PDF
Citation
Print
INTRODUCTION
Hematuria, defined as the presence of red blood cells (RBCs) in the urine, is an essential clinical finding associated with various renal and urological conditions [1, 2]. RBC morphology is a crucial diagnostic tool for differentiating between glomerular and non-glomerular hematuria [3, 4]. Glomerular hematuria is often associated with conditions such as glomerulonephritis, whereas non-glomerular hematuria typically arises from lower urinary tract disorders, including infections, malignancies, and calculi. Traditionally, RBC morphology assessment relies on manual microscopy, classifying RBCs as either isomorphic or dysmorphic. However, this method is labor-intensive, operator-dependent, and is subject to interobserver variability, leading to inconsistencies in the classification accuracy.
To overcome these limitations, automated systems, such as the Fully Automated Urine Particle Analyzer UF-5000, have been developed. Based on flow cytometry, this analyzer generates scattergram data that provide insights into the size, shape, and internal composition of urinary particles. Unlike imaging-based systems, scattergram data suggest particle characteristics based on computational analysis. However, few studies have explored the clinical applications of scattergram data.
Urinary RBC size and volume are critical markers for distinguishing glomerular from non-glomerular hematuria using automated analyzers [5–11]. RBC distribution and scattergram-based indices—particularly small RBC fractions—provide clinically useful discrimination against microscopic reference standards. While the UF-5000 analyzer can distinguish RBCs based on size and related characteristics using the RBC-Info feature and additionally provides built-in research parameters (e.g., small RBCs, large RBCs, lysed RBCs, non-lysed RBCs) [12–15], performance can vary across settings and may be limited by specimen characteristics, underscoring the need for robust validation.
Machine learning (ML) has emerged as a promising tool for enhancing diagnostic accuracy in medical applications in recent years [16–18]. By applying ML algorithms to UF-5000 scattergram data, complex patterns and features can be extracted, which may not be readily apparent using traditional diagnostic methods or human analysis. We aimed to improve urinary RBC morphology classification by utilizing ML models for scattergram analysis.
Accurately distinguishing between glomerular and non-glomerular RBCs using an ML-enhanced automated approach may have significant clinical implications. Such a system could facilitate faster and more reliable diagnoses of kidney and urinary tract diseases than existing systems, serving as a complementary tool to kidney biopsy. Additionally, these advancements may contribute to the standardization of urinary RBC morphological analysis. We assessed the integration of ML with UF-5000 scattergram data to address the existing challenges in urinary RBC morphology evaluation, laying a foundation for future innovations in renal and urological diagnostics.
MATERIALS AND METHODS
Samples
Urinalysis results from 648 ambulatory and hospitalized patients at The University of Tokyo Hospital (Tokyo, Japan) between August and October 2021 were randomly collected for this retrospective study. All urine samples were immediately analyzed using an UF-5000 instrument (Sysmex Corp., Kobe, Japan) and manual microscopy. Based on morphology, the RBCs in the samples were categorized as either dysmorphic or isomorphic, following the criteria outlined in Fig. 1. Ultimately, 185 samples were included in the analysis. Supplemental Data Table S1 presents the characteristics of the participants for the training set.
An independent validation dataset comprising 1,093 consecutive urinalysis results was collected at our institution between November 2023 and March 2024, using the same inclusion criteria as those used for the training set. These samples were entirely independent of the training dataset and were included without preselection to reflect routine clinical testing conditions. All samples were analyzed using the UF-5000 analyzer and manual microscopy. Scattergram data were obtained under the same conditions as those used for the ML model development. Supplemental Data Table S2 presents the characteristics of the participants for the validation set.
Manual microscopic examination
Four certified medical technologists specializing in urinalysis performed manual microscopic examinations according to the Japanese Committee for Clinical Laboratory Standards (JCCLS) [19]. Ten milliliters of urine was centrifuged at 500×g for 5 min, and the supernatant was decanted. The sediment (approximately 0.2 mL) was thoroughly resuspended, and 15 μL was transferred onto a glass slide and covered with an 18×18-mm cover glass. Urinary RBCs were classified as either isomorphic or dysmorphic. Dysmorphic RBCs were further categorized into three levels—minority, moderate, and majority—based on their proportions, as recommended by the JCCLS guidelines. The criteria are outlined in the Supplemental Data Table S3.
Quantification of UF-5000 scattergram data
UF-5000 employs flow cytometry with a blue semiconductor laser to detect urine sediment components using forward-scattered light, side-scattered light, side fluorescence, and depolarized side-scattered light. Urinary RBCs were identified on scattergrams, in which using depolarized side-scatter (X-axis, reflecting birefringence) is plotted against forward scatter (Y-axis, reflecting particle size). Representative microscopic and scattergram images for isomorphic and dysmorphic hematuria are shown in Supplemental Data Fig. S1.
The UF-5000 RBC-Info feature categorizes RBCs as “Isomorphic,” “Mixed,” or “Dysmorphic” based on RBC-P70Fsc (forward scatter, Fsc) and RBC-Fsc-DW (distribution width, DW). RBC-P70Fsc is defined as the 70th percentile of the forward-scatter signal intensity of RBC events, and RBC-Fsc-DW as the distribution width of the RBC forward-scatter signal. Additional parameters included the counts of small, large, lysed, and non-lysed RBCs, which are automatically categorized and output by the UF-5000 analyzer based on internal scattergram analysis algorithms, using manufacturer-defined thresholds. Various ratios of each RBC subtype (e.g., small RBC/total RBC ratio) were manually calculated from the absolute counts provided by the UF-5000 instrument.
Scattergram images were imported into ImageJ 1.54 (NIH, Bethesda, MD, USA) for feature extraction, including minimum and maximum values, mean, standard deviation, median, skewness, and kurtosis for both axes. Regression equation parameters (slope, intercept, and coefficient of determination) were also computed.
ML model development
To construct a predictive model for RBC morphology classification, we used DataRobot (v9.1, https://www.datarobot.com/), an automated ML platform. The platform explored multiple algorithm families, including logistic regression and elastic net, random forests, gradient boosting (LightGBM), XGBoost, support vector machines, k-nearest neighbors, and shallow neural networks. Model selection was based on the logarithmic loss (LogLoss) metric, which evaluates the classification probability accuracy. Cross-validation was performed automatically in DataRobot using a five-fold strategy to assess generalizability and to avoid overfitting. Hyperparameters were tuned internally using the built-in optimization routines of DataRobot, which apply techniques such as grid search or Bayesian optimization, depending on the model type and data characteristics.
Statistical analysis
To facilitate comparisons between UF-5000 and manual microscopy results, urinary RBC counts were converted using the formula particles (p)/high-power field=0.18×p/µL. Statistical analyses were conducted using EZR (Easy R, v1.68) [20]. Cochran’s Q test was used to assess the differences in concordance rates across more than two classifiers, and post-hoc pairwise comparisons were performed using McNemar’s test (exact version when discordant counts were small), with Holm–Bonferroni adjustment as applicable. For two-classifier comparisons of concordance rates, McNemar’s test alone was used. The Mann–Whitney U-test was used to compare variables among groups. ROC curve analysis was performed using StatFlex (v6.0) to evaluate the diagnostic performance of the UF-5000 parameters in distinguishing RBC morphology. Statistical significance was set to P<0.05.
RESULTS
The small RBC/total RBC ratio as a key indicator
The small RBC/total RBC ratio showed the highest accuracy, achieving an area under the ROC curve (AUROC) of 0.97 (Table 1). Fig. 2 illustrates ROC curves and cutoff values for the small RBC/total RBC ratio, RBC-P70-Fsc, and RBC-Fsc-DW. Compared with RBC-Info, the small RBC/total RBC ratio demonstrated superior accuracy in RBC morphology classification in terms of the concordance rate (94.1% vs. 88.1%), AUROC (0.97 vs. 0.88), sensitivity (93.2% vs. 86.3%), specificity (94.6% vs. 89.3%), positive predictive value (PPV; 91.9% vs. 84.0%), negative predictive value (NPV; 95.5% vs. 90.9%), positive likelihood ratio (PLR; 17.39 vs. 8.06), and negative likelihood ratio (NLR; 0.07 vs. 0.15) (Table 2).
Scattergram data analysis and RBC size parameters
Supplemental Data Table S4 presents a comparison of variables derived from the quantification of UF-5000 scattergram data between samples with dysmorphic and isomorphic RBCs. X-axis skewness and kurtosis (1.50 vs. 1.10, P=0.005; 6.10 vs. 3.15, P=0.005, respectively) and Y-axis skewness (0.10 vs. –0.80, P<0.001) were significantly greater in the dysmorphic RBC group than in the isomorphic RBC group. Conversely, the minimum and maximum values, mean, median, and kurtosis of the Y-axis (17.00 vs. 19.00, 174.00 vs. 186.50, 85.00 vs. 125.15, 85.50 vs. 135.00, and –0.60 vs. 0.25, respectively) and the intercept of the regression equation (80.80 vs. 126.25) were lower in the dysmorphic RBC group than in the isomorphic RBC group, with P<0.001 for all comparisons.
High concordance rate of the ML model
To improve RBC morphology classification further, we developed an ML model using the DataRobot. The model was trained using 20 quantitative features extracted via scattergram analysis (Supplemental Data Table S4). These features reflect the geometric and statistical properties of RBC distribution patterns across the scattergram axes, including conventional histogram descriptors (e.g., minimum, maximum, and median) and higher-order statistical features, such as skewness and kurtosis along the X- and Y-axes. No clinical or demographic variables (Table 1) were included in the model input. During model generation, 70 candidate models were automatically generated. The model labeled “Light Gradient Boosting on ElasticNet Predictions” achieved the best performance, with a LogLoss of 0.207 (Supplemental Data Fig. S2A), and was selected for further evaluations. Feature impact analysis showed that the Y-axis median (Y-median) was the most influential variable, followed by Y-skewness and Y-minimum, suggesting that the vertical distribution patterns of RBCs in the scattergram were key determinants in the ML-based classification (Supplemental Data Fig. S2B). When validated against manual microscopic examination, the ML model achieved a concordance rate of 93.9%, sensitivity of 94.9%, specificity of 93.3%, and AUROC of 0.95.
To assess its performance further, the ML model was tested on an independent dataset comprising 1,093 samples with both manual microscopic examination results and RBC-Info data. Among RBC-Info outputs labeled as “Mixed” or “Dysmorphic”, we reclassified all the “Mixed” and “Dysmorphic” cases as “Dysmorphic” because 88.6% of the samples were classified as “Dysmorphic” by microscopy. Notably, “Mixed” is considered suggestive of dysmorphic RBCs, and manual confirmation is generally recommended. Although this reclassification may have introduced interpretive bias, it was conducted to reflect predominant manual findings and maintain consistency in the subgroup analysis. The model demonstrated improved classification accuracy compared with RBC-Info, achieving a concordance rate of 95.2% vs. 92.1% and an AUROC of 0.95 (0.94–0.97) vs. 0.92 (0.91–0.94) (Table 3). Although the ML model had lower sensitivity (94.4% [92.0–96.3]) than RBC-Info (95.0% [92.7–96.7]), other parameters were improved, with a specificity of 95.9% (94.0–97.4) vs. 89.7% (87.0–92.0), PPV of 95.2% (92.9–96.9) vs. 88.6% (85.6–91.2), and PLR of 23.29 (15.73–34.48) vs. 9.22 (7.26–11.70).
Effects of combining the ML model with RBC-Info on RBC classification performance
To determine whether integrating RBC-Info with the ML model would improve classification accuracy, we conducted a simulation using DataRobot. When evaluated against manual microscopic examination, logical OR/AND combinations of the ML model and RBC-Info outperformed RBC-Info alone (concordance rates: 92.7% and 94.6%; AUROC: 0.93 [0.92–0.95] and 0.94 [0.93–0.96], respectively) (Table 3).
Model performance across RBC morphology subgroups
To evaluate model performance based on the degree of dysmorphic RBC representation, we stratified the cases into four subgroups based on the proportion of urinary RBCs in manual microscopy: isomorphic, dysmorphic-minority, dysmorphic-moderate, and dysmorphic-majority. As shown in Fig. 3A, the concordance rates between the ML model and manual classification were 95.9%, 88.8%, 98.9%, and 93.3% in the isomorphic (568/592), dysmorphic-minority (190/214), dysmorphic-moderate (269/272), and dysmorphic-majority groups (14/15), respectively.
Although the overall model performance was high, the lower concordance in the dysmorphic-minority group suggests a need for further improvement of the system for samples with a small proportion of dysmorphic RBCs. A comparative analysis of scattergram-derived parameters between the concordant and discordant groups classified by the ML model relative to manual microscopic examination is provided in Supplemental Data Table S5. To explore the underlying reasons for misclassification, we analyzed the key features used by the model (i.e., Y-median, Y-skewness, and Y-minimum) in matched versus mismatched samples. As shown in Fig. 3B–D, mismatched cases in both the isomorphic and dysmorphic-minority groups displayed significantly different values from matched cases, with the mismatched dysmorphic-minority cases often showing feature profiles closer to those of isomorphic RBCs. In addition, representative discordant cases (Supplemental Data Fig. S3) either had a very low number of dysmorphic RBCs or exhibited overlapping morphological features with isomorphic RBCs, complicating classification.
DISCUSSION
Current assessment of urinary RBC morphology is limited by variability and interobserver differences inherent to manual microscopy. In this study, we identified three complementary solutions that improve classification accuracy: the small RBC/total RBC ratio derived from UF-5000 outputs, a set of novel scattergram-derived features, and ML algorithms built on these features. The results revealed that, together, these approaches enhance the accuracy of urinary RBC classification and may help differentiate between glomerular and non-glomerular hematuria.
First, the small RBC/total RBC ratio was the most effective parameter for distinguishing RBC morphology. Compared with RBC-Info, this ratio significantly improved classification performance across multiple metrics, including the concordance rate, sensitivity, specificity, PPV, NPV, PLR, and NLR (Table 2). These results suggest that the small RBC/total RBC ratio is a promising alternative to conventional RBC-Info for automated RBC morphological analysis using the UF-5000 analyzer. This finding aligns with those in previous studies [12, 13], reinforcing its validity as a reliable diagnostic parameter.
Second, this study represents the first quantitative analysis of UF-5000 scattergram data. Our analysis revealed that several Y-axis features, which correlated with particle size, were significantly smaller in the dysmorphic RBC group than in the isomorphic RBC group (Supplemental Data Table S3). Notably, the negative skewness along the Y-axis suggests that the histogram peak shifts toward lower values, indicating a higher concentration of smaller RBCs. Additionally, lower kurtosis in the Y-axis implies a flatter histogram peak, reflecting greater heterogeneity in RBC morphology. These findings suggest that RBC size plays a critical role in differentiating dysmorphic RBCs from isomorphic RBCs. In clinical practice, scattergram and histogram data obtained from the UF-5000 analyzer may serve as valuable adjuncts to manual microscopic examination for urinary RBC classification.
Finally, the ML model developed using DataRobot achieved a high concordance rate (93.9%) and AUROC (0.95) compared with manual microscopy. Validation using an independent dataset confirmed the model’s robustness, with a concordance rate of 95.2% and an AUROC of 0.95 (Table 3). Importantly, the ML model did not include the small RBC/total RBC ratio as an input, although this parameter has been validated as a useful screening indicator [10]. The improved performance of the ML model was achieved solely via the scattergram-derived geometric and statistical features. The combined use of both methods did not improve the overall concordance. As shown in Table 3, logical combinations, such as RBC-Info AND the ML model, improved specificity (97.1%) but decreased sensitivity (91.6%), resulting in a slightly lower concordance (94.6%) than that achieved with the ML model alone. This outcome may reflect conflicting decision boundaries between the two classifiers, and future studies should explore more sophisticated ensemble approaches. Nonetheless, these findings suggest that combining traditional parameters with ML approaches can optimize RBC morphology classification and improve the clinical utility of the UF-5000.
The reduced concordance rate in cases with a minority proportion of dysmorphic RBCs (Fig. 3) indicates that further refinement of the system is needed to enhance its accuracy for this subgroup. As shown in representative cases in Supplemental Data Fig. S3, some discordant samples exhibited either a very low number of dysmorphic RBCs or a mixture of dysmorphic and isomorphic morphological features on scattergram analysis, which likely contributed to the classification difficulty. Although quantitative parameters, such as Y-median and Y-skewness, were useful overall, these features alone were insufficient to reliably distinguish minority-type dysmorphic RBCs from isomorphic RBCs (Fig. 3B–D and Supplemental Data Table S5). To overcome this limitation, future efforts should focus on developing a dedicated secondary classifier tailored to detect subtle glomerular features or identify novel features better suited for borderline cases.
Notably, the ML model developed in this study was based on UF-5000-specific output parameters. However, the model is not generalizable to data from other urine analyzers. Nevertheless, the analytical methodology itself remains broadly applicable and may be adapted to other platforms with similar data structures, pending appropriate calibration and validation.
This study had two key limitations. First, it was conducted as a single-center, retrospective analysis, which limits the generalizability of the proposed ML model. Further validation across multiple institutions and in prospective settings is warranted. Second, the correlation between RBC morphological assessments derived from urine sediment analysis and histological diagnoses was incomplete. Some samples lacked a definitive etiological classification for hematuria. However, as biopsies are invasive and not routinely performed in most patients, microscopic examination remains the most practical method in clinical laboratories. Future studies should evaluate the proposed algorithm using patient data with histological diagnoses to strengthen its clinical applicability.
In conclusion, the scattergram data-based ML model significantly improves the reliability and accuracy of urinary RBC morphological evaluation using the UF-5000. These methods have the potential to standardize urinary RBC classification and serve as valuable tools for diagnosing and monitoring renal and urological diseases. By improving classification, they may reduce unnecessary kidney biopsies, enable earlier differentiation between glomerular and non-glomerular hematuria, and support more standardized, faster decision-making in nephrology and urology practice, thereby linking model outputs more directly to patient management.
XML Download