Translational Prediction of Hemodialysis Clearance Integrating Molecular Structure and Dialysis Conditions

ABSTRACT

Quantitative prediction of drug removal by hemodialysis is essential for rational dose design in patients with renal failure. We aimed to develop a machine learning‐based quantitative structure–pharmacokinetic relationship (QSPKR) model to predict hemodialysis clearance (CL_HD) using molecular descriptors and dialysis conditions. CL_HD data were compiled from published clinical studies, yielding a dataset comprising 46 compounds; 457 measurements; 7 dialysis membrane types; and dialysis parameters, including blood flow rate (Q _B) and dialysate flow rate (Q _D). Recursive feature elimination revealed six key predictors, which were used to construct and compare five models: lasso regression, random forest, support vector machine, averaged neural network, and extreme gradient boosting (XGBoost) models. Feature contributions were interpreted using Shapley additive explanations (SHAP) analysis. Among the models, XGBoost demonstrated the highest predictive performance (root mean square error = 51.71 mL/min; R ² = 0.614). The SHAP analysis revealed that lipophilicity was the most influential predictor. The effects of Q _B and Q _D on CL_HD were nonlinear and saturable, consistent with the established dialysis clearance theory. CL_HD can be predicted from molecular structure and dialysis parameters using a QSPKR‐based machine learning approach, providing a practical framework for estimating drug clearance and assisting in the preliminary evaluation of drug dialyzability when empirical data are limited.

Study Highlights

• What is the current knowledge on the topic?

  • ○
    Drug removal during hemodialysis varies widely depending on molecular properties and dialysis conditions. Dose adjustments are typically based on empirical parameters such as protein binding and volume of distribution, while quantitative prediction of hemodialysis clearance (CL_HD) during drug development remains limited.
• What question did this study address?

  • ○
    Can hemodialysis clearance be quantitatively predicted from molecular descriptors and dialysis operational parameters using a machine learning‐based translational framework?
• What does this study add to our knowledge?

  • ○
    This study establishes an integrative model linking molecular structure with dialysis parameters to predict CL_HD. The results demonstrate that nonlinear interactions between drug physicochemical properties and dialysis flow conditions govern drug removal, and that these relationships can be captured using machine learning with interpretable feature attribution.
• How might this change clinical pharmacology or translational science?

  • ○
    The proposed framework enables prospective estimation of drug exposure during hemodialysis by integrating predicted CL_HD into pharmacokinetic models. Accordingly, the framework should be regarded as a tool for preliminary assessment of drug dialyzability rather than for individual dose selection until further validation is available.

Machine learning‐based prediction of hemodialysis clearance (CL_HD) integrating molecular descriptors and dialysis conditions. XGBoost showed the best predictive performance, with lipophilicity as the most influential predictor. The model supports preliminary estimation of CL_HD and assessment of drug dialyzability when empirical dialysis data are limited.

1. Introduction

Hemodialysis is a renal replacement therapy used to remove harmful substances, including toxins and metabolic waste products, from the blood. It is commonly used in patients with various clinical conditions, such as acute renal failure, chronic kidney disease, and drug‐induced poisoning. In acute renal failure, hemodialysis supports renal function by efficiently clearing uremic toxins and managing fluid overload. Daily intermittent hemodialysis improves survival and accelerates the resolution of acute renal failure compared with conventional alternate‐day hemodialysis by offering better uremic control with fewer complications (hypotension) [1]. For patients with chronic kidney disease, particularly end‐stage renal disease, hemodialysis is life‐sustaining as native kidney function declines. These patients usually need multiple medications for comorbidities, and hemodialysis alters drug pharmacokinetics by affecting absorption, distribution, metabolism, and excretion, thereby requiring careful management to avoid drug accumulation and adverse events [2]. Regarding drug‐induced poisoning and critical illness, hemodialysis and other renal replacement therapies, such as continuous venovenous hemofiltration, can remove toxic substances and certain drugs, such as the antibiotic linezolid, to prevent elevated and harmful serum concentrations [3]. Despite its clinical significance, quantitative prediction of hemodialysis clearance (CL_HD) during drug development remains limited, as dialysis‐specific pharmacokinetic data are scarce in the early stages.

Pharmacokinetic parameters critical for predicting drug removal by hemodialysis include protein binding, volume of distribution, and elimination half‐life. Drugs with low protein binding and volume of distribution are more readily removed by hemodialysis because only unbound drugs are dialyzable, and drugs confined primarily to the vascular compartment are more accessible to the dialysis membrane. Various dialysis modalities and membranes differ in their drug removal efficiency. For instance, extended daily dialysis and continuous renal replacement therapy have distinct effects on drug clearance compared with conventional hemodialysis.

Empirical dosing adjustments are often based on drug protein binding and distribution volume; however, these can be imprecise owing to the complex interplay among patient factors, dialysis parameters (e.g., blood flow rate [Q _B], dialysate flow rate [Q _D], membrane type, and surface area), and drug properties [4]. More precise quantitative predictions should incorporate measured dialytic clearance, plasma protein‐binding fraction, volume of distribution, and dialysis session characteristics. Although empirical extrapolation using protein binding and distribution volume is commonly employed, an accurate quantitative prediction of drug removal during hemodialysis requires pharmacokinetic data integration. This includes CL_HD measurements, consideration of dialysis timing relative to drug administration, and attention to dialysis modality and membrane characteristics. Such predictive methods remain underdeveloped during drug development, highlighting the need for focused pharmacokinetic studies in patients undergoing hemodialysis to optimize dosing and improve therapeutic outcomes [5, 6].

As drugs can be removed during hemodialysis, dosing should account for the amount of drug cleared by the procedure, particularly when administration occurs during dialysis. Even with the same dialysis membrane, the extent of removal varies with drugs, making dose optimization extremely challenging. However, during pharmaceutical development, only a few studies have investigated drug pharmacokinetics under hemodialysis therapy, which is a unique pathological condition. When the rate of drug removal by hemodialysis is unknown, dosing is determined empirically by extrapolating from the drug's protein‐binding rate in plasma and distribution volume. Therefore, a quantitative prediction method for drug removal by hemodialysis is crucial.

Drug removal by hemodialysis is influenced by the affinity of the drug for the dialysis membrane and blood constituents. The affinity between drugs and dialysis membranes is related to drug physicochemical properties and membrane characteristics (e.g., pore size, charge, and membrane area); however, only a few quantitative studies have been conducted [7]. The analytical approach involving the use of mathematical models, artificial intelligence, and machine learning to represent the relationship among the molecular structure, physicochemical properties, and pharmacokinetics of a compound is termed quantitative structure–pharmacokinetic relationship (QSPKR) analysis [8]. Numerous mathematical models have been used to predict the plasma clearance, volume of distribution, and protein‐binding rate of compounds using the QSPKR analysis. However, only multiple regression analyses or machine learning models have been used for removal rates during hemodialysis, with no reports on CL_HD [7, 9]. Therefore, we developed a machine learning‐based QSPKR model to predict CL_HD by integrating molecular descriptors with dialysis parameters, supporting pharmacokinetic modeling and preliminary assessment of drug dialyzability. This study provides a translational framework linking molecular structure and dialysis conditions to enable prospective estimation of drug exposure during hemodialysis.

2. Methods

2.1. Data Collection and Preprocessing

We collected CL_HD data from published articles to create a database. A literature search was conducted using PubMed and Google Scholar with the term “hemodialysis clearance.” Studies were included if they met the following criteria: (1) involved patients undergoing hemodialysis; (2) reported individual CL_HD values; and (3) provided details regarding CL_HD, Q _B, Q _D, dialysis membrane type, and the name of the compound administered under hemodialysis conditions. For each literature source, the method used to determine CL_HD was reviewed during data extraction. The included studies did not use a single uniform method to determine CL_HD. The reported CL_HD values were derived using different approaches, including arteriovenous concentration‐based methods, dialysate recovery‐based methods, or model‐based estimation from blood or plasma concentration data. CL_HD values were included only when they were explicitly reported as dialysis clearance or could be identified as clearance attributable to hemodialysis. The values were used as reported, without recalculation. When multiple CL_HD measurements were reported for a single patient, each measurement was included as a separate observation with its corresponding dialysis conditions and CL_HD value. Dialysis membrane type was assigned according to the original publication whenever patient‐level information was available. For theophylline, the membrane type was not reported, and Q _D values ranged from 325 to 350 mL/min; therefore, the M1 membrane (cuprophan), the most frequently represented membrane in the dataset, was assumed, and 333 mL/min was assigned as the representative value. For 8 of the 44 vancomycin CL_HD measurements, measurement‐specific Q _B values were not reported; therefore, the reported mean Q _B value from the original publication was used.

2.2. Calculation of Molecular Descriptors

Structural data for each drug were obtained from the PubChem database (http://www.ncbi.nlm.nih.gov/pcsubstance/). Additionally, structural data for the water molecules and counter ions were eliminated by processing the disposal salt. Molecular descriptors, including pharmacokinetic properties, were calculated using ADMET Predictor 10.4 (Simulations Plus Inc., Lancaster, CA) using the two‐dimensional structure of each drug. Of the descriptors calculated by ADMET Predictor, the 123 represented as numerical or categorical variables were retained for further analysis. The descriptors are shown in Table S1.

2.3. Feature Selection

The response variable was CL_HD (mL/min) and the explanatory variables were Q _B (mL/min), Q _D (mL/min), 7 dialysis membrane types (M1, cuprophan; M2, cellulose acetate; M3, polysulfone; M4, regenerated cellulose; M5, polyacrylonitrile; M6, hemophane; M7, polyamide), and 123 molecular descriptors. The CL_HD values were natural logarithm‐transformed before model construction. The dataset was divided into training (80%) and test (20%) sets on a compound basis, ensuring that no identical compounds appeared in both sets. The training data were standardized (centered and scaled), and the resulting scaling parameters were applied to the test dataset. To improve model generalization and interpretability, we performed feature selection on the training dataset containing 125 candidate features, excluding the seven dialysis membrane type variables. Furthermore, we used recursive feature elimination (RFE), a wrapper method that recursively tests smaller feature sets to select the optimal subset. Random forest (RF) was used as the external estimator for RFE to account for potential nonlinear interactions between variables. RFE was performed before model development to determine the optimal subset of features. Candidate feature subset sizes of 5–15, 20, 25, 30, 35, 40, and 50 were evaluated. The optimal number of features was selected as the subset size that achieved the lowest cross‐validated root mean square error (RMSE). Based on this criterion, six features were retained for subsequent model development.

2.4. Construction of CL_HD Models

Using the six features selected by RFE and seven dialysis membrane types, we developed and compared five predictive models, namely lasso regression (LASSO; linear baseline), RF, support vector machine (SVM), averaged neural network (avNNet), and extreme gradient boosting (XGBoost) models. Hyperparameters for each model were optimized using an RMSE‐based grid search with leave‐one‐compound‐out cross‐validation (LOCO‐CV). In LOCO‐CV, all observations from one compound were excluded from training and used exclusively for validation in each iteration. The hyperparameter combination yielding the lowest mean cross‐validated RMSE on the natural‐log‐transformed CL_HD scale was selected as optimal. RMSE was calculated using the following equation:

$$\text{RMSE}=\sqrt{\frac{\sum {\left(\text{predicted}-\text{observed}\right)}^2}{n}};$$ (1)

where n refers to the number of observations evaluated. The final model was then refitted on the entire training dataset with the selected hyperparameters. Model development was performed using the caret package (version 7.0‐1). All analyses were conducted using R (version 4.4.1) in the RStudio environment (version 2025.09.2, Build 418).

2.5. Model Evaluation

The predictability of each model was assessed using the RMSE and mean absolute error (MAE), calculated using the following equation:

$$\mathrm{MAE}=\frac{\sum \left|\text{predicted}-\text{observed}\right|}{n};$$ (2)

where n refers to the number of observations evaluated. The linearity between the predicted and reported CL_HD values was evaluated using the coefficient of determination (R ²). The predicted values were re‐transformed using Duan's smearing estimator. Predictability was assessed using the test dataset.

To interpret the prediction mechanism of the best‐performing model, we used Shapley additive explanations (SHAP) values to quantify the contribution of each feature. To further evaluate this model, diagnostic plots were generated using the back‐transformed predicted values. These included observed versus predicted plots for the training dataset and LOCO‐CV, a residual versus predicted plot, and a residual histogram. Residuals were calculated on the original CL_HD scale as the difference between observed and predicted CLHD (observed − predicted CL_HD).

3. Results

3.1. Selection of Features Using RFE

The CL_HD dataset, obtained from previous studies, comprised 46 compounds, 457 measurements, and seven dialysis membranes. The histogram of CL_HD is shown in Figure S1. The details of CL_HD, Q _B, Q _D, and dialysis membrane types for the target compounds are summarized in Table S2. RFE effectively reduced the dimensionality of the initial dataset while retaining the predictive information. From the original variable pool, RFE identified the following key features for CL_HD prediction: Moriguchi model of the octanol–water partition coefficient (MlogP), total number of halogen atoms (N_Halogen), maximum sigma Fukui index (EEM_MaxF), Q _B, Q _D, and medium ellipsoidal radius of the molecule based on its size and shape (T_Radb). Selecting these variables indicates that dialysis conditions, molecular size, lipophilicity, and electronic factors are instrumental in determining target properties.

3.2. Comparative Analysis of Model Performance

We evaluated the predictive performance of the five models (LASSO, RF, SVM, avNNet, and XGBoost) using six selected features. The performance metrics (RMSE, MAE, and R ²) for the independent test set are summarized in Table 1 and visualized in Figure 1. Among the evaluated models, XGBoost demonstrated the highest predictive accuracy, achieving the lowest RMSE (51.71) and the highest R ² (0.614) values, followed by avNNet. The diagnostic plots and residual histogram are shown in Figures S2–S5.

TABLE 1.

Comparison of models.

Model	RMSE (mL/min)	MAE (mL/min)	R 2
Extreme gradient boosting	51.71	37.04	0.614
Averaged neural network	52.35	39.37	0.476
Random forest	53.46	37.99	0.584
Support vector machine	55.92	42.88	0.455
Lasso regression	57.44	42.88	0.444

Abbreviations: MAE, mean absolute error; R ², coefficient of determination; RMSE, root mean square error.

FIGURE 1.

Predictive performance of various machine learning models on the test dataset. The solid line represents unity (y = x) and the dotted lines indicate the two‐fold error limits.

3.3. Interpretability and Feature Importance

To elucidate the decision‐making process of the best‐performing model (XGBoost), we conducted a SHAP analysis. The SHAP summary plot (Figure 2) ranks the features according to their mean absolute SHAP values. MlogP was the most dominant predictor, followed by T_Radb and N_Halogen. Furthermore, the SHAP dependence plot for Q _B and Q _D (Figure 3) revealed distinct nonlinear relationships. For Q _B, the positive contribution increased linearly between 150 and 400 mL/min; however, the rate of increase declined above this threshold. Similarly, the effect of Q _D increased linearly up to 500 mL/min, followed by a plateau at higher flow rates.

FIGURE 2.

SHAP summary plot derived from the training dataset. EEM_MaxF, maximum sigma Fukui index; M1, cuprophan; M1–M7, dialysis membrane type; M2, cellulose acetate; M3, polysulfone; M4, regenerated cellulose; M5, polyacrylonitrile; M6, hemophane; M7, polyamide; MlogP, Moriguchi model of the octanol–water partition coefficient; N_Halogen, total number of halogen atoms; Q _B, blood flow rate; Q _D, dialysate flow rate; SHAP, Shapley additive explanations; T_Radb, medium ellipsoidal radius of the molecule based on its size and shape.

FIGURE 3.

SHAP dependence plot derived from the training dataset. MlogP, Moriguchi model of the octanol–water partition coefficient; Q _B, blood flow rate; Q _D, dialysate flow rate; SHAP, Shapley additive explanations.

4. Discussion

CL_HD varies depending on the affinity between the drug and dialysis membrane, as well as the affinity between the drug and biological components. For instance, studies on aminoglycosides, such as gentamicin and tobramycin, have determined that these hydrophilic drugs with a relatively low protein‐binding ability are substantially removed by hemodialysis, with CL_HD often exceeding systemic clearance, leading to significant drug removal during dialysis sessions [6, 10]. Similarly, the amikacin pharmacokinetic study has shown a substantial decrease in half‐life during hemodialysis owing to effective drug removal, emphasizing the need for dosage adjustment [11]. In contrast, micafungin pharmacokinetics is unaffected by dialysis; therefore, dosage adjustment is not required [12]. On the basis of these findings, the effects of dialysis on pharmacokinetics can vary significantly among drugs. The extent of drug elimination varies widely among drugs, even with identical dialysis membranes, complicating the prediction of drug clearance during hemodialysis. Studies on meropenem and vancomycin pharmacokinetics during extended daily dialysis have demonstrated significant drug removal, indicating that dosing regimens from intermittent hemodialysis cannot be directly applied and should be tailored, possibly with therapeutic drug monitoring [5]. Additionally, drug clearance is influenced by membrane characteristics and residual renal function in continuous renal replacement therapy, requiring individualized dosing considerations [4, 13].

The prediction of dialysis‐based drug removal rates has been reported using methods based on pharmacokinetic parameters, such as protein binding rates [7]. The prediction method developed in this study, based on molecular structures, may provide preliminary estimates without requiring new compound‐specific dialysis experiments. This study demonstrated the feasibility of predicting CL_HD using six selected features and the dialysis membrane type. Combining the predicted CL_HD with established pharmacokinetic models may facilitate drug concentration prediction during dialysis. Additionally, this integrative approach may support estimation of drug removal, accounting for the molecular characteristics and operational parameters of the dialysis process. Consequently, it provides a quantitative framework for estimating CL_HD during drug development and may assist in the preliminary evaluation of drug dialyzability when empirical data are limited. The proposed model is not intended to replace clinical pharmacokinetic studies but to provide a quantitative, structure‐based estimate of CL_HD during drug development or the early phase of clinical use.

In this study, compound‐based data splitting and independent test validation were used to minimize information leakage and assess generalization across chemical structures, despite the limited number of compounds. XGBoost showed the highest predictive accuracy (RMSE = 51.71) among the evaluated models in the test set. Notably, the nonlinear models (XGBoost, RF, avNNet, and SVM) outperformed the linear baseline (lasso regression, RMSE = 57.44). This performance gap strongly suggests that the relationship between the six features and CL_HD involves complex nonlinear interactions that cannot be captured using simple linear regression. The SHAP dependence plot (Figure 3) exhibits a nonlinear relationship between CL_HD and features Q _B and Q _D, explaining why tree‐based ensemble models, such as XGBoost, outperformed the linear lasso model. XGBoost naturally models feature interactions because each decision tree adds splits based on different features in sequence, creating hierarchical combinations of decision rules. This hierarchical splitting structure can uncover nonlinear interactions between features without explicit feature engineering. In contrast, while neural networks can model complex interactions, their ability to represent sharp, piecewise decision boundaries is often less direct unless specifically designed with appropriate architectures [14]. The superior performance of XGBoost over avNNet suggests that the model benefits from the decision tree structure, which effectively captures sharp thresholds and interactions between specific features [15]. This advantage allows XGBoost to model complex, nonlinear relationships within the data more effectively than linear methods do. Additionally, the ensemble approach of XGBoost reduces overfitting by aggregating multiple decision trees, enhancing generalizability [14, 15]. Consequently, these characteristics make XGBoost particularly well‐suited for datasets with intricate feature interactions and heterogeneous patterns [15]. Additional diagnostic plots were generated to examine the predictive behavior of the best‐performing XGBoost model. In the training dataset, the observed and predicted CL_HD values showed good agreement, whereas the LOCO‐CV analysis exhibited greater dispersion, reflecting the more stringent evaluation of model generalizability to unseen compounds. The residual‐versus‐predicted plot did not reveal a pronounced nonlinear pattern or marked heteroscedasticity. However, a slight downward trend in residuals was observed at higher predicted CL_HD values, suggesting a modest tendency toward overprediction in this range. The residual histogram was approximately centered on zero, although a slight negative shift and several large residuals were observed. These results indicate that the model captured the overall trend in CL_HD but that prediction uncertainty remains, particularly for observations in the higher CL_HD range.

Dialysis clearance is generally expressed using Michaels' equation [16], a function of Q _B, Q _D, and the overall mass transfer area coefficient (KoA), as follows:

$${\mathrm{CL}}_{\mathrm{HD}}=Q_{\mathrm{B}}·\frac{1-\exp \left[\mathrm{KoA}\left(1/Q_{\mathrm{B}}-1/Q_{\mathrm{D}}\right)\right]}{Q_{\mathrm{B}}/Q_{\mathrm{D}}-\exp \left[\mathrm{KoA}\left(1/Q_{\mathrm{B}}-1/Q_{\mathrm{D}}\right)\right]}.$$ (3)

We have previously reported a model for predicting CL_HD using Michaels' equation and mixed‐effects models, in which KoA was expressed as a function of logBB, the decimal logarithm of the brain‐to‐blood partition coefficient [17]. In the SHAP dependence plot (Figure 3), the influence of Q _B and Q _D on CL_HD exhibited a convex curve, consistent with Michaels' equation. In the present model, MlogP demonstrated the greatest effect on CL_HD prediction (Figure 2). Lipophilicity facilitates penetration across the blood–brain barrier, resulting in a positive correlation between logP (including MlogP) and logBB. Therefore, although MlogP was selected over logBB in this study, the finding remains consistent with our previous report [17], as both descriptors reflect similar physicochemical properties of the compounds.

Several pharmacokinetic models have been developed to estimate CL_HD. These models represent the patient's intrinsic drug elimination as plasma clearance (when dialysis is not performed) and describe drug elimination during dialysis as the sum of plasma clearance and CL_HD [18]. For drugs without reported pharmacokinetic models in patients undergoing dialysis, blood drug concentrations may be simulated by combining previously reported pharmacokinetic models with the CL_HD model developed in this study. The integration of structure‐based CL_HD prediction into existing population pharmacokinetic models may enable prospective simulation of drug exposure during dialysis sessions. Consequently, this combined modeling strategy may support further pharmacokinetic evaluation when dialysis‐specific data are limited. In addition, predicted CL_HD may provide supportive information for assessing the potential effectiveness of dialysis in cases of overdose or drug toxicity [19], although further validation is required before such predictions can be used for clinical decision‐making.

This study has some limitations. First, the proposed model cannot predict the individual differences in CL_HD. In the constructed model, CL_HD was calculated as the same value under identical conditions for the compound, Q _B, Q _D, and dialysis membrane. This approach limits the model's ability to account for the variability in patient‐specific factors that may influence CL_HD. The present model focuses on compound‐specific and dialysis condition‐dependent determinants of CL_HD rather than on individual patient variability, making it complementary to population pharmacokinetic approaches. Second, it does not incorporate individual physiological parameters, such as blood components, which may influence CL_HD. Third, dynamic changes in compound properties and membrane characteristics are not integrated into the model, potentially affecting its clinical applicability. Fourth, the methods used to determine CL_HD were not uniform across the included literature sources. Some studies derived CL_HD from arteriovenous concentration differences across the dialyzer, whereas others used dialysate recovery‐based methods and model‐based estimation. These methodological differences may have contributed to unexplained variability and to the residual prediction error observed in this study. Fifth, the CL_HD distribution was imbalanced, with relatively few observations in the higher CL_HD range. Because hyperparameters were selected using RMSE on the natural‐log scale rather than the original CL_HD scale, model selection did not directly minimize prediction error in mL/min. The limited representation of high‐CL_HD observations may also have contributed to reduced predictive accuracy in this range. Therefore, predictions in the higher CL_HD range should be interpreted with caution. Finally, differences in dataset composition may have contributed to the remaining prediction bias. Observed CL_HD values were higher in the test dataset than in the training dataset, with medians of 117.00 mL/min (interquartile range, 75.83–171.25) and 68.70 mL/min (31.70–102.20), respectively. In contrast, the mean N_Halogen value was lower in the test dataset than in the training dataset (0.01087 vs. 0.6877), although the median was zero in both. In the training‐set SHAP analysis, lower N_Halogen values were generally associated with negative contributions to the model output. Moreover, the limited representation of some membrane types in the training dataset may have constrained the learning of membrane‐specific relationships, contributing to uncertainty in predictions involving these membrane types. These findings suggest that differences in the predictor and outcome distributions, in conjunction with the relationships learned by the model, may have contributed to underprediction in the test dataset. However, the contribution of these differences was not quantified, and summary statistics alone cannot establish inadequate coverage of the multivariable predictor space. Underprediction of CL_HD could lead to underestimation of dialysis‐related drug removal if the predicted values were used directly to inform dosing. Therefore, the model should not be used as the sole basis for individual dose selection.

In this study, we constructed a machine learning model to predict CL_HD based on the QSPKR analysis. Among the multiple models evaluated, XGBoost showed the best predictive performance in the independent test dataset. Future studies should focus on integrating more comprehensive clinical data to further improve model accuracy and applicability. This approach provides a quantitative framework for preliminary assessment of drug dialyzability in patients undergoing hemodialysis. However, further validation is required before the model can be used to inform individual dose adjustment.

Author Contributions

T.A., T.K., A.S., H.O., J.Y., T.I., K.K., and Y.T. wrote the manuscript. T.A. and Y.T. designed the research. T.A., A.S., and H.O. performed the research. T.A. and A.S. analyzed the data.

Funding

This work was supported by JSPS KAKENHI grant number JP24K09969.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Summary of descriptors.

Table S2: Summary of compounds.

Figure S1: Distribution of observed hemodialysis clearance. Histograms of observed CL_HD values are shown for (A) the full dataset, (B) the training dataset, and (C) the test dataset.

Figure S2: Observed versus predicted hemodialysis clearance values in the training dataset. The solid line represents unity, dotted lines indicate the two‐fold error limits, and blue line represents the fitted regression line.

Figure S3: Observed versus predicted hemodialysis clearance values obtained from leave‐one‐compound‐out cross validation. The solid line represents unity, dotted lines indicate the two‐fold error limits, and blue line represents the fitted regression line.

Figure S4: Residual plots obtained from leave‐one‐compound‐out cross validation. Residuals were calculated as observed CL_HD minus predicted CL_HD on the original scale. The blue line represents the fitted regression line.

Figure S5: Histogram of residuals obtained from leave‐one‐compound‐out cross validation. Residuals were calculated as observed CL_HD minus predicted CL_HD on the original scale.