can be estimated [36]. Hypokalemia showed a tendency to drop significantly in the group with Gap_DW of more than 2 kg. Based on these results, caution is needed when interpreting and applying DW_BIS to patients with hypokalemia and/or malnutrition.

These blood chemistry findings may not be related to malnutrition. Even though there was no tendency in LTI and BMI according to the Gap_DW increase in the box plots, these blood chemistry findings are associated with a decrease in the effective circulating volume and an increase in interstitial edema, consequently making the prediction of DW_CP difficult.

There are some limitations to this study. First, the DW measurements using the BIS device were performed by a single person; therefore, there is a possibility that an error in the measurement result may be included. Second, because the amount of data for the group with Gap_DW of more than 2 kg is small, it is not suitable for training a machine-learning predictive model for OH_CP. However, it is important to obtain a predictive model and proper factors other than the BIS factors to improve the treatment of patients in this group. Third, in this study, an analysis by gender and age was not performed. Such an analysis in the future will enable a more detailed approach for these patient groups. Fourth, it was not possible to evaluate whether the external cohort population showed the same pattern as a single-center study. Fifth, the patient’s mortality and morbidity were not analyzed in relation to water control through more accurate measurement of dry weight. In future studies, prospective studies and multicenter studies will be conducted, and the long-term prognosis of patients will need to be analyzed.

In conclusion, we found the feature importance of groups according to differences in Gap_DW. It was confirmed that ECW and malnutrition-related blood chemistry findings, including hemoglobin, serum total protein, serum albumin, serum creatinine, serum phosphorus, and serum potassium, were important features. Machine learning made better predictions of DW_CP based on DW_BIS in patients with Gap_DW of less than 2 kg. It is difficult to predict DW_CP using machine learning for patients with Gap_DW of more than 2 kg; therefore, it is necessary to analyze additional appropriate features for future extreme cases.

In the future, if more patients are included to increase the prediction accuracy using machine learning, this technique will be helpful in establishing the appropriate DW for patients. Machine-learning predictive models can be helpful to establish aOH_CP.

Data Availability

Data cannot be shared publicly because data contains potentially identifying or sensitive patient information. Data are available from the Institutional Data Access / Ethics Committee of Chungnam National University Hospital (contact via irb@cnuh.co.kr) for researchers who meet the criteria for access to confidential data.

Funding Statement

This research was supported by National Institute for Mathematical Sciences (NIMS) grant funded by the Korea government, 2021 (No. NIMS-B21910000) and Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2017R1D1AB03035061). However, the funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

1.Chang S-T, Chen C-C, Chen C-L, Cheng H-W, Chung C-M, Yang T-Y. Changes of the cardiac architectures and functions for chronic hemodialysis patients with dry weight determined by echocardiography. Blood Purif. 2004;22(4):351–359. 10.1159/000080031 [DOI] [PubMed] [Google Scholar]
2.Wizemann V, Wabel P, Chamney P, Zaluska W, Moissl U, Rode C, et al. The mortality risk of overhydration in haemodialysis patients. Nephrol Dial Transplant. 2009;24(5):1574–1579. 10.1093/ndt/gfn707 [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Charra B. Does empirical long slow dialysis result in better survival? If so, how and why? ASAIO J. 1993. [PubMed] [Google Scholar]
4.Sinha AD, Agarwal R. Opinion: can chronic volume overload be recognized and prevented in hemodialysis patients? The pitfalls of the clinical examination in assessing volume status. Semin Dial. 2009;22(5):480–482. 10.1111/j.1525-139X.2009.00641.x [DOI] [PubMed] [Google Scholar]
5.Onofriescu M, Siriopol D, Voroneanu L, Hogas S, Nistor I, Apetrii M, et al. Overhydration, cardiac function and survival in hemodialysis patients. PLoS One. 2015;10(8):e0135691. 10.1371/journal.pone.0135691 [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Vertes V, Cangiano JL, Berman LB, Gould A. Hypertension in end-stage renal disease. NEJM. 1969;280(18):978–981. 10.1056/NEJM196905012801802 [DOI] [PubMed] [Google Scholar]
7.Fishbane S, Natke E, Maesaka JK. Role of volume overload in dialysis-refractory hypertension. Am J Kidney Dis. 1996;28(2):257–261. 10.1016/s0272-6386(96)90309-1 [DOI] [PubMed] [Google Scholar]
8.Amato R, Hlebovy D, King B, Salai P. Fluid removal. Obtaining estimated dry weight during hemodialysis. ANNA Core Curriculum for Nephrology Nursing. 2008:692–703. [Google Scholar]
9.Sands JJ, Usvyat LA, Sullivan T, Segal JH, Zabetakis P, Kotanko P, et al. Intradialytic hypotension: Frequency, sources of variation and correlation with clinical outcome. Hemodial Int. 2014;18(2):415–422. 10.1111/hdi.12138 [DOI] [PubMed] [Google Scholar]
10.Daugirdas JT, Blake PG, Ing TS. Handbook of dialysis. 4th ed: Lippincott Williams & Wilkins; 2012. [Google Scholar]
11.San Miguel S. Haemodialysis dry weight assessment: A literature review. Ren Soc Aust J. 2010;6(1):19–24. [Google Scholar]
12.Poggi A, Maggiore Q. Cardiothoracic ratio as a guide to ultrafiltration therapy in dialyzed patients. Int J Artif Organs. 1980;3(6):332–337. [PubMed] [Google Scholar]
13.Vitturi N, Dugo M, Soattin M, Simoni F, Maresca L, Zagatti R, et al. Lung ultrasound during hemodialysis: the role in the assessment of volume status. Int Urol Nephrol. 2014;46(1):169–174. 10.1007/s11255-013-0500-5 [DOI] [PubMed] [Google Scholar]
14.Agarwal R, Bouldin JM, Light RP, Garg A. Inferior vena cava diameter and left atrial diameter measure volume but not dry weight. Clin J Am Soc Nephrol. 2011;6(5):1066–1072. 10.2215/CJN.09321010 [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Passauer J, Petrov H, Schleser A, Leicht J, Pucalka K. Evaluation of clinical dry weight assessment in haemodialysis patients using bioimpedance spectroscopy: a cross-sectional study. Nephrol Dial Transplant. 2010;25(2):545–551. 10.1093/ndt/gfp517 [DOI] [PubMed] [Google Scholar]
16.Chamney PW, Krämer M, Rode C, Kleinekofort W, Wizemann V. A new technique for establishing dry weight in hemodialysis patients via whole body bioimpedance. Kidney Int. 2002;61(6):2250–2258. 10.1046/j.1523-1755.2002.00377.x [DOI] [PubMed] [Google Scholar]
17.Gupta A, Liu T, Shepherd S, Paiva W. Correction: Using Statistical and Machine Learning Methods to Evaluate the Prognostic Accuracy of SIRS and qSOFA. Healthc Inform Res. 2018;24(3):250. 10.4258/hir.2018.24.3.250 [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Desautels T, Hoffman J, Barton C, Mao Q, Jay M, Calvert J, et al. Pediatric severe sepsis prediction using machine learning. bioRxiv. 2017:223289. [Google Scholar]
19.Arefan D, Mohamed AA, Berg WA, Zuley ML, Sumkin JH, Wu S. Deep learning modeling using normal mammograms for predicting breast cancer risk. Med Phys. 2020;47(1):110–118. 10.1002/mp.13886 [DOI] [PMC free article] [PubMed] [Google Scholar]
20.Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al., editors. Lightgbm: A highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst; 2017. [Google Scholar]
21.Chen T, Guestrin C, editors. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining; 2016 August; San Francisco California USA: Association for Computing Machinery, New York, NY, United States.
22.Ho TK, editor Random decision forests. Proceedings of 3rd international conference on document analysis and recognition; 1995 Aug 14–16; Montreal, Quebec, Canada, Canada: IEEE; 2002.
23.Jaeger JQ, Mehta RL. Assessment of dry weight in hemodialysis: an overview. J Am Soc Nephrol. 1999;10(2):392–403. [DOI] [PubMed] [Google Scholar]
24.Lee SW, Song JH, Kim GA, Lim HJ, Kim M-J. Plasma brain natriuretic peptide concentration on assessment of hydration status in hemodialysis patient. Am J Kidney Dis. 2003;41(6):1257–1266. 10.1016/s0272-6386(03)00358-5 [DOI] [PubMed] [Google Scholar]
25.Ando Y, Yanagiba S, Asano Y. The inferior vena cava diameter as a marker of dry weight in chronic hemodialyzed patients. Artificial Organs. 1995;19(12):1237–1242. 10.1111/j.1525-1594.1995.tb02292.x [DOI] [PubMed] [Google Scholar]
26.Cox-Reijven PL, Kooman JP, Soeters PB, van der Sande FM, Leunissen KM. Role of bioimpedance spectroscopy in assessment of body water compartments in hemodialysis patients. Am J Kidney Dis. 2001;38(4):832–838. 10.1053/ajkd.2001.27703 [DOI] [PubMed] [Google Scholar]
27.Dumler F. Hypoalbuminemia is a marker of overhydration in chronic maintenance patients on dialysis. ASAIO J. 2003;49(3):282–286. 10.1097/01.mat.0000065465.52748.bb [DOI] [PubMed] [Google Scholar]
28.Zaki DS, Mohamed RR, Mohammed NA, Abdel-Zaher RB. Assessment of malnutrition status in hemodialysis patients. Clinical Medicine and Diagnostics. 2019;1:0.20. [Google Scholar]
29.Jones CH, Akbani H, Croft DC, Worth DP. The relationship between serum albumin and hydration status in hemodialysis patients. J Ren Nutr. 2002;12(4):209–212. 10.1053/jren.2002.35295 [DOI] [PubMed] [Google Scholar]
30.Garagarza C, Valente A, Caetano C, Oliveira T, Ponce P, Silva AP. Hypophosphatemia: nutritional status, body composition, and mortality in hemodialysis patients. International urology nephrology. 2017;49(7):1243–1250. 10.1007/s11255-017-1558-2 [DOI] [PubMed] [Google Scholar]
31.Hung SC, Kuo KL, Peng CH, Wu CH, Wang YC, Tarng DC. Association of fluid retention with anemia and clinical outcomes among patients with chronic kidney disease. J Am Heart Assoc. 2015;4(1):e001480. 10.1161/JAHA.114.001480 [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Androne A-S, Katz SD, Lund L, LaManca J, Hudaihed A, Hryniewicz K, et al. Hemodilution is common in patients with advanced heart failure. Circ. 2003;107(2):226–229. 10.1161/01.cir.0000052623.16194.80 [DOI] [PubMed] [Google Scholar]
33.Patel SS, Molnar MZ, Tayek JA, Ix JH, Noori N, Benner D, et al. Serum creatinine as a marker of muscle mass in chronic kidney disease: results of a cross-sectional study and review of literature. J Cachexia Sarcopenia Muscle. 2013;4(1):19–29. 10.1007/s13539-012-0079-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Molnar MZ, Streja E, Kovesdy CP, Bunnapradist S, Sampaio MS, Jing J, et al. Associations of body mass index and weight loss with mortality in transplant‐waitlisted maintenance hemodialysis patients. Am J Transplant. 2011;11(4):725–736. 10.1111/j.1600-6143.2011.03468.x [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Vinge E, Lindergård B, Nilsson–Ehle P, Grubb A. Relationships among serum cystatin C, serum creatinine, lean tissue mass and glomerular filtration rate in healthy adults. Scand J Clin Lab Invest. 1999;59(8):587–592. 10.1080/00365519950185076 [DOI] [PubMed] [Google Scholar]
36.Nakhoul GN, Huang H, Arrigain S, Jolly SE, Schold JD, Nally JV Jr, et al. Serum potassium, end-stage renal disease and mortality in chronic kidney disease. J Am Soc Nephrol. 2015;41(6):456–463. 10.1159/000437151 [DOI] [PMC free article] [PubMed] [Google Scholar]

PLoS One. doi: 10.1371/journal.pone.0250467.r001

Decision Letter 0

Bhagwan Dass

10 Feb 2021

PONE-D-20-40081

A novel approach for adjustment of to dry weight in adjustments for dialysis patients using machine learning.

PLOS ONE

Dear Dr. Choi,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.This is an interesting study as your team has tried a newer approach of dry weight in adjustments for dialysis patients using machine learning .After reviewing this paper and also the reviewer comments several issues needs to be addressed before this study can be considered for publication.Please address the comments and concerns by the reviewers and after that your paper will be reviewed again before being considered for publication.

Please submit your revised manuscript by Mar 27 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.
A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.
An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols

We look forward to receiving your revised manuscript.

Kind regards,

Bhagwan Dass, MD

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. To comply with PLOS ONE submission guidelines, in your Methods section, please provide additional information regarding the methodological and statistical analyses performed for the computational tests. For more information on PLOS ONE's expectations for statistical reporting, please see https://journals.plos.org/plosone/s/submission-guidelines.#loc-statistical-reporting.

3. We note that you have indicated that data from this study are available upon request. PLOS only allows data to be available upon request if there are legal or ethical restrictions on sharing data publicly. For information on unacceptable data access restrictions, please see http://journals.plos.org/plosone/s/data-availability#loc-unacceptable-data-access-restrictions.

In your revised cover letter, please address the following prompts:

a) If there are ethical or legal restrictions on sharing a de-identified data set, please explain them in detail (e.g., data contain potentially identifying or sensitive patient information) and who has imposed them (e.g., an ethics committee). Please also provide contact information for a data access committee, ethics committee, or other institutional body to which data requests may be sent.

b) If there are no restrictions, please upload the minimal anonymized data set necessary to replicate your study findings as either Supporting Information files or to a stable, public repository and provide us with the relevant URLs, DOIs, or accession numbers. Please see http://www.bmj.com/content/340/bmj.c181.long for guidelines on how to de-identify and prepare clinical data for publication. For a list of acceptable repositories, please see http://journals.plos.org/plosone/s/data-availability#loc-recommended-repositories.

We will update your Data Availability statement on your behalf to reflect the information you provide.

4. Please amend either the title on the online submission form (via Edit Submission) or the title in the manuscript so that they are identical.

5.Thank you for stating the following in the Acknowledgments Section of your manuscript:

"This research was also supported by National Institute for Mathematical Sciences (NIMS) grant funded

by the Korea government, 2020 (No. NIMS-2020B900000). This research was supported by Basic

Science Research Program through the National Research Foundation of Korea (NRF) funded by the

Ministry of Education (NRF-2017R1D1AB03035061)."

We note that you have provided funding information that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.

Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:

"he funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript."

Please include your amended statements within your cover letter; we will change the online submission form on your behalf.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #3: No

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: No

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: No

Reviewer #3: No

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Choi and coworkers are submitting a novel approach for adjusting dry weight in hemodialysis patients using machine learning support. For this purpose, they developed a retrospective study using a data set of 1672 hemodialysis (HD) patients in which they compared dry weight defined clinically (clinical proper) to dry weight predicted or estimated from bioimpedance spectroscopy device (BIS). They estimated the gap between these two methods and they clustered such difference in three categories (<1; 1-2;>2 liters). With machine learning support, they identified and weighted factors (clinical data, lab test, body composition from BIS) being associated with each of these groups. They concluded that machine learning support, may help to improve dry weight clinical estimate using BIS and additional factors (clinical data, lab values, body composition) particularly in the lowest fluid overload patient categories. In addition, malnutrition related factors may explain most of gap discrepancy in dry weight estimate.

This novel approach relying on machine learning and artificial intelligence is interesting, and seems appealing as support tool for guiding physicians in managing more precisely dialysis patients. This approach may be used as an example of featuring future of medicine.

Now the study raises several concerns that should be addressed:

1. Use of such machine learning and intelligence artificial tool to identify factors that may influence bad estimate of dry weight such malnutrition is not sufficient to validate this approach.

2. It is not clear from reading, how the authors defined the ideal dry weight of patients? Was it the fluid overload estimated from the BIS device that was chosen as target to set suitable or ideal dry weight or the clinical judgement or a combination? That should be defined clearly since it is confusing along the manuscript.

3. If the ideal or suitable dry weight was established on the clinical and/or biologic assessment, therefore the authors should define which criteria and threshold values were used to define the three categories of gap discrepancies.

4. If the target dry weight was established from BIS measurement then it is easier to understand but still clinical criteria used to define fluid status needs to be listed with their threshold values.

5. From a methodological point of view, it would be also interesting to validate the algorithm that was developed in the assessment of fluid overload within an external cohort of patients.

6. Clinical outcomes of these three categories of patients would require further analyses. This is important to value the support of machine learning in reducing intradialytic morbidity (ie, incidence of intradialytic hypotension) and/or improving medium- or long-term morbidity (ie, hospitalization for pulmonary edema) and mortality (all cause or cardiovascular mortality).

7. What will be the clinical implication in the future of using this tool? How the authors envisage to develop and use such tool? Are they planning to perform a prospective interventional study with which aims?

Reviewer #2: It is a research paper that has been hard work with a large number of patients, but it requires extensive revision in order for the general reader to understand.

I have several questions.

I think you defined DWcp as usual meaning, but what is the ‘prediction of DWcp’ using machine learning?

You said OHcp is the gap between DWcp and the body weight (pre-dialysis?), but what does OHcp mean by machine learning?

You said that you derived a more clinically accurate DW using machine learning, but where is the data?

What new data is included in Group Two?

DWgap Group 2 has mixed definitions. Be sure to indicate.

More detail explanation is needed for the general readers to understand machine learning.

What does the number next to diabetes or HTN on the figure 2? Please indicate the statistical values.

I do not understand what correlation figure 2 represents. The graph showing the correlation seems to be missing.

For the title ‘The factors influencing GapDW’, these factors do not affect GapDW, but are values that depend on the GapDW group.

Maybe larger gap group tend to be in high overhydration status, so that they have large ECW and lower concentrations of materials.

Reviewer #3: Kim et al. have gathered an impressive dataset: BIS-measurements from as many as 1672 hemodialysis patients, undergoing hemodialysis at a single center. The authors stated they would “predict the clinical dry weight, by using machine learning for volume status derived from BIS and clinical information”.

Major comments:

1) I do not understand the definition of the “DW gap”. The authors state that this is “the absolute value of the body weight minus DWBIS from DWCP in units of kg”, but I find this definition unclear. For example, if an 80 kg person has a DWBIS of 77 kg and a DWCP of 78 kg. I am assuming the gap is 1 kg. What happens if an 80 kg person had a DWBIS of 78 kg and a DWCP of 77 – is the gap still 1 kg?

2) The authors state that “in general, dry weight (DW) is defined as the lowest weight a patient on chronic hemodialysis can tolerate [4].” (1) The reference is a 1980 publication by Henderson and clinically outdated. A more appropriate definition is given by Agarwal: (2) (also not the newest paper available, but much more modern than Henderon’s).

3) The authors state that “In this study, the concept of clinically proper DW (DWCP) was used to distinguish it from DWBIS. DWCP was defined as the post-dialysis weight in which the patient had a clinically stable water state (no hypotension during dialysis or edema after dialysis).” Sounds very good, but what was done to ensure this condition in 1672 hemodialysis patients?

4) The authors state: “Additionally, by correcting these influential factors, we attempted to correct the difference between the DWBIS and the clinically appropriately adjusted DW to ultimately predict the correct DW.” A confusing sentence, but even I take it seriously, may I ask how this was done in a retrospective study?

5) The authors state: “A total of 1,672 patients were included in the study.” […] “The study was conducted […] between January 2010 and September 2015.” […] “There are some limitations to this study. First, the DW measurements using the BIS device were performed by a single person [36]”

� The reference (REF 36) does not fit here. Can the authors please explain, using a patient flow chart, how exactly that one person was able to perform BIS in all 1,672 patients? Did one person measure 1-2 consecutive patients per day, thus between 300 and 600 a year? How did the authors deal with the effect of time? Did the dry weight protocol change from beginning to end of the retrospective study period?

6) The authors state that “in patients with a GapDW value of more than 2 kg, the proportion of women was high, the age was high, and there were many patients with diabetes (Table 1).” Related to my comment (1), I do not understand how the dry weight gap was defined. Was the clinical dry weight “off target” (too high, if BIS was considered the gold standard) in older women with diabetes?

7) The authors state that “There were 672 (40%) men.” This percentage is very low for a usual hemodialysis center, as the majority of dialysis patients worldwide are men, not women. Can the authors please report the country statistics for Korea, and how their center fits in within their country’s data?

8) The authors state: “It is obvious from this result that, when data with large GapDW are included and as GapDW increases, it is increasingly difficult to predict the target property (OHCP) as a BIS measurement property.” I have no clue what this sentence means.

9) The authors state: “Box plots were used to reveal the distributions of the BIS-based data and the clinical data according to the gap differences based on GapDW. In the box plots, the mean value of ECW showed a tendency to decrease as GapDW increased. Conversely, the overall distribution of hemoglobin, serum total protein, serum albumin, serum creatinine, phosphorus, and serum potassium had lower values as GapDW increased. The ECW value reflects overhydration. As GapDW increases, ECW tends to increase.”

� Can the authors please clarify what they mean? In my understanding, the two sentences that I underlined (“In the box plots, the mean value of ECW showed a tendency to decrease as GapDW increased.” versus “As GapDW increases, ECW tends to increase.”) say exactly the opposite of one another.

Overall judgement:

Many of my points raised here above are so major that they unfortunately jeopardize the entire study itself. However, I see an even more fundamental problem with this analysis: If I understand correctly, the authors are using BIS dry weight data to predict clinical dry weight, and the machine learning algorithm is a fancy way of trying to relate one thing (BIS technology) with another (clinical dry weight assessment and management). In my opinion, this undertaking is unfortunately useless, clinically. Many studies from the early years of the BCM have shown that dialysis patients are “not on target”, meaning the clinical dry weight differs from the BIS-derived dry weight. The fact that it is difficult to perform clinical dry weight management is the very reason that BIS is informative, on top of clinical judgement, in the first place. A machine learning algorithm should not be used to identify factors that predict the clinical dry weight, deviating from the BIS-derived dry weight. Instead, the authors can take the classical approach: establish a hypothesis regarding the factors that they feel are important (malnutrition, BMI, sex/gender, malnutrition, inflammation, frailty) and check whether these factors differ between patients who are “on” or “off” the BIS-derived dry weight target. Unfortunately, this analysis will not be new. But in my opinion, the novel machine learning algorithm presented here causes more confusion than it helps the clinician.

Minor:

10) The authors state: “ECW, ICW, and TBW were calculated using a fluid model [23].” The BCM result is actually obtained that way, but this sentence sounds as if the authors had done the calculation by themselves.

References

1. Henderson LW: Symptomatic hypotension during hemodialysis. Kidney Int 1980;17:571-576

2. Agarwal R, Weir MR: Dry-weight: a concept revisited in an effort to avoid medication-directed approaches for blood pressure control in hemodialysis patients. Clin J Am Soc Nephrol 2010;5:1255-1260

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

PLoS One. 2021 Apr 23;16(4):e0250467. doi: 10.1371/journal.pone.0250467.r002

Author response to Decision Letter 0

30 Mar 2021

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

ANS) The manuscripts have been modified according to PLOS one style.

ANS) We analyzed typical baseline statistics, mean and standard deviation as a characteristics of the data. It is general and clear in statistical analyses. Anyhow, we added more words for more detail in method section, especially, in its subsection, ‘Measurements of clinical parameters’.

We will update your Data Availability statement on your behalf to reflect the information you provide.

ANS) As you indicated, we contain following contents in cover letter.

Patient’s data contain potentially identifiable or sensitive patient information. The IRB has maintained ethical and legal restrictions on patient data acess. If you want to access individual patient data, you need to get approval from Chungnam National University Hospital IRB (255 Munwharo, Junggu, Daejeon, South Korea, 35015, +82422806781).

4. Please amend either the title on the online submission form (via Edit Submission) or the title in the manuscript so that they are identical.

ANS) As your recommendation, we corrected that

5.Thank you for stating the following in the Acknowledgments Section of your manuscript:

"This research was also supported by National Institute for Mathematical Sciences (NIMS) grant funded by the Korea government, 2020 (No. NIMS-2020B900000). This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (NRF-2017R1D1AB03035061)."

We note that you have provided c that is not currently declared in your Funding Statement. However, funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form.

Please remove any funding-related text from the manuscript and let us know how you would like to update your Funding Statement. Currently, your Funding Statement reads as follows:

"The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript."

Please include your amended statements within your cover letter; we will change the online submission form on your behalf.

ANS) As your recommendation, we deleted the funding information in Acknowledgments Section. We added the our amended statements within your cover letter.

Reviewers' comments:

5. Review Comments to the Author

Now the study raises several concerns that should be addressed:

1. Use of such machine learning and intelligence artificial tool to identify factors that may influence bad estimate of dry weight such malnutrition is not sufficient to validate this approach.

ANS) Traditional statistical analysis gives many important analytical results. Boxplot, which is one of Exploratory Data Analytics (EDA) is good analytics tool to get interquartile behavior of given data and a pattern in different groups. In this study, we mainly used box plot to identify important factors. However, because of linear property of statistical analysis, it is hard to get hidden nonlinear relations. For the reason, we applied a machine learning approach to find hidden features with its ‘feature importance’ functionality, and then provided its results. Combining statistical analysis and machine learning could get more information and more chance find new factors, especially, important factor for target property, for example, DWCP or OHCP in this study. Because of a little ambiguity, which is the review’s comment, we remedy the statement in the discussion section, as follows.

“ In conclusion, machine learning is useful to understand the feature importance of groups according to differences in GapDW. Machine-learning predictive models can be used to establish a clinical diagnosis support system for OHCP.”

� “ In conclusion, we found the feature importance of groups according to differences in GapDW. Machine-learning predictive models can be helpful to establish OHCP.”

ANS) In this study, the ideal dry weight (clinically proper dry weight (DWCP) in this study) was determined by applying the concepts of several studies that define the existing dry weight [references 1,2,3]. The lowest post-dialysis body weight was set during or after dialysis with no signs or symptoms of overhydration or dehdyration. The ideal dry weight was set by combining the evaluation of the clinical situation while referring to the dry weight predicted through the BIS device. For example, in the case of overhydration, the judgment included peripheral edema, generalized edema, or chest discomfort, or when pleural effusion or pulmonary edema was not observed in chest PA. Hypotension during or after dialysis, musle spasm during dialysis, dizziness It was evaluated based on the presence or absence of triggers. The following content has been added to manuscipt.

Reference

1) Henderson LW. Symptomatic Hypotension During Hemodialysis. Kidney international. 1980.

2) Jaeger JQ, Mehta RL. Assessment of Dry Weight in Hemodialysis: An Overview. J Am Soc Nephrol. 1999;10(2):392-403.

3) Charra B, Laurent G, Chazot C, Calemard E, Terrat J-C, Vanel T, et al. Clinical Assessment of Dry Weight. Nephrology Dialysis Transplantation. 1996;11(supp2):16-19.

ANS) There is neither recommended values nor RCT for range of clinical stable dry weight in hemodialysis patients. And, there is no RCT on how much volume change is allowed in hemodialysis patients from an appropriate dry body weight, but when the weight gain between dialysis session is greater than 4-4.5% of the body weight, the risk of ventricular hypertrophy, cardiovascular death, cerebral side effects, all causes mortality is higher (reference 1). For this reason, the U.S. Kidney Foundation generally requires that the weight gain between dialysis session be kept within 2 kg (reference 2). Based on these details, a dry weight error of 2 kg or more was determined as a value with a high possibility of causing side effects due to clinically obvious excessive volume change in hemodialysis patients.

For dry weight errors within 1kg, the BIS measurement error range of volume status is suggested as -1.1~1.1L (referece 3). Based on this information, we set the group of an error of less than 1 kg of dry weight as control.

1) Bossola M, Pepe G, Vulpio CJJoRN. The frustrating attempt to limit the interdialytic weight gain in patients on chronic hemodialysis: New insights into an old problem. 2018;28(5):293-301.

2) Daugirdas JT, Blake PG, Ing TS. Handbook of dialysis: Lippincott Williams & Wilkins; 2012.

3) Passauer J, Petrov H, Schleser A, Leicht J, Pucalka KJNDT. Evaluation of clinical dry weight assessment in haemodialysis patients using bioimpedance spectroscopy: a cross-sectional study. 2010;25(2):545-551.

ANS) The criteria for establishing clinical dry weight are the same as those described in the answer to question 2. In order to know the critical value of clinical dry weight, it is necessary to measure the weight which the patient's clinical factors change during dialysis are unstable. However, it is difficult to know the threshold of clinical dry weight. Because of the safety of patients, doctor do not induce the patient to be unstable clinical situation for evaluating range of dry weight. Deliberately lowering or raising dry weight until complications are triggered in order to establish precise criteria for threshold values can be a problem in terms of patient safety. For example, a patient with a weight of 68 kg started dialysis and continue to reduce the weight by 1-2 kg for a dialysis session and find the weight which patient is clinically stable. If the stable dry weight is 64 kg, we do not set the lower dry weight to know the clinical unstable point for patient safety. However, in general, clinical situation remained stable at a range from -0.5kg to 1kg in clinically stable dry weight.

5. From a methodological point of view, it would be also interesting to validate the algorithm that was developed in the assessment of fluid overload within an external cohort of patients.

ANS) This study is a retrospective, a single-center study. It has not been analyzed in an external cohort. We added this point as a limitation of this study. We consider a prospective study including a group of patients from other institutions.

ANS) In this study, we focused on prediction of clinical dry weight through machine learning. And the information of the episode of hypotension and mid- to long-term morbidity and mortality during actual dialysis are not collected. Especially we did not permit the information of death and hospitalization of patients from IRB. So, we cannot analysis the intradialytic morbidity and long term mortality. This comment is indicated as a part of the limitation of this research.

ANS) Although we did not present it in this study, it is clinically valuable to predict an accurate dry weight because volume status in dialysis patients is already closely related to the patient's prognosis in previous studies.

Several previous studies have addressed limits on dry weight on determining by clinical judgment. In order to compensate for this, there has been an opinion that predicting dry weight using the BIS method. Although BIS based dry weight is useful, but it often incorrectly predicts clinical dry weight. Therefore, it is necessary to find the better prediction tool for clinical stable dry weight.

In this study, we wanted to know whether a more accurate clinical dry weight could be presented using machine learning based on the various clinical factors and BIS data of the patient.

Prediction using machine learning will be important because the accuracy increases as the number of data is richer, so it will be important to have a large number of data not only in our institution but also in multi centers. In future prospective studies, the goal is to proceed as a multicenter, large-scale study, and the primary goal is to increase the accuracy of prediction using a large amount of data. In addition, we want to evaluate whether this machine learning based DW prediction improve cardiovascular complications or patient’s survival.

Reviewer #2: It is a research paper that has been hard work with a large number of patients, but it requires extensive revision in order for the general reader to understand.

I have several questions.

1. I think you defined DWcp as usual meaning, but what is the ‘prediction of DWcp’ using machine learning?

ANS) The main subject is DWcp, which is easily calculated to OHcp with the body weight. In developing a predictive model in machine learning, it is important to get a better error (small error) rather than accuracy. Sometimes, the good accuracy does not imply an expected good resolution in error. For a good resolution in prediction, it is more efficient to get target factor in small range and decimal point for establishing a predictive model using machine learning. OHcp is the factor in this study. So, we first developed the machine learning predictive model for OHcp rather than DWcp. Hence, we predicted OHcp and then easily calculated DWcp, which is the ‘prediction of DWcp’.

2. You said OHcp is the gap between DWcp and the body weight (pre-dialysis?), but what does OHcp mean by machine learning?

ANS) As we answered in reviewer’s comment 1 just above, we made OHcp the target factor of machine learning rather than DWcp. And then we calculated/recoverd the ‘prediction of DWcp’ with given the body weight and OHcp.

3. You said that you derived a more clinically accurate DW using machine learning, but where is the data?

ANS) Because of the relation among DWcp, OHcp, and the body weight (which is the given input factor), DWcp and OHcp are one-to-one correspondence with the given body weight. Especially, the error of OHcp in prediction is equivalent to that of DWcp in prediction. So, the better (smaller) error of OHcp using machine learning means a more clinically accurate DW.

4. What new data is included in Group Two?

ANS) Group 2 contains the data of Group 1. So, the new data in Group 2 means 1 kg ≦ GapDW < 2 kg added to group 1. This new data ranges from 1 kg to 2 kg are in Table 1.

5. DWgap Group 2 has mixed definitions. Be sure to indicate.

ANS) we correct the confusing and incorrect expression. Group 2 means <2kg.

6. More detail explanation is needed for the general readers to understand machine learning.

ANS) As your recommendation, we added the following contents in method section.

Since many external sources on machine learning exist, detailed information on the introduction and methodology of machine learning is not included. Briefly, machine learning is a very effective method of learning a given data to predict a specific aim attribute (OHcp in this paper), and using it to predict the corresponding aim attribute when new patients data are inputted. A lot of research is being conducted, and it is widely used in artificial intelligence technology.In this study, we mainly used decision tree-based learning methods (LightGBM, XGBoost, RF) in machine learning predictions. These tree-based methods are more efficient for table type data structure and relatively small data set, especially, RF is better than others for given small data set. Even though deep learning is more famous in machine learning applications, it requires bigger data set for learning model and getting certain prediction accuracy.

7. What does the number next to diabetes or HTN on the figure 2? Please indicate the statistical values.

ANS) In the box plot, the 1/0 categorical data must be displayed as shown in the figure. Therefore, it is difficult to correct the picture for this part, and instead, the percentage of patients is presented on the graph.

8. I do not understand what correlation figure 2 represents. The graph showing the correlation seems to be missing.

ANS) I removed it because it misrepresented.

9. For the title ‘The factors influencing GapDW’, these factors do not affect GapDW, but are values that depend on the GapDW group.

ANS) we changed to “ The factors affecting GapDW”

10. Maybe larger gap group tend to be in high overhydration status, so that they have large ECW and lower concentrations of materials.

Ans) Yes, we agree. Thank you for your comment.

Major comments:

ANS) As you indicated, the predicted dry weight value by the BIS method may be higher or lower than the clinical dry weight. We did not distinguish theses separately. The focus was mainly on the difference (Gap) between clinical dry weight and BIS based predicted dry weight. We aimed to reduce this gap through machine learning.

ANS) The concept of dry weight proposed by Agarwal is more concrete, but it is difficult to regard it as a completely different concept from the concept proposed by Henderson. The dry weight (DW) defined by Henderson is defined as the lowest weight a patient on chronic hemodialysis can tolerate. The dry weight as defined by Agarwal is the lowest tolerated postdialysis weight achieved via gradual change in postdialysis weight at which there are minimal signs or symptoms of hypovolemia or hypervolemia. As you commented, we modified it in manuscipt. However, since it is almost similar to the concept of dry weight (DW) defined by Henderson, it does not seem necessary to change in the clinically proper dry weight (DWCP) defined in this study.

ANS) Basically, the hemodialysis medical staffs in this hospital reviews all the symptoms, signs, and events that occurred newly after dialysis immediately before the dialysis day the patient visited. All of the vital sign measured prior to dialysis is checked for stability, and abnormal symptoms or signs are checked through physical examination and history taking such as palpation and auscultation. In addition, the blood pressure during dialysis is basically checked every hour, but when the dry weight is reduced or increased by measuring body composition (for 1-2 weeks), blood pressure is measured at 30-minute intervals during dialysis, and blood pressure fluctuations are large or blood pressure. If there is a deterioration, it is systemized to notify the doctor in charge immediately. In addition, after changing the dry weight, the chest x-ray was checked to check for pulmonary edema or pleural effusion. It was not taken with an exact protocol in relation to the dry weight change date, but it was mainly taken between 1 week and 2 weeks. Depending on the patient, if excessive body water was suspected, continuous chest x-ray was performed every dialysis day.

ANS) As you indicated, that sentence is misleading. So, we changed the sentence as following,

Additionally, we evaluated whether a more accurate clinical dry weight could be presented using machine learning based on the various clinical factors and BIS data of the patient.

◊ The reference (REF 36) does not fit here. Can the authors please explain, using a patient flow chart, how exactly that one person was able to perform BIS in all 1,672 patients? Did one person measure 1-2 consecutive patients per day, thus between 300 and 600 a year? How did the authors deal with the effect of time? Did the dry weight protocol change from beginning to end of the retrospective study period?

ANS) For Reference 36 insertion, it is a duplicate insertion. Thanks for the right point. I deleted it.

In hospitals where the research was conducted, specialized nurses trained to perform outpatient examinations, etc. reside in the internal medicine department. During the study period, one professional nurse performed the same task, and the dedicated nurse was in June 2020. Dedicated nurses worked daily from 9am to 5pm, Monday through Friday. The number of patients examined was 0-7 per day, which varied from day to day. The examination was not performed while the nurse was absent, and the period of absence was very short, so there were no significant restrictions on the performance of the examination.

Outpatient dialysis patients visited the internal medicine outpatient clinic before dialysis on the day of dialysis and were instructed to measure dry weight through the BIS device, and dialysis was performed immediately after the dry weight measurement.

Dry weight measurement using the BIS device used the same protocol.

ANS) BIS measurement method is not Gold Standard. It should be considered that the clinical dry weight is the gold standard. Although the clinical dry weight is gold standard, finding this value actually takes a lot of time and trial and error because it means that the patient's condition is stable while continuing to dialysis. On the other hand, the BIS method is an easy method because you only have to invest about 10 minutes to measure it, and there is a relatively accurate part in evaluating the patient's hydration status.

However, there are cases where the BIS method is not well suited for predicting the actual clinical dry weight, and in this study, it means that the frequency of patients who do not fit well in the case of elderly and female diabetes in this study is high. The fact that the reviewers did not understand the definition of GAPDW may have determined that the BIS method's prediction of water status is inconsistent due to a mixture of overestimating or underestimating water status than the actual clinical status. However, in this study, we did not focus on the over-underestimation of the patient's hydration status, but whether the BIS method measurement can reduce excessive deviation from the actual clinical dry weight value through machine learning. And we focused on the characteristics of people who are excessively deviated.

This means that elderly and diabetic women often deviate from the actual dry weight value.

ANS) According to a 2018 report from the Insan Memorial Dialysis Registry (ESRD Registry Committee) collected by the Korean Society of Nephrology, the total number of hemodialysis patients was 77,617, of which 59% were male. In this center, there are variations every year, but the proportion of males in outpatient hemodialysis clinics is around 50%. However, if ward dialysis patients account for 35-45% of the total number of outpatients, it is considered to have a similar value to the Korean Society of Nephrology. In addition, the low proportion of men in this study is thought to have been derived because the proportion of men removed from the exclusion criteria was high, and the data were not separately adjusted to the gender ratio. This study is a single-center study, and we believe that this representation problem can be supplemented when multi-center studies are conducted in the future.

ANS) It can be seen that the probability of prediction for the target attribute OHcp (Table 2) by utilizing machine learning methods is probabilistically, the larger the GapDW is in the group containing the larger value.

I used the expression that DWgap is more difficult because the absolute error is increasing in the group with large DWgap, and the large group is increasing in the group with large DWgap.

◊ Can the authors please clarify what they mean? In my understanding, the two sentences that I underlined (“In the box plots, the mean value of ECW showed a tendency to decrease as GapDW increased.” versus “As GapDW increases, ECW tends to increase.”) say exactly the opposite of one another.

ANS)As GapDW increases, ECW tends to increase. Modified in Manuscipt.

Overall judgement:

Minor:

ANS) The following sentences have been deleted to eliminate misunderstandings.

References

1. Henderson LW: Symptomatic hypotension during hemodialysis. Kidney Int 1980;17:571-576

2. Agarwal R, Weir MR: Dry-weight: a concept revisited in an effort to avoid medication-directed approaches for blood pressure control in hemodialysis patients. Clin J Am Soc Nephrol 2010;5:1255-1260

PLoS One. doi: 10.1371/journal.pone.0250467.r003

Decision Letter 1

Bhagwan Dass

7 Apr 2021

A novel approach to dry weight adjustments for dialysis patients using machine learning

PONE-D-20-40081R1

Dear Dr. Choi,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Bhagwan Dass, MD

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

PLoS One. doi: 10.1371/journal.pone.0250467.r004

Acceptance letter

Bhagwan Dass

12 Apr 2021

PONE-D-20-40081R1

A novel approach to dry weight adjustments for dialysis patients using machine learning

Dear Dr. Choi:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Bhagwan Dass

Academic Editor

PLOS ONE

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

[pone.0250467.ref001] 1.Chang S-T, Chen C-C, Chen C-L, Cheng H-W, Chung C-M, Yang T-Y. Changes of the cardiac architectures and functions for chronic hemodialysis patients with dry weight determined by echocardiography. Blood Purif. 2004;22(4):351–359. 10.1159/000080031 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref002] 2.Wizemann V, Wabel P, Chamney P, Zaluska W, Moissl U, Rode C, et al. The mortality risk of overhydration in haemodialysis patients. Nephrol Dial Transplant. 2009;24(5):1574–1579. 10.1093/ndt/gfn707 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref003] 3.Charra B. Does empirical long slow dialysis result in better survival? If so, how and why? ASAIO J. 1993. [PubMed] [Google Scholar]

[pone.0250467.ref004] 4.Sinha AD, Agarwal R. Opinion: can chronic volume overload be recognized and prevented in hemodialysis patients? The pitfalls of the clinical examination in assessing volume status. Semin Dial. 2009;22(5):480–482. 10.1111/j.1525-139X.2009.00641.x [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref005] 5.Onofriescu M, Siriopol D, Voroneanu L, Hogas S, Nistor I, Apetrii M, et al. Overhydration, cardiac function and survival in hemodialysis patients. PLoS One. 2015;10(8):e0135691. 10.1371/journal.pone.0135691 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref006] 6.Vertes V, Cangiano JL, Berman LB, Gould A. Hypertension in end-stage renal disease. NEJM. 1969;280(18):978–981. 10.1056/NEJM196905012801802 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref007] 7.Fishbane S, Natke E, Maesaka JK. Role of volume overload in dialysis-refractory hypertension. Am J Kidney Dis. 1996;28(2):257–261. 10.1016/s0272-6386(96)90309-1 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref008] 8.Amato R, Hlebovy D, King B, Salai P. Fluid removal. Obtaining estimated dry weight during hemodialysis. ANNA Core Curriculum for Nephrology Nursing. 2008:692–703. [Google Scholar]

[pone.0250467.ref009] 9.Sands JJ, Usvyat LA, Sullivan T, Segal JH, Zabetakis P, Kotanko P, et al. Intradialytic hypotension: Frequency, sources of variation and correlation with clinical outcome. Hemodial Int. 2014;18(2):415–422. 10.1111/hdi.12138 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref010] 10.Daugirdas JT, Blake PG, Ing TS. Handbook of dialysis. 4th ed: Lippincott Williams & Wilkins; 2012. [Google Scholar]

[pone.0250467.ref011] 11.San Miguel S. Haemodialysis dry weight assessment: A literature review. Ren Soc Aust J. 2010;6(1):19–24. [Google Scholar]

[pone.0250467.ref012] 12.Poggi A, Maggiore Q. Cardiothoracic ratio as a guide to ultrafiltration therapy in dialyzed patients. Int J Artif Organs. 1980;3(6):332–337. [PubMed] [Google Scholar]

[pone.0250467.ref013] 13.Vitturi N, Dugo M, Soattin M, Simoni F, Maresca L, Zagatti R, et al. Lung ultrasound during hemodialysis: the role in the assessment of volume status. Int Urol Nephrol. 2014;46(1):169–174. 10.1007/s11255-013-0500-5 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref014] 14.Agarwal R, Bouldin JM, Light RP, Garg A. Inferior vena cava diameter and left atrial diameter measure volume but not dry weight. Clin J Am Soc Nephrol. 2011;6(5):1066–1072. 10.2215/CJN.09321010 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref015] 15.Passauer J, Petrov H, Schleser A, Leicht J, Pucalka K. Evaluation of clinical dry weight assessment in haemodialysis patients using bioimpedance spectroscopy: a cross-sectional study. Nephrol Dial Transplant. 2010;25(2):545–551. 10.1093/ndt/gfp517 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref016] 16.Chamney PW, Krämer M, Rode C, Kleinekofort W, Wizemann V. A new technique for establishing dry weight in hemodialysis patients via whole body bioimpedance. Kidney Int. 2002;61(6):2250–2258. 10.1046/j.1523-1755.2002.00377.x [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref017] 17.Gupta A, Liu T, Shepherd S, Paiva W. Correction: Using Statistical and Machine Learning Methods to Evaluate the Prognostic Accuracy of SIRS and qSOFA. Healthc Inform Res. 2018;24(3):250. 10.4258/hir.2018.24.3.250 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref018] 18.Desautels T, Hoffman J, Barton C, Mao Q, Jay M, Calvert J, et al. Pediatric severe sepsis prediction using machine learning. bioRxiv. 2017:223289. [Google Scholar]

[pone.0250467.ref019] 19.Arefan D, Mohamed AA, Berg WA, Zuley ML, Sumkin JH, Wu S. Deep learning modeling using normal mammograms for predicting breast cancer risk. Med Phys. 2020;47(1):110–118. 10.1002/mp.13886 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref020] 20.Ke G, Meng Q, Finley T, Wang T, Chen W, Ma W, et al., editors. Lightgbm: A highly efficient gradient boosting decision tree. Adv Neural Inf Process Syst; 2017. [Google Scholar]

[pone.0250467.ref021] 21.Chen T, Guestrin C, editors. Xgboost: A scalable tree boosting system. Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining; 2016 August; San Francisco California USA: Association for Computing Machinery, New York, NY, United States.

[pone.0250467.ref022] 22.Ho TK, editor Random decision forests. Proceedings of 3rd international conference on document analysis and recognition; 1995 Aug 14–16; Montreal, Quebec, Canada, Canada: IEEE; 2002.

[pone.0250467.ref023] 23.Jaeger JQ, Mehta RL. Assessment of dry weight in hemodialysis: an overview. J Am Soc Nephrol. 1999;10(2):392–403. [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref024] 24.Lee SW, Song JH, Kim GA, Lim HJ, Kim M-J. Plasma brain natriuretic peptide concentration on assessment of hydration status in hemodialysis patient. Am J Kidney Dis. 2003;41(6):1257–1266. 10.1016/s0272-6386(03)00358-5 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref025] 25.Ando Y, Yanagiba S, Asano Y. The inferior vena cava diameter as a marker of dry weight in chronic hemodialyzed patients. Artificial Organs. 1995;19(12):1237–1242. 10.1111/j.1525-1594.1995.tb02292.x [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref026] 26.Cox-Reijven PL, Kooman JP, Soeters PB, van der Sande FM, Leunissen KM. Role of bioimpedance spectroscopy in assessment of body water compartments in hemodialysis patients. Am J Kidney Dis. 2001;38(4):832–838. 10.1053/ajkd.2001.27703 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref027] 27.Dumler F. Hypoalbuminemia is a marker of overhydration in chronic maintenance patients on dialysis. ASAIO J. 2003;49(3):282–286. 10.1097/01.mat.0000065465.52748.bb [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref028] 28.Zaki DS, Mohamed RR, Mohammed NA, Abdel-Zaher RB. Assessment of malnutrition status in hemodialysis patients. Clinical Medicine and Diagnostics. 2019;1:0.20. [Google Scholar]

[pone.0250467.ref029] 29.Jones CH, Akbani H, Croft DC, Worth DP. The relationship between serum albumin and hydration status in hemodialysis patients. J Ren Nutr. 2002;12(4):209–212. 10.1053/jren.2002.35295 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref030] 30.Garagarza C, Valente A, Caetano C, Oliveira T, Ponce P, Silva AP. Hypophosphatemia: nutritional status, body composition, and mortality in hemodialysis patients. International urology nephrology. 2017;49(7):1243–1250. 10.1007/s11255-017-1558-2 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref031] 31.Hung SC, Kuo KL, Peng CH, Wu CH, Wang YC, Tarng DC. Association of fluid retention with anemia and clinical outcomes among patients with chronic kidney disease. J Am Heart Assoc. 2015;4(1):e001480. 10.1161/JAHA.114.001480 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref032] 32.Androne A-S, Katz SD, Lund L, LaManca J, Hudaihed A, Hryniewicz K, et al. Hemodilution is common in patients with advanced heart failure. Circ. 2003;107(2):226–229. 10.1161/01.cir.0000052623.16194.80 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref033] 33.Patel SS, Molnar MZ, Tayek JA, Ix JH, Noori N, Benner D, et al. Serum creatinine as a marker of muscle mass in chronic kidney disease: results of a cross-sectional study and review of literature. J Cachexia Sarcopenia Muscle. 2013;4(1):19–29. 10.1007/s13539-012-0079-1 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref034] 34.Molnar MZ, Streja E, Kovesdy CP, Bunnapradist S, Sampaio MS, Jing J, et al. Associations of body mass index and weight loss with mortality in transplant‐waitlisted maintenance hemodialysis patients. Am J Transplant. 2011;11(4):725–736. 10.1111/j.1600-6143.2011.03468.x [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0250467.ref035] 35.Vinge E, Lindergård B, Nilsson–Ehle P, Grubb A. Relationships among serum cystatin C, serum creatinine, lean tissue mass and glomerular filtration rate in healthy adults. Scand J Clin Lab Invest. 1999;59(8):587–592. 10.1080/00365519950185076 [DOI] [PubMed] [Google Scholar]

[pone.0250467.ref036] 36.Nakhoul GN, Huang H, Arrigain S, Jolly SE, Schold JD, Nally JV Jr, et al. Serum potassium, end-stage renal disease and mortality in chronic kidney disease. J Am Soc Nephrol. 2015;41(6):456–463. 10.1159/000437151 [DOI] [PMC free article] [PubMed] [Google Scholar]

PERMALINK

A novel approach to dry weight adjustments for dialysis patients using machine learning

Hae Ri Kim

Hong Jin Bae

Jae Wan Jeon

Young Rok Ham

Ki Ryang Na

Kang Wook Lee

Yun Kyong Hyon

Dae Eun Choi

Roles

Abstract

Background and aims

Methods

Results

Conclusions

Introduction

Materials and methods

Study design

Patient selection

Body composition measurements

Definitions

Measurements of clinical parameters

Machine-learning adjustment

Results

Baseline characteristics

Table 1. Baseline characteristics.

Prediction of clinical DW using machine learning

Table 2. Prediction using machine learning.

Fig 1. Feature importance for the two groups (Group 1: GapDW < 1 kg and Group 2: GapDW < 2 kg).

The factors affecting GapDW

Fig 2.

Discussion

Data Availability

Funding Statement

References

Decision Letter 0

Bhagwan Dass

Roles

Author response to Decision Letter 0

Decision Letter 1

Bhagwan Dass

Roles

Acceptance letter

Bhagwan Dass

Roles

Associated Data

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases

Fig 1. Feature importance for the two groups (Group 1: Gap_DW < 1 kg and Group 2: Gap_DW < 2 kg).

The factors affecting Gap_DW