Machine learning-based prediction of tubulointerstitial lesions in diabetic kidney disease: a multicenter validation study

Abstract

Background

Tubulointerstitial lesions (TILs) play a crucial role in the progression of diabetic kidney disease (DKD). Current clinical prediction methods rely heavily on invasive biopsies and fall short of providing a comprehensive, multidimensional assessment.

Methods

This multicenter study, conducted from 2010 to 2024, involved patients with biopsy-confirmed DKD. We employed logistic regression and the Least Absolute Shrinkage and Selection Operator (LASSO) with 10-fold cross-validation to identify key predictive indicators. A machine learning-based nomogram was developed and model performance is evaluated by discrimination, calibration, and decision curve analysis (DCA).

Results

A total of 337 were divided into training (n = 180), internal validation (n = 78), and external validation (n = 79) cohorts. Multivariate analysis confirmed SCr (OR = 1.02, 95% confidence interval [CI]:1.01–1.03), HDL (OR = 0.68, 95% CI: 0.52–0.85), and severe glomerular hyperplasia (OR = 9.79, 95% CI: 1.41–80.89) as independent predictors. The AUC for the training, internal validation, and external validation cohorts were 0.93, 0.86, and 0.94, respectively, indicating high calibration accuracy (p > 0.05). DCA demonstrated a 75% higher net clinical benefit compared to traditional models at risk thresholds exceeding 18.

Conclusions

In order to avoid repeated renal biopsy puncture, by integrating kidney function, metabolic markers, and pathological features, the model provides more accurate diagnosis and treatment suggestions for predicting the risk of TILs and further deterioration of renal function. This facilitates the early identification of high-risk patients and supports targeted interventions, marking a paradigm shift toward precision nephrology in the management of DKD.

1. Introduction

Diabetic kidney disease (DKD) is the leading cause of chronic kidney disease (CKD) and end-stage kidney disease (ESKD) globally, accounting for approximately 50% of ESKD cases [1,2]. Early intervention can delay DKD progression; however, the insidious onset and lack of specific biomarkers often result in delayed diagnosis. Treatment options for advanced kidney failure remain limited [3]. Consequently, understanding the key drivers of DKD progression and developing early warning tools are essential for improving patient outcomes.

Recent research has highlighted tubulointerstitial lesions (TILs) as the “final common pathway” in the transition from DKD to ESKD [4,5]. The pathological process of TILs involves early tubular epithelial cell hypertrophy and tubular basement membrane thickening, eventually progressing to interstitial fibrosis, inflammatory infiltration, glomerulosclerosis, and ultimately, irreversible loss of kidney function [6,7]. The pathogenesis involves multiple pathways such as mitochondrial dysfunction, oxidative stress, and microcirculation disorders [7]. However, identifying TILs in clinical settings relies heavily on invasive kidney biopsies, which pose significant risks for DKD patients due to advanced age and cardiovascular complications, including risks of bleeding (1% to 7%) and arteriovenous fistula (5% to 10%) [8,9]. Furthermore, international guidelines recommend biopsies only in atypical cases, resulting in up to 30% to 40% of DKD patients remaining undiagnosed [10,11]. This “diagnostic dilemma,” combined with the pathological complexity of TILs (interacting conditions such as anemia, hypertension, and dyslipidemia), exacerbates the challenges in accurate risk assessment [5,12,13].

Given these challenges, the development of predictive models is essential to dynamically understand the progression and overcome bottlenecks in the management of DKD. Traditional statistical methods, like univariate regression, suffer from issues such as collinearity and overfitting, limiting their ability to handle multidimensional clinical-pathological data. Machine learning algorithms, however, address these issues by extracting key predictors through feature compression and regularization optimization [14]. Based on this, this study has innovatively constructed a multi-dimensional TILs risk prediction framework by integrating metabolic markers (high-density lipoprotein [HDL]), kidney function parameters (serum creatinine [SCr]) and pathological features (severe glomerular hyperplasia) for the first time. The study further developed a visual nomogram tool to rapidly stratify individualized risks, provide an evidence-based basis for early targeted interventions, and ultimately promote the precise transformation of DKD management.

2. Materials and methods

2.1. Study design and participants

The multicenter retrospective cohort study was conducted across four medical institutions in the Xinjiang Uygur Autonomous Region, China, from January 2010 to December 2024. A total of 337 patients with pathologically confirmed DKD and comprehensive clinical data were enrolled. All participants met the International Society of Pathology criteria for DKD.

Patients were categorized according to the Tubulointerstitial Injury Index (TII) [15]: No TILs group (TII 0–1 points, interstitial lesion area <10%, n = 68) and TILs group (TII ≥ 1–4 points, interstitial lesion area ≥10%, n = 269). From the main center (People’s Hospital of Xinjiang Uygur Autonomous Region), 258 patients were randomly assigned in a 7:3 ratio to a training set (n = 180) and an internal validation set (n = 78), ensuring balanced baseline characteristics between groups (p > 0.05). An external validation set comprised 79 patients from three independent medical institutions, representing diverse regions and timeframes to validate model generalizability (Figure 1).

Figure 1.

Flow diagram for the training set and validation set. In this multicenter retrospective study, candidate predictors were first identified by univariate logistic regression, refined through least absolute shrinkage and selection operator (LASSO) regression, and subsequently entered into a multivariable logistic model to determine independent risk factors and construct a clinical prediction model. The discriminative and calibration performance of this model was benchmarked against three additional machine-learning algorithms—decision tree (DT), extreme gradient boosting (XGBoost), and support vector machine (SVM). A user-friendly nomogram was then developed to facilitate bedside risk assessment and individualized clinical decision-making. Institution 2–4: the First People’s Hospital of Kashgar Prefecture, the People’s Hospital of Yache County, and the Friendship Hospital of Yili Prefecture.

2.2. Data collection

Demographic characteristics collected included age, sex, body mass index (BMI). Disease characteristics comprised the duration of diabetes, blood pressure level, presence of diabetic retinopathy (DR), and medication history (Renin-angiotensin-aldosterone system inhibitors [RAASi], Sodium-glucose cotransporter-2 inhibitors [SGLT2i], Non-steroidal mineralocorticoid receptor antagonists [nsMRA], Traditional Chinese Medicine [TCM]) at the time of kidney biopsy. Laboratory parameters included routine blood tests, kidney function, and metabolic indexes (lipids, glycosylated hemoglobin [HbA1c]). Inflammatory markers and urinalysis parameters were also recorded. Pathological assessments involved Periodic Acid-Schiff (PAS) staining to determine the degree of mesangial hyperplasia and Masson’s trichrome staining to quantify the extent of TILs according to TII.

To ensure data completeness of at least 80%, multiple imputation using chained equations filled in missing data, provided the missing rate was below 20% (UTP, HbA1c, Uβ₂-MG, TG, TC, HDL and LDL). The missing value of Uβ₂-MG was found to be greater than 20% in the external validation data collection, and it was discarded in the subsequent analysis. Extreme outliers were managed using the two-sided 5% truncated Winsorization method, preserving 95% of the data distribution interval.

2.3. Ethics and criteria

The study protocol was approved by the Ethics Committee of Xinjiang Uygur Autonomous Region People’s Hospital (approval number: KY2025041503) and adhered to the Declaration of Helsinki and the Chinese Code of Ethics for Biomedical Research. Written informed consent was obtained from all participants, and data were anonymized to protect privacy.

2.4. Model development

To develop a robust prediction model and systematically identify the core predictors of diabetic TILs, this study employed Univariate logistic regression (ULR) and Least Absolute Shrinkage and Selection Operator (LASSO) combined with 10-fold cross-validation to screen important features from 22 clinical and pathological candidate variables. This regularization process was followed by Multivariable Logistic Regression (MLR) to determine independent risk factors, and the final model was constructed using the minimum Akaike Information Criterion (AIC), expressed as odds ratios (OR) and 95% confidence intervals (CI). Based on these analyses, risk prediction models were established, and a visual nomogram was constructed.

2.5. Model validation

Model performance was evaluated through discrimination (assessed by the Area Under the Receiver Operating Characteristic [ROC] Curve [AUC]), calibration (Hosmer-Lemeshow test [HL test], p > 0.05), and clinical utility (Decision Curve Analysis [DCA], Clinical Impact Curve [CIC]). Robustness was verified using bootstrap resampling with 1,000 iterations to correct for the risk of model overfitting. The other three institutional cases were used as external queues to validate model performance.

2.6. Statistical analysis

Continuous variables were presented as mean ± standard deviation or median (IQR). Categorical variables were expressed as frequency (%). The sample size was calculated using G*Power 3.1 software with an effect size f² of 0.15, an α level of 0.05, and a β level of 0.2, ensuring statistical power greater than 80%. All statistical analyses were performed using R version 4.4.2 (R Foundation for Statistical Computing) and Jingding Diagnostic Clinical Prediction Model Software (DCPM V6.01, Hefei Tading Information Technology Co., Ltd.). The significance threshold for hypothesis testing was set to a two-sided p < 0.05.

3. Results

3.1. Clinical and pathological characteristics of DKD patients in the training and validation sets

The baseline characteristics, including gender distribution, age, BMI, and duration of diabetes, showed no significant differences among the three cohorts (all p > 0.05). However, the prevalence of hypertension was slightly lower in the external validation set (p = 0.01), and mean arterial pressure (MAP) was slightly higher (p = 0.009). The occurrence of DR and medication history (RAASi, TCM) were consistent across the three cohorts (p > 0.05). In contrast, the use of SGLT2i (p = 0.01) and nsMRA (p = 0.03) was higher in the external validation set, reflecting regional differences in clinical practice (Table 1).

Table 1.

Baseline characteristics of training and validation sets in DKD patients.

Parameter	ALL (N = 337)	Training set (N = 180)	Internal validation set (N = 78)	External validation set (N = 79)	P
Gender, n (%)					0.255
Male	242 (71.81%)	136 (75.56%)	52 (66.67%)	54 (68.35%)
Female	95 (28.19%)	44 (24.44%)	26 (33.33%)	25 (31.65%)
Age (years), median [IQR]	52.00 [45.00–59.00]	52.00 [45.00–58.00]	52.50 [45.00–58.75]	53.00 [48.00–60.50]	0.440
BMI (kg/m²), median [IQR]	26.34 [23.72–28.73]	26.17 [23.98–28.37]	26.09 [23.23–28.37]	27.41 [23.33–30.88]	0.136
Hypertension, n (%)					0.010
No	45 (13.35%)	16 (8.89%)	11 (14.10%)	18 (22.78%)
Yes	292 (86.65%)	164 (91.11%)	67 (85.90%)	61 (77.22%)
SBP (mmHg), mean ± SD	144.37 ± 23.19	145.70 ± 23.35	140.60 ± 23.38	145.08 ± 22.55	0.257
DBP (mmHg), median [IQR]	83.00 [74.00–91.00]	84.50 [74.00–92.25]	79.50 [74.00–88.00]	85.00 [77.00–92.00]	0.087
MAP (mmHg), mean ± SD	104.63 ± 16.05	104.61 ± 15.19	100.73 ± 14.65	108.54 ± 18.35	0.009
DR, n (%)					0.397
No	168 (49.85%)	95 (52.78%)	34 (43.59%)	39 (49.37%)
Yes	169 (50.15%)	85 (47.22%)	44 (56.41%)	40 (50.63%)
RAASi, n (%)					0.440
No	111 (32.94%)	58 (32.22%)	30 (38.46%)	23 (29.11%)
Yes	226 (67.06%)	122 (67.78%)	48 (61.54%)	56 (70.89%)
SGLT2i, n (%)					0.001
No	231 (68.55%)	138 (76.67%)	51 (65.38%)	42 (53.16%)
Yes	106 (31.45%)	42 (23.33%)	27 (34.62%)	37 (46.84%)
nsMRA, n (%)					0.007
No	250 (74.18%)	142 (78.89%)	60 (76.92%)	48 (60.76%)
Yes	87 (25.82%)	38 (21.11%)	18 (23.08%)	31 (39.24%)
TCM, n (%)					0.351
No	327 (97.03%)	175 (97.22%)	74 (94.87%)	78 (98.73%)
Yes	10 (2.97%)	5 (2.78%)	4 (5.13%)	1 (1.27%)
Diabetes duration (months) median [IQR]	116.00 [60.00–168.00]	120.00 [60.00–180.00]	109.00 [55.76–165.00]	108.00 [45.00–156.00]	0.641

Abbreviations: BMI: Body mass index; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; MAP: Mean arterial pressure; DR: diabetic retinopathy.

RAASi: Renin-angiotensin-aldosterone system inhibitors; SGLT2i: Sodium-glucose cotransporter-2 inhibitors; nsMRA: Non-steroidal mineralocorticoid receptor antagonists; TCM: Traditional Chinese Medicine; IQR: Interquartile range.

*Note: Diabetes duration was calculated from the date of diabetic diagnosis to the kidney biopsy.

There were no significant differences in eGFR, SCr and blood urea nitrogen (BUN) levels among the three groups. However, the metabolic indexes in the external validation set exhibited distinct pattern: levels of HDL (p < 0.001), Low-Density Lipoprotein (LDL, p = 0.020) and TG (p < 0.001) showed significant differences from those in the main center cohort, potentially associated with regional dietary habits or lipid-lowering treatments. Hematological parameters, inflammatory markers and Urinalysis were consistent among the groups (all p > 0.05). Data on urinary beta-2 microglobulin (Uβ₂-MG) was not available for the external validation set (Table 2).

Table 2.

Clinical and pathological parameters of training and validation sets in DKD patients.

Parameter	ALL (N = 337)	Training set (N = 180)	Internal validation set (N = 78)	External validation set (N = 79)	P
Kidney function, median [IQR]
eGFR (mL/min/1.73 m2)	55.76 [25.83–90.70]	55.58 [25.62–91.05]	54.95 [24.54–95.95]	57.60 [28.49–87.56]	0.959
SCr (µmol/L)	127.40 [88.23–222.50]	128.85 [89.47–233.80]	137.10 [83.72–228.65]	117.00 [87.85–196.50]	0.606
BUN (mmol/L)	9.10 [6.62–13.61]	8.89 [6.64–12.68]	9.84 [6.04–14.32]	9.13 [6.76–12.93]	0.935
UA (µmol/L)	364.00 [301.50–416.40]	371.82 [310.78–418.58]	368.56 [324.06–428.38]	337.00 [274.82–399.75]	0.025
Urinalysis
UTP (g/24h), median [IQR]	4.48 [2.44–6.44]	4.59 [2.44–6.25]	4.38 [2.20–6.53]	4.24 [3.20–6.44]	0.775
Uβ₂-MG (mg/L) median [IQR]	/	2.27 [0.41–6.96]	3.22 [0.23–6.96]	/	/
Hematuria*, n (%)					0.638
No	103 (30.56%)	56 (31.11%)	26 (33.33%)	21 (26.58%)
Yes	234 (69.44%)	124 (68.89%)	52 (66.67%)	58 (73.42%)
Metabolic profile, median [IQR]
TG (mmol/L)	2.78 [1.72–4.40]	3.54 [1.91–4.68]	3.15 [1.85–4.14]	2.06 [1.42–2.74]	<0.001
TC (mmol/L)	4.41 [2.96–5.50]	4.02 [2.19–4.97]	4.41 [2.89–5.51]	4.76 [3.87–6.40]	<0.001
HDL (mmol/L)	1.21 [0.92–2.33]	1.29 [0.94–2.62]	1.46 [0.98–3.55]	1.00 [0.82–1.24]	<0.001
LDL (mmol/L)	3.24 [2.31–4.66]	3.42 [2.38–4.72]	3.55 [2.15–6.03]	2.66 [2.21–3.63]	0.020
HbA1c (%),	7.60 [6.50–9.00]	7.35 [6.50–8.70]	7.90 [6.32–9.00]	8.30 [6.90–9.45]	0.074
Hematology, median [IQR]
RBC (1012/L)	3.99 [3.31–4.55]	4.03 [3.34–4.56]	3.86 [3.09–4.43]	3.89 [3.45–4.54]	0.306
HB (g/L)	117.53 [96.00–137.00]	118.00 [97.00–138.25]	116.00 [91.50–133.00]	115.00 [100.00–136.00]	0.578
Alb (g/L)	30.60 [25.30–36.30]	30.05 [25.38–36.64]	30.75 [24.66–36.38]	31.70 [26.84–35.60]	0.911
CRP (mg/L)	2.56 [1.10–5.40]	2.50 [1.14–5.20]	2.50 [1.10–5.68]	3.11 [0.88–5.60]	0.950
WBC (109/L)	7.51 [6.01–9.28]	7.50 [6.04–9.51]	7.62 [6.03–8.92]	7.48 [5.88–9.21]	0.681
PLT (109/L)	252.00 [203.00–303.00]	249.50 [194.00–304.50]	250.00 [217.25–302.50]	257.00 [201.00–303.50]	0.852
Kidney pathology
Mild hyperplasia† (%), median [IQR]	43% [22–72]	44% [21–75]	45% [21–77]	38% [24–62]	0.691
Severe hyperplasia† (%), median [IQR]	57% [27–78]	56% [25–79]	54% [23–80]	60% [36–76]	0.782
GBM†, n (%)					0.001
No	49 (14.54%)	34 (18.89%)	14 (17.95%)	1 (1.27%)
Yes	288 (85.46%)	146 (81.11%)	64 (82.05%)	78 (98.73%)
K-W nodules†, n (%)					0.018
No	140 (41.54%)	83 (46.11%)	35 (44.87%)	22 (27.85%)
Yes	197 (58.46%)	97 (53.89%)	43 (55.13%)	57 (72.15%)
Hypertensive damage†, n (%)					<0.001
No	222 (65.88%)	146 (81.11%)	62 (79.49%)	14 (17.72%)
Yes	115 (34.12%)	34 (18.89%)	16 (20.51%)	65 (82.28%)

Abbreviations: eGFR: Estimated Glomerular Filtration Rate; UA: Uric Acid; BUN: Blood Urea Nitrogen; SCr: Serum creatinine; UTP: Urinary Total Protein; Uβ₂ MG: Urinary Beta-2 Microglobulin; TG: Triglycerides; TC: Total Cholesterol; HDL: High-Density Lipoprotein; LDL: Low-Density Lipoprotein; HbA1c. Glycosylated Hemoglobin A1C; RBC: Red Blood Cells; HB: Hemoglobin; Alb: Albumin; CRP: C-Reactive Protein; WBC: White Blood Cells; PLT: Platelets; GBM: Glomerular Basement Membrane; K-W nodules: Kimmelstiel-Wilson. Nodules; IQR: Interquartile range.

^*Note: Hematuria defined as ≥5 red blood cells per high-power field. / : Meaning no result.

^† Pathological definitions.

Mild hyperplasia: Under PAS staining, mesangial proliferation in the glomerulus is less than 50%.

Severe hyperplasia: Mesangial proliferation in the glomerulus exceeds 50%.

GBM: Glomerular basement membrane thickening with diagnostic thresholds: average male thickness >430 nm, average female thickness >395 nm.

K-W nodules: Nodular glomerulosclerosis observed on PAS staining.

Hypertensive damage: Arteriolar hyalinosis, ischemic glomerulosclerosis, or interstitial fibrosis.

Pathologically, the incidence of glomerular basement membrane (GBM) thickening (98.73% vs. 82.05%), Kimmelstiel-Wilson (K-W) nodules (72.15% vs. 55.13%), and hypertensive renal impairment (82.28% vs. 20.51%) was markedly higher in the external validation set compared to the main center group. With the awareness of the disease and the importance of health, diabetic patients are more likely to receive kidney biopsy than before, and the diagnosis rate of hypertensive kidney damage is significantly higher than usual. Conversely, the distribution of mild and severe mesangial hyperplasia was comparable across the groups (all p > 0.05), signifying more advanced pathological damage in the external cohort. Notably, the training set demonstrated high consistency with the internal validation set regarding demographic, biochemical, and pathological features (Table S1 of Supplementary Materials), affirming the validity of the randomization process. The population from external validation differed from the internal data in terms of disease spectrum, diagnosis and treatment methods, regional differences, and patient characteristics. This difference can evaluate the generalization ability of the model, predict its performance in the real world, and ensure the reliability and security of the model.

Despite regional differences in diagnosis and treatment within the external validation set, the consistency with the main center cohort in core indicators such as renal function and key pathological features supports the model’s generalizability.

3.2. Screening of predictors

Twenty-two indicators were screened out by ULR (Select P threshold 0.1, Table S2 of Supplementary Materials), and continued to be screened by LASSO regression which aimed at reducing model complexity, initially retained 12 potential predictors by minimizing cross-validation error (λ.min = 0.012). Using the one standard error criterion (λ.1se = 0.051), redundant variables were further compressed to identify seven key features: eGFR, SCr, BUN, HB, TG, HDL, and severe mesangial hyperplasia in the glomeruli (Figure 2). The selection of lambda_1se corresponding indicators into the model reduces the complexity of the model and reduces the risk of overfitting. Stronger penalties are applied to the parameters, which brings the coefficients closer to zero, which enhances the robustness of the model, reduces the variance, and avoids local optimum. This method effectively addresses multicollinearity issues in high-dimensional data, reducing model variables by 41.67% from 12 to 7.

Figure 2.

Diagram of LASSO coefficients for 22 clinical features. [A] lasso coefficient path: the regression coefficient varies with log(λ). [B] lasso regularization path: the mean square deviation of the model varies with log(λ). [C] lambda.min corresponds to 12 factors. [D] lambda.1se yields 7 predictors.

3.3. Establishment of predictive models and plausibility analysis

MLR identified five independent predictors: SCr, HDL, severe mesangial hyperplasia (mesangial expansion area >50%), TG, and HB (Figure 3). Elevated SCr and TG increased the risk of TILs, while higher HDL levels had a protective effect. Severe mesangial hyperplasia was a significant pathological driver of TILs (OR > 9). Although TG (OR = 1.38, p = 0.091) and HB (OR = 0.98, p = 0.088) were not statistically significant, their effect directions suggest potential pro-fibrotic roles for lipid metabolism disorders and a protective role for anemia correction, necessitating further validation with larger samples.

Figure 3.

Forest map of multivariate regression analysis three independent risk factors were screened out: SCr, HDL, severe mesangial hyperplasia.

3.4. Validation and performance evaluation of predictive models

1. Discrimination analysis

In the training set, the model exhibited an AUC of 0.93 (95% CI: 0.89–0.97), with a sensitivity of 89.7% and specificity of 83.7%. For the internal validation set, the AUC was 0.86, with a sensitivity of 94.4% and specificity of 68.3% (Figure 4A). Bootstrap resampling (1,000 iterations) resulted in a corrected AUC of 0.91 for the training set and 0.86 for the internal validation set (Figure 4C and D). The external validation cohort achieved an AUC of 0.94 (95% CI: 0.86–1.00), with a sensitivity of 90.9% and specificity of 83.8%. Bootstrap resampling for the external validation set confirmed an AUC of 0.91 (95% CI: 0.84–0.98), affirming the model’s excellent generalizability (Figure 4B and E).

Figure 4.

ROC curves for the training set and validation sets. [A] ROC curves for training and internal validation sets; [B] ROC curves for the external validation set; [C, D, E] bootstrap analysis for training, internal validation set, and external validation sets, respectively.

1. Calibration analysis

Calibration analysis demonstrated close alignment between predicted probabilities and observed risks. The HL test revealed high consistency between predicted and actual risks, with χ² = 3.33 for the training set (Figure 5A), χ² = 13.75 for the internal validation set (Figure 5B), and χ² = 4.59 for the external validation set (Figure 5C), indicating no significant statistical bias (degrees of freedom [df] = 8). Calibration curves further visualized prediction accuracy, particularly in the medium to high-risk range (20%–80% prediction probability), where observed risk overlapped nearly perfectly with predicted values.

Figure 5.

Calibration curves for predictive probability of TILs. [A] training set. [B] internal validation set. [C] external validation set. All P-values in HL tests were greater than 0.05, indicating no statistical deviation between predicted and observed values.

1. Clinical application

DCA was conducted to quantify the net benefit of the nomogram in predicting TILs risk. In the training set, when the risk threshold probability ranged between 18%–80%, the net benefit of interventions guided by the model was significantly higher than that of “full intervention” or “no intervention” strategies, with the maximum net benefit increasing by 75% (Figure 6A). Consistent results were observed for the internal and external validation sets, with stable net benefit rates within the 15%–80% risk threshold range (Figure 6B and C), confirming the model’s broad applicability across different geographies and populations.

Figure 6.

DCA and CIC for the nomogram. [A, B, and C] DCA for the training set, internal validation set, and external validation set, respectively, illustrating increased clinical net benefit compared to “all-occurring” or “non-occurring” scenarios. [D, E, F] CIC for the training set, internal validation set, and external validation set, respectively.

Further predictions of risk alignment with actual case distribution were visualized through the CIC. In the training set, when the risk threshold was 0.1, the true positive rate was 90.34% (TP+FP = 880, n = 795). Raising the risk threshold to 0.4 increased the true positive ratio to 95.18% (TP+FP = 830, n = 790), indicating high alignment between model predictions and actual cases (Figure 6D). Similar trends were noted in both the internal validation set (Figure 6E) and external validation set (Figure 6F), further validating the model’s clinical decision-making performance.

These results demonstrate that the nomogram provides significant net benefit across a wide range of clinical risk thresholds (>15%) and exhibits high accuracy in screening high-risk patients (true positive rate > 95%).

3.5. Multi-model performance evaluation and clinical decision-making

In order to evaluate the possibility of the best model for univariate-LASSO-multivariate logistic regression analysis, the AUC, calibration and net benefit, specificity, sensitivity, recall, F1, positive predictive value and negative predictive value were evaluated compared with various algorithms such as Decision Tree (DT), eXtreme Gradient Boosting (XGB), Support Vector Machine (SVM). XGB performed better in many aspects (Figure 7).

Figure 7.

Multi-model performance evaluation training set: [A] AUC comparison of models [B] Comparison of calibration curves [C] Comparison of net benefit [D] specificity, sensitivity, recall, F1, positive predictive value and negative predictive value of models validation set: [E] AUC comparison of models [F] Comparison of calibration curves [G] Comparison of net benefit [H] specificity, sensitivity, recall, F1, positive predictive value and negative predictive value of models.

In order to further strengthen the interpretive nature of the model and the preliminary translation of clinical decision-making, we constructed a nomogram to predict the risk of diabetic TILs using these three independent predictors. The model quantifies the contribution of each variable to TILs risk, allowing individualized risk assessment (Figure 8A). In patients whose nomogram-derived risk score exceeds 45 points or whose predicted probability of progression surpasses 0.2, prompt initiation or up-titration of SGLT2 inhibitors is advised, coupled with stringent monitoring of renal function. Consistent with the Clinical Practice Guideline for the Detection and Management of DKD [16], repeat renal biopsy should be considered when any of the following are encountered: gross hematuria, rapidly progressive renal failure, or an accelerated decline in eGFR—defined as a reduction greater than 30% within a 3-month interval.

Figure 8.

[A] Nomogram predicting the probability of TILs; model plausibility analysis: [B] training set, [C] validation set; group 0: no obvious TILs; group 1: TILs. [D] Nomo-ROC curve evaluation.

The prediction model’s rationality was analyzed based on the Nomoscore, revealing significant differences between groups with and without TILs in both the training and validation sets (Figure 8B and C). The model’s plausibility was further supported by an AUC of 0.89, with specificity and sensitivity of 0.84 and 0.82, respectively, indicating strong model performance (Figure 8D).

4. Discussion

The progression of DKD is influenced by a combination of pathophysiological mechanisms [17–19]. Emerging evidence underscores the critical role of diabetic TILs in driving DKD progression to ESKD [7,20]. Despite significant advancements in the understanding of these pathological mechanisms, effective clinical prediction tools remain lacking. Our study addresses this gap by integrating machine learning algorithms with multidimensional clinical data to develop and validate a novel nomogram model for predicting the risk of diabetic TILs. This model highlights the interplay between renal function, metabolic disorders, and pathological features, serving as a valuable tool for clinical risk stratification.

The study identified several independent risk factors for TILs, including severe mesangial hyperplasia, SCr, and HDL. Among them, HDL demonstrated significant protective effects. Elevated SCr not only indicates impaired glomerular filtration but also contributes to interstitial fibrosis by activating renal tubular epithelial apoptosis pathways, such as the TGF-β1/Smad3 axis [21]. Consistent with previous research, higher SCr levels were associated with more pronounced TILs, confirming SCr as a robust predictor of kidney function deterioration.

Our predictive model revealed that patients with significant glomerular hyperplasia experienced more severe tubulointerstitial damage. Glomerular hyperplasia can directly harm renal tubular epithelial cells, impairing their absorptive and secretory functions [22]. Severe mesangial hyperplasia was significantly correlated with K-W nodules and GBM thickening, revealing a spatiotemporal relationship between glomerulosclerosis and TILs. Mechanistically, mesangial expansion >50% induces tubular epithelial-mesenchymal transition through the Angiotensin II/Angiotensin type 1 receptor axis, while GBM thickening exacerbates local hypoxia and inflammatory responses [23,24].

Lipid metabolism is critical for renal health, as proper lipid metabolism and mitochondrial function in renal tubular cells prevent lipotoxicity-induced mitochondrial dysfunction and endoplasmic reticulum stress, which can lead to tubulointerstitial fibrosis [25,26]. Our study found higher levels of TG, LDL, and TC in patients with significant TILs, while HDL served as a protective factor. Although direct literature linking HDL to TILs is limited, HDL’s anti-inflammatory, antioxidant, and endothelial function-enhancing properties suggest a protective role against renal tubulointerstitial damage [27,28].

Hemoglobin levels emerged as protective, with higher RBC counts and hemoglobin levels observed in patients without significant TILs. Renal anemia often accompanying tubulointerstitial injury, results from abnormal iron metabolism and decreased erythropoietin production [29,30]. Lower hemoglobin levels in patients with severe interstitial fibrosis and tubular atrophy negatively affect long-term outcomes in DKD, highlighting TILs as a predictor of anemia [31,32]. Correcting anemia presents a potential target for delaying TILs progression.

Our study also emphasized clinical management considerations, such as the higher levels of serum albumin and UTP in patients with significant TILs, which can exacerbate TILs by directly damaging proximal tubular epithelial cells. Proteinuria can also trigger the release of inflammatory mediators through lysosomal pathways, forming a “glomerulo-tubular crosstalk” feedback loop [22,33]. Additionally, the combined impact of hypertension and diabetes further damages the renal tubulointerstitium, with over 91% of patients with significant TILs having a history of hypertension and prolonged diabetes, triggering inflammatory cascades [34–36]. Clinically, strict blood pressure and blood glucose control are crucial in preventing and delaying TILs.

Compared with traditional indicators (such as eGFR, serum cystatin C and proteinuria), the model based on machine learning can integrate multi-dimensional indicators, capture nonlinear relationships, comprehensively assess kidney risk, and improve the accuracy of prediction. At the same time, the prediction results and risk trends are visually displayed through charts and other methods, which is easy for doctors and patients to understand [37,38]. The machine learning-based KidneyIntelX model (concentrations of three biomarkers and their ratios and seven clinical variables) has attracted much attention. In contrast, the risk stratification, dynamic evaluation and clinical application value of our model are weak [39,40]. However, our model does not predict the risk of deterioration or failure of renal function in general, but directly predicts TIL, and includes pathological indicators, which has high prediction accuracy. Also, this study has several limitations. Firstly, as a multicenter retrospective study, the sample size was relatively small due to limited clinical data availability and constraints in medical records. Consequently, some potential risk factors were not included in the analysis. Additionally, the restricted number of available cases made it unfeasible to stratify and analyze the differences between early (TII 1-2) and late (TII 3-4) stages of TILs in detail. Future research should focus on integrating advanced techniques such as single-cell sequencing and urine exosome miRNA detection to enhance the model’s prediction performance. Combining these methodologies could provide deeper insights into the molecular mechanisms underlying TILs and improve the accuracy of risk prediction models for DKD.

5. Conclusion

By integrating machine learning algorithms and multidimensional clinical data, this study pioneered the construction and validation of a predictive model for diabetic TILs risk. The model demonstrated excellent predictive performance and calibration (AUC > 0.85), with SCr, lipid metabolism (HDL) and glomerular pathology (severe mesangial hyperplasia) as core indicators. Its clinical application can facilitate the early identification of high-risk patients, providing a theoretical basis for a comprehensive management strategy encompassing blood pressure, proteinuria, anemia, and metabolic control. Future research should combine metabolomics and artificial intelligence-based pathological analysis to further explore the glomerulo-tubule interaction mechanism, transforming DKD prevention and treatment from empirical intervention to targeted therapy.

Funding Statement

This work was supported by the third batch of the “2 + 5 Key Talent Program – Tianshan Talents” high‑level medical and health talent initiative, through the project “Research on the mechanism of secondary kidney disease and establishment of a follow‑up system for kidney transplantation” (Grant No. TSYC202401A013).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.