A SuperLearner approach to predicting diabetic kidney disease upon the initial diagnosis of T2DM in hospital

doi:10.21203/rs.3.rs-3549169/v1

Download PDF

Research Article

A SuperLearner approach to predicting diabetic kidney disease upon the initial diagnosis of T2DM in hospital

https://doi.org/10.21203/rs.3.rs-3549169/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Diabetic kidney disease (DKD) is a serious complication of diabetic mellitus (DM) and is asymptomatic until the advanced stage. We developed and validated a predictive model for DKD in patients with an initial diagnosis of type 2 Diabetes Mellitus (T2DM) based on real-world data.

Methods

Data from 3,291 patients (1740 men, 1551 women) newly diagnosed with T2DM at Ningbo Municipal Hospital of Traditional Chinese Medicine (2011–2023) were retrospectively collected. Data were randomly divided into training and validation cohorts. Forty-six readily available medical characteristics at initial diagnosis of T2DM from the electronic medical record (EMR) were used to develop prediction models based on linear, non-linear, and SuperLearner approaches. Model performance was evaluated using the area under the curve (AUC). Shapley additive explanation (SHAP) was used to interpret the best-performing models.

Results

Of the 3291 participants, 563 (17.1%) were diagnosed with DKD during a median follow-up of 2.53 years. The SuperLearner model had the highest AUC (0.7138, 95% CI [0.673, 0.7546]) for the holdout internal validation set in predicting any DKD stage. Top-ranked features were WBC_Cnt*, Neut_Cnt, Hct, and Hb. High WBC_Cnt, low Neut_Cnt, high Hct, and low Hb levels were associated with an increased risk of developing DKD.

Conclusions

We developed and validated a DKD risk prediction model for patients with newly diagnosed T2DM. Using routinely available clinical measurements, the SuperLearner model can predict DKD during hospital visits. Prediction accuracy and SHAP-based model interpretability may help improve early detection, targeted interventions, and prognosis of patients with DM.

Type 2 diabetes

Diabetic kidney disease

Real-world data

Machine learning

Model interpretability

Risk estimation

Diabetic kidney disease (DKD) is a common microvascular complication of diabetes mellitus (DM) and has been recognized as the leading cause of end-stage renal disease (ESRD)^[1]. It is always asymptomatic until it develops into an advanced stage^[2]. Although DKD can be slowed down through medication at an early stage, lifestyle changes, and careful management of blood sugar levels^[3], its advanced stages are often irreversible and may result in ESRD, which typically requires dialysis or kidney transplantation^[4].

With approximately 30% of patients with DM developing DKD locally and globally^[5–7], epidemiological data have highlighted the prevalence of DKD. Furthermore, DKD has become the leading cause of dialysis^[8], suggesting the severity of this problem. What has worsened the situation is the large portion of patients with abnormal clinical or laboratory measurements who were undiagnosed with DKD^[9]. Therefore, a DKD early prediction model that can be used for risk communication is essential.

There is an emerging literature on developing and validating various risk models for predicting DKD^[10–12], including those specifically for Asians or Chinese populations^{[13, 14]}. However, several factors limit the validity and clinical application of these models.

First, from the study design perspective, the selection of the study population in most previous studies was either unmatched with real-world hospital-visited patients with T2DM or introduced bias due to inappropriate requirements for data completeness. For instance, in certain studies, patients without baseline estimated glomerular filtration rate (eGFR) were excluded^[14–18]. This exclusion criterion could potentially introduce selection bias, as individuals who underwent creatinine testing due to physician suspicion of kidney disease were more likely to have pre-existing renal conditions^{[19, 20]}. Consequently, the areas under the curve (AUC) of the models developed based on these cohorts may have been overestimated.

From a methodology perspective, the final best-performing models in previous studies were mostly classifiers of a single type, such as linear (e.g., lasso) or non-linear (e.g., random forest)^[11]. However, various potential predictors and their interactions could be even more complex, suggesting they are beyond a single algorithm. To date, no model can achieve higher predictive accuracy than any single algorithm by weighting the contributions of each algorithm accordingly (e.g., SuperLearner). In addition to modelling challenges, several other factors hinder such investigations, including risk predictor selection based on univariate screening and the no mention of handling of missing data^[21]. Hence, there are several gaps in the research methodology.

This study applied and compared the performance of multiple machine learning models to patients with T2DM to address these gaps and fused SuperLearner to identify how the predictions of various single algorithms can be combined to improve the prediction performance. Then, we used SHapley Additive exPlanations (SHAP) to improve the model interpretability by delineating how each feature contributes to the prediction outcome at the patient level.

Study design and population

This was a retrospective cohort study of patients initially diagnosed with T2DM at the Ningbo Municipal Hospital of Traditional Chinese Medicine between 2011 and 2023. All patients with at least one occurrence of T2DM were screened using ICD-9 or ICD-10 codes in the outpatient and inpatient departments. The baseline was defined as the time of earliest diagnosis of T2DM in each patient (also called ‘T0’ hereafter). A 6-month T0-centered interval was designated the patient’s baseline time window.

Patients were included if they met the following criteria: (1) age > = 18 years; (2) had at least one hospital visit (regardless of visit type) following the end of the baseline time window.

This study applied the following exclusion criteria: (1) diagnosis of DKD at any time before the end of the baseline time window, identified by a urinary albumin-to-creatinine ratio (UACR) ≥ 30 mg/g (urine Microalbuminuria [mAlb] > = 30 mg, Total protein [TP] − 24 Hour Urine > = 180 mg ), eGFR < 60 mL/min/1.73 m² (CKD-EPI) ^[22], or protein in the urine dipstick test ≥ 1 +; (2) pregnancy within the baseline time window; (3) presence of active infections within the baseline time window; (4) active cancer or malignancy within the baseline time window; (5) autoimmune disease within the baseline time window; (6) involvement of other renal diseases (e.g., urinary tract infection, polycystic kidney disease, glomerulonephritis) any time before the end of the baseline time window.

Notably, we did not set a threshold for the required number of measurements for the key index (e.g., eGFR) or the length of follow-up. Figure 1 shows the study consort diagram.

The ethics board of Ningbo Municipal Hospital of Traditional Chinese Medicine approved this study. The requirement for written informed consent was waived due to the retrospective nature of this study.

Candidate predictor variables and outcomes

The baseline data included 74 routinely measured characteristics extracted from the Medical Data Intelligence Platform related to patients' initial T2DM diagnosis. Candidate predictor variables were classified into the following categories: (1) demographics (e.g., age at T0 and gender), (2) medical history (e.g., hypertension), (3) vital signs (e.g., Systolic Blood Pressure [SBP]), and (4) laboratory tests (e.g., routine blood tests, urinalysis, and biochemical analysis). All baseline predictor variables were collected within the patient’s baseline time window, and if there were multiple testing values for one parameter, the value closest to T0 was selected. A complete list of the predictor variables, attributes of each variable in the dataset, and their abbreviations in the analysis are displayed in Table S1.

The primary endpoint was the occurrence of DKD, regardless of the specific stage it reached, for the first time during follow-up. Specifically, the event was defined as any of the following: (1) persistent albuminuria (UACR ≥ 30 mg/g) (urine Microalbuminuria [mAlb] > = 30 mg, Total protein [TP] − 24 Hour Urine > = 180 mg) over 3 months; (2) eGFR < 60 mL/min/1.73 m²; (3) urinary protein in routine urine tests ≥ 1 +; (4) diagnosis of DKD according to ICD-9 or ICD-10 codes.

Modelling

Patients with outliers in any predictor variables based on clinical experience were removed when preparing the data for modelling. Variables with more than 70% missing data were excluded. Next, patients with missing data for more than 30% of the remaining variables were excluded (the DKD incidence rate did not change significantly before and after exclusions). Missing values were then handled using a multivariate imputation strategy to prepare the data matrix for modeling^[23]. The variables used in the analyses were converted to numeric or binary values (e.g., 1 = male, 0 = female). The primary outcome variable was converted to either zero (negative, no DKD) or one (positive, DKD present).

The cohort dataset was randomly assigned to a training (70%, N = 2303) or validation (30%, N = 988) set before model development. Machine learning algorithms were used to develop prediction models. Typical linear models (i.e., multivariate logistic regression [MLR] and Lasso^[24]), typical non-linear models (i.e., random forest [RF] and extreme gradient boosting [XGBoost]), and SuperLearner were used for model development^[25]. Subsequently, the optimal parameters of the machine-learning algorithms were obtained through cross-validation using the training set. Finally, the models' performances were compared using the validation set, and the model with the highest area under the curve (AUC) was selected as the final best-performing model.

SHAP was used to interpret the results of the best-performing prediction model by computing the contribution of each variable to the prediction^[26].

Model evaluation and validation

The performance of the best model (SuperLearner) was evaluated using AUC. The accuracy of the optimal cut-off value was assessed using sensitivity, specificity, and positive and negative predictive values (PPV and NPV, respectively).

The validation strategy was described in our previous study as below “By applying the final best-performing model to make predictions, patients in the holdout validation cohort were classified into two prognostic groups (i.e., high-risk group vs. low-risk group) based on their predicted probability of DKD and the selected cutoff probability. Their survival curves were compared using the Kaplan–Meier method”^[27].

Statistical analysis

Unpaired two-tailed t-tests and Wilcoxon tests were used to compare the distributions of continuous variables. The median and quantile values were compared for variables that did not follow a normal distribution. Chi-squared tests were used to quantify the relationships between categorical variables (e.g., label balance between the training and validation sets). Multivariate imputation using the chained equations (MICE) package was used for missing-value imputation. SuperLearner is implemented in the ‘SuperLearner’ R package. Statistical significance was set at P = 0.05. Statistical analyses were performed using R version 4.0.1 (R Foundation for Statistical Computing, Vienna, Austria).

Patient characteristics

Overall, 3,291 patients were included in the analysis (the initial diagnosis of T2DM was evenly distributed over the 12-year study period). The median age of these patients was 61 years (interquartile range [IQR], 50–69 years), and 1,551 (47.1%) were female. The median follow-up time (non-normal) was 2.53 years (IQR, 0.97–5.88 years) after the initial diagnosis of T2DM (i.e., T0). During the follow-up, 563 (17.1%) patients were diagnosed with DKD (Fig. 1).

Predictor variables and outcomes

After filtering for missing rate-based data, we identified 3,291 patients and 47 variables (46 predictor variables and one outcome variable). A comparison of those with and without DKD is shown in Table 1. RBC_Cnt, Hb, Hct, LDL, HDL, Ca, MCHC, Lymph_Cnt, TBIL, PLTHct, TC, ALT, Lymph_Per, PLT_Cnt were significantly higher in patients without DKD than in those with DKD. There was a significant positive correlation with age, urine pH, Mono_Per, PDW, DDimer, AST/ALT, RDWCV, creatinine, MCV, and Neut_Per among those with DKD. No significant differences were found in other variables between the groups. All clinical variables were well-balanced between the training and validation sets (Table S2).

Table 1

Demographic, clinical, and biological characteristics of patients with and without diabetic kidney disease
Characteristics	DKD (N = 563)	Non-DKD (N = 2728)	p-value
Demographics (N = 2)
Age	66 (54–75)	60 (50–68)	< 0.001
Gender			0.413
Female	256	1295
Male	307	1433
Lifestyle (N = 2)
is_drinking	62 (11.01%)	360 (13.20%)	0.180
is_smoking	106 (18.83%)	518 (18.99%)	0.977
Blood routine (N = 21)
WBC_Cnt	6.1 (5.1–7.8)	6.3 (5.2–7.7)	0.495
Neut_Cnt	3.8 (2.9–5.1)	3.8 (2.9–5)	0.724
Hct	39.4 (36–42.6)	40.5 (37.5–43.9)	< 0.001
Hb	133 (121–145)	137.5 (127–150)	< 0.001
Lymph_Per	27.3 (20.9–33.55)	29.2 (22.6–35.6)	< 0.001
Baso_Cnt	0.02 (0–0.03)	0.02 (0.01–0.03)	< 0.001
RBC_Cnt	4.26 (3.89–4.71)	4.44 (4.10–4.85)	< 0.001
MCV	91.9 (88.9–95.1)	91.2 (88.48–93.8)	< 0.001
Lymph_Cnt	1.6 (1.3–2.1)	1.8 (1.4–2.2)	< 0.001
MCH	31.1 (29.8–32.3)	31 (30–32)	0.436
Neut_Per	62.8 (56–70.15)	61.7 (54.7–68.6)	0.049
Mono_Cnt	0.4 (0.3–0.5)	0.4 (0.3–0.5)	0.001
Mono_Per	6.8 (5.5–8.1)	6.1 (5.1–7.5)	< 0.001
PDW	16.5 (16.2–16.9)	16.4 (16.1–16.7)	< 0.001
Eos_Per	1.8 (0.9–3)	1.6 (0.9–2.7)	0.104
PLTHct	0.19 (0.15–0.22)	0.2 (0.16–0.23)	< 0.001
MCHC	338 (332.75–343.25)	340 (334–346)	< 0.001
PLT_Cnt	190 (154–230)	198.75 (163–239.6)	< 0.001
MPV	9.8 (9.1–10.75)	9.8 (9–10.7)	0.632
Baso_Per	0.4 (0.3–0.6)	0.4 (0.3–0.6)	0.454
RDWCV	13 (12.6–13.5)	12.8 (12.4–13.4)	< 0.001
Biochemical analysis (N = 13)
LDL	2.27 (1.75–2.76)	2.45 (1.91–3.01)	< 0.001
TC	4.34 (3.7–5.09)	4.58 (3.86–5.27)	< 0.001
Crea	61 (51–73)	59 (49.875–69)	< 0.001
HbA1c	7.6 (6.5–9.6)	7.4 (6.5–9)	0.118
HDL	1.13 (0.9–1.355)	1.185 (0.99–1.43)	< 0.001
TBIL	10.5 (7.65–14)	11.1 (8.3–15)	< 0.001
PA	236.9 (192.65–286.55)	242.8 (196.675–291.225)	0.248
TG	1.43 (1.055–2.06)	1.44 (1.02–2.1)	0.838
ALT	19 (13–29)	21 (14–32)	< 0.001
AST	19 (16–27)	20 (16–27)	0.966
ASTALT	1 (0.8–1.4)	0.9 (0.7–1.2)	< 0.001
Ca	2.23 (2.13–2.32)	2.25 (2.16–2.34)	< 0.001
Eos_Cnt	0.1 (0.06–0.2)	0.1 (0.05–0.17)	0.124
Urinalysis (N = 7)
pH	6.5 (6–7)	6 (5.5–6.5)	< 0.001
glu_qual	204 (36.23%)	1058 (38.78%)	0.278
ket_qual	47 (8.35%)	299 (10.96%)	0.078
uro_qual	6 (1.07%)	23 (0.843%)	0.790
nit_qual	15 (2.66%)	48 (1.76%)	0.209
ob_qual	69 (12.26%)	287 (10.52%)	0.257
bil_qual	3 (0.533%)	8 (0.293%)	0.620
Blood coagulation function (N = 1)
DDimer	133 (78–268.5)	105 (61.75–202)	< 0.001

development, evaluation, and validation

After the missing-value imputation, no significant outliers were observed in the feature set. There was no significant difference in the incidence of DKD between the training and validation sets (17.1% vs. 17.1%, P = 1).

All variables (N = 46) were used as inputs for the five machine learning algorithms to predict the risk of DKD occurrence. The discrimination abilities were compared, and SuperLearner had the highest AUC (0.714, 95% CI 0.673–0.755) (Fig. 2A) and was thus chosen as the final best-performing model. In the validation cohort, the sensitivity, specificity, PPV, and NPV for differentiating DKD were 0.7337, 0.5910, 0.2702, and 0.9149, respectively (Table 2).

Table 2

Prediction accuracy of the SuperLearner estimating the risk of diabetic kidney disease in patients with type 2 diabetes mellitus and normal renal function
	Value (95% CI)
Variable	Training cohort (N = 2303)	Validation Cohort (N = 988)
#DKD	394 (17.11%)	169 (17.11%)
AUC	0.9378	0.7138 (0.673, 0.7546)
Cutoff probability	0.15	0.15
Sensitivity, %	95.94 (93.49, 97.66)	73.37 (66.04, 79.87)
Specificity, %	67.73 (65.58, 69.83)	59.10 (55.64, 62.49)
PPV, %	38.03 (35.00, 41.13)	27.02 (23.00, 31.33)
NPV, %	98.78 (98.02, 99.30)	91.49 (88.78, 93.73)
Positive likelihood ratio	2.9732 (2.7775, 3.1827)	1.7938 (1.5869, 2.0277)
Negative likelihood ratio	0.0600 (0.0371, 0.0970)	0.4506 (0.3486, 0.5824)

To further validate the model, patients in the validation cohort (N = 988) were divided into two prognostic groups based on their DKD predicted probability by the best-performing model: the high-risk (459/988, predicted probability > 0.15) and low-risk (529/988, predicted probability < = 0.15) groups. Their survival curves were compared using the Kaplan-Meier method^[28]. The difference between these groups was statistically significant (P < 0.01) (Fig. 2B).

Explanation of risk factors

SHAP was used to interpret the results of SuperLearner by computing the contribution of each variable to the prediction^[13]. The SHAP summary plot (beeswarm) is shown in Fig. 3A. The importance plot ranked the variables contributing to DKD risk prediction from most to least important as patients’ baseline WBC_Cnt, Neut_Cnt, Hct, Hb and so on.

The SHAP dependence plots of the four selected variables for inclusion in SuperLearner are shown in Fig. 3B (the remaining top 20 plots are shown in Figure S1). Higher SHAP values are associated with an increased risk of developing DKD. Older patients with higher WBC_Cnt, lower Neut_Cnt, higher Hct, and lower Hb levels had an increased risk of developing DKD. The feature value associated with zero SHAP can be used as a reference to determine the desired value of the variable. In other words, it is the tipping point to distinguish between positive and negative contributions to DKD risk.

This study included hospital-visiting patients with an initial T2DM diagnosis and without renal diseases and employed SuperLearner and SHAP to reveal the potentially complex relationship between predictors and DKD outcomes. Compared with existing studies on developing prognostic models to predict DKD in patients with T2DM^{[10, 12]}, we discovered new predictive markers (e.g., WBC_Cnt and urine pH) by applying relatively new analytical methods to somewhat new and more representative populations. The study design ensured that the model pertained to a real-world setting^[29], enhanced clinical application, and the discovered association between new markers and DKD provides new directions for future research.

This study design enhanced the practicality and validity of the final model. First, we considered hospital visits in patients with T2DM; this population was neither general (e.g., the general healthy population) nor too specific (e.g., clinical research)^[13]. Moreover, T2DM patients have a higher risk of developing DKD, increasing their need for DKD screening^{[30, 31]}. Patients initially diagnosed with T2DM typically undergo routine clinical tests, providing readily available measurements that make it easier to estimate their risk.^{[16, 19, 20]}. Second, the EMR-based automatic data extraction, transformation, and loading (ETL) provides a set of characteristics from various sources, making data-driven discovery possible, as evidenced by the higher ranking of many previously uncommonly detected markers, such as WBC_Cnt. The higher ranking of routine blood characteristics allows for exploring the association between DKD and blood dysfunction, such as amenia.

Compared to previous studies that only established a risk model using a single classifier, this study included analyses using linear, non-linear (including ensemble learning algorithms), and SuperLearner. Based on the model performance comparison, SuperLearner was superior to non-linear models which were generally superior to linear models (Fig. 2A) in the validation set. Since multiple characteristics interact, they are more complex than linear or non-linear factors, contributing to the progression and risk of DKD^{[32, 33]}. Although this is consistent with clinical intuition gained from clinical practice, this study demonstrated these findings using data for the first time. Notably, to reduce the potential for selection bias, we did not exclude patients based on their eGFR or other key indices (e.g., UACR); instead, we conducted data-driven filtering. This increased the difficulty of improving the prediction model’s discriminative capability, measured using the AUC. This is because patients who are required to be tested for creatine usually have a higher likelihood of renal disease and suspicion of renal disease by physicians, making it easier to identify those with positive outcomes for the prediction model. Nevertheless, a comparable AUC (0.7 in our study vs. 0.6–0.8 in previous studies) was obtained, thus improving the practicality and validity of the final model^[12].

For a more complex model, greater interpretability is required. One of the strengths of SHAP is its individualised explanation, which means it provides more explanation at a higher granular level^[34]. In addition, the dependence plot is an accumulation of individualised explanations, which depicted the general association between variables and the risk of DKD and revealed a desired value for the variables. For example, the risk of DKD increased sharply at 60–62 years of age (Fig. 3B), consistent with previous studies' results ^[13].

SuperLearner and SHAP confirmed the prognostic value of several common risk factors. The present study showed that age, HbA1c levels, and serum creatinine levels were positively correlated with DKD risk (Fig. 3A, B). Previous studies have identified these markers as important variables for predicting long-term microvascular complications of diabetes^{[10, 13, 35, 36]}. Specifically, older T2DM patients with high HbA1c and creatinine levels had a higher risk of developing DKD. The SHAP dependence plots also provided tipping points where the DKD risk contribution of these variables switched from negative to positive; for example, 8 for HbA1c and 60 for age (Fig. 3B).

Infection, inflammation, and immunity are critical risk factors that affect DKD outcomes^[37]. However, their role in early-stage DKD development has seldom been quantitatively investigated. In this study, WBC_Cnt and Neut_Cnt, reliable markers of systemic inflammation, infections, and immunity^{[38, 39]}, were ranked the highest among all potential predictors (Fig. 3A). We found that elevated WBC_Cnt and reduced Neut_Cnt were risk factors for DKD in patients newly diagnosed with T2DM and that these two characteristics at baseline were proportional to the risk of DKD even if they were in the clinically normal range. The identified association and previous studies indicated that control of inflammatory activity and attention to immune status are critical to T2DM outcomes^[40–42]. Patients with elevated WBC_Cnt and reduced Neut_Cnt at baseline should be closely monitored (e.g., more frequent routine screening for DKD).

In the present study, low Hb levels were associated with an increased risk of developing DKD. Although previous studies have demonstrated the predictive value of Hb for the risk of renal function decline in the early stages of DKD^[43], the negative association between baseline Hb concentrations and the risk of follow-up DKD has often been ignored^{[13, 44]}. We speculate that patients with diabetes often have anaemia, which can lead to renal hypoxia and accelerate the progression of diabetic nephropathy^{[45, 46]}. In contrast, other studies have shown that high Hb levels are also associated with an increased risk of diabetic nephropathy^[47]. Maintaining an appropriate haemoglobin level is important for managing and preventing diabetic nephropathy.

We also found that higher baseline LDL levels were associated with a lower risk of progression to DKD. A previous study identified a similar association with marginal significance (P = 0.09)^[13]. From a clinical perspective, elevated LDL levels are normally associated with an increased risk of DKD^[48]. However, the current study showed the opposite association. We suspect this was due to the intervention provided. Since controlling lipid levels is significant, and lipids are one of the most intervenable indices^{[49, 50]}, patients with higher LDL levels may have received more aggressive lipid-specific interventions. Consequently, LDL levels were lowered, leading to a lower incidence of DKD in the follow-up period.

Notably, new predictive markers were identified in this study, such as urine pH. Urine pH, which reflects the kidney's ability to regulate the body's acid-base balance, and is related to diet, medication, and other factors^[51], was positively associated with the risk of developing DKD. However, this finding has not been reported to date. Since urine pH detection is inexpensive and practical, it can serve as an auxiliary biomarker for DKD risk estimation. Additionally, this discovery provides valuable information for future research.

The findings from the present study have several clinical implications. First, based on our real-world study design and stringent inclusion and exclusion criteria, the model yielded a moderately good and realistic AUC, indicating good discriminatory ability. At the optimal threshold probability, the sensitivity and specificity were 0.7337 and 0.5910, respectively. Clinicians can adjust the threshold for better sensitivity (i.e., a lower risk threshold) or specificity (i.e., a higher risk threshold) case-by-case. Second, the dependence plots generated by SHAP provide index-wise cutoff points for DKD risk alteration, which, when combined with mechanism validation, can be used to facilitate clinical decision-making, such as targeted interventions.

This study has several limitations. First, the analysis was based on patient data from a single centre, which may limit the generalisability of these findings to other populations. Multicentre studies and external validations are required to confirm these results. Second, a prospective study is required to confirm the usefulness of the developed model. Nonetheless, the study design and large dataset generated by our Medical Data Intelligence Platform, the data-driven analytical approach, and the interpretability greatly improved the validity and robustness of the final model.

Using real-world data and novel analytical methods, we developed a SHAP-based explainable SuperLearner model to predict DKD in patients with T2DM with hospital visits. The resulting model revealed discoveries via risk estimation (e.g., markers), which can improve patient outcomes through easy-to-conduct early detection and more targeted interventions.

WBC_Cnt White blood cell count

Neut_Cnt Neutrophil Counts

Hct Hematocrit

Hb Hemoglobin

Lymph_Per Lymphocyte percentage

Baso_Cnt Basophil counts

RBC_Cnt Red blood cell count

MCV Mean corpuscular volume

Lymph_Cnt Lymphocyte count

MCH Mean Corpuscular Hemoglobin

Neut_Per Neutrophil percentage

LDL Low density lipoprotein

age Age

TC Total cholesterol

Crea Creatinine

HbA1c Haemoglobin A1c

Mono_Cnt Monocyte count

HDL High density lipoprotein

Mono_Per Monocyte percentage

PH Urine potentialofhydrogen

PDW Platelet Distribution Width

Eos_Per Eosinophil percentage

TBIL Total bilirubin

PLTHct Plateletcrit

PA Prealbumin

TG Triglyceride

is_smoking Smoking

MCHC Mean corpuscular hemoglobin concentration

ALT Alanine transaminase

PLT_Cnt Platelet count

AST Aspartate aminotransferase

DDimer D-Dimer

glu_qual Glucose qualitative

MPV Mean platelet volume

Baso_Per Basophil percentage

RDW-CV (RDWCV) Red cell volume distribution width -coefficient of variation

is_drinking Drinking

ket_qual Ketone qualitative

AST/ALT (ASTALT) Aspartate aminotransferase/Alanine transaminase

Ca Calcium

Eos_Cnt Eosinophil count

uro_qual Urobilinogen qualitative

nit_qual Nitrite qualitative

gender Gender

ob_qual Occult blood qualitative

bil_qual Bilirubin qualitative

Ethics approval and consent to participate

Consent for publication

Not applicable.

Availability of data and material

Data will be made available on reasonable request.

Conflicts of Interest

The authors have no conflicts of interest to declare.

Funding

The following funds and projects supported this study: 1. Key R&D Projects Funded by Zhejiang Province (2023C03038); 2. Ningbo Top Medical and Health Research Program (No.2022030309); 3. Co-construction Key Laboratory for Kidney Essence Deficiency of National and Zhejiang Provincial Administration of Traditional Chinese Medicine (GZY-ZJ-SY-2306); 4. Ningbo Major Research and Development Plan Project (2022Z135); 5. Ningbo public welfare science and technology plan project (2022S083).

Authors' contributions

XP, YW, XDC, XML, CL and HYW designed this study, reviewed and revised the manuscript. XP, XML, XHF participated in patients’ management. LFL, JMX led the development of the medical data intelligence platform. XML and CL collected clinical data and followed the patients. Statistical analysis was performed by CL and YW. XML, CL, YW interpreted data and drafted the manuscript. All authors contributed to the article and approved the final manuscript.

Acknowledgements

We want to thank the patients at Ningbo Municipal Hospital of Traditional Chinese Medicine who participated in this study.

ORCID

Xiaomeng Lin: 0000-0002-4672-8457

Chao Liu: 0000-0002-6708-4654

Huaiyu Wang: 0000-0002-9093-8793

Xiaohui Fan: 0000-0002-6336-3007

Linfeng Li: 0000-0003-4949-8906

Changlin Li: 0009-0005-1987-0115

Yao Wang: 0000-0003-4121-7313

Xudong Cai: 0000-0002-2132-3242

Xin Peng: 0000-0002-6168-9505

(*Note: Since there are numerous features involved, all abbreviations of the clinical variables are included in Table S1, and a list of abbreviations is provided at the end of the text.)

Tuttle KR, Bakris GL, Bilous RW, Chiang JL, De Boer IH, Goldstein-Fuchs J, et al. Diabetic kidney disease: a report from an ADA Consensus Conference. Diabetes care. 2014;37(10):2864-83.
Mu X, Wu A, Hu H, Zhou H, Yang M. Prediction of Diabetic Kidney Disease in Newly Diagnosed Type 2 Diabetes Mellitus. Diabetes, Metabolic Syndrome and Obesity. 2023:2061-75.
McGrath K, Edi R. Diabetic kidney disease: diagnosis, treatment, and prevention. American family physician. 2019;99(12):751-9.
Braun MM, Khayat M. Kidney Disease: End-Stage Renal Disease. FP essentials. 2021;509:26-32.
Gheith O, Farouk N, Nampoory N, Halim MA, Al-Otaibi T. Diabetic kidney disease: world wide difference of prevalence and risk factors. Journal of nephropharmacology. 2016;5(1):49.
Society CD, Association CM. A nationwide retrospective analysis on chronic diabetic complications and related macrovascular diseases of in-patients with diabetes during 1991-2000. Zhongguo yi xue ke xue yuan xue bao Acta Academiae Medicinae Sinicae. 2002;24(5):447-51.
Pelle MC, Provenzano M, Busutti M, Porcu CV, Zaffina I, Stanga L, et al. Up-date on diabetic nephropathy. Life. 2022;12(8):1202.
Bouhairie VE, McGill JB. Diabetic kidney disease. Missouri medicine. 2016;113(5):390.
Wei J, Wang B, Shen F-j, Zhang T-t, Duan Z, Zhou D-m. Diagnostic value of triglyceride and cystatin C ratio in diabetic kidney disease: a retrospective and prospective cohort study based on renal biopsy. BMC nephrology. 2022;23(1):1-8.
Jiang W, Wang J, Shen X, Lu W, Wang Y, Li W, et al. Establishment and validation of a risk prediction model for early diabetic kidney disease based on a systematic review and meta-analysis of 20 cohorts. Diabetes Care. 2020;43(4):925-33.
González-Rocha A, Colli VA, Denova-Gutiérrez E. Peer Reviewed: Risk Prediction Score for Chronic Kidney Disease in Healthy Adults and Adults With Type 2 Diabetes: Systematic Review. Preventing Chronic Disease. 2023;20.
Slieker RC, van der Heijden AA, Siddiqui MK, Langendoen-Gort M, Nijpels G, Herings R, et al. Performance of prediction models for nephropathy in people with type 2 diabetes: systematic review and external validation study. bmj. 2021;374.
Dong Z, Wang Q, Ke Y, Zhang W, Hong Q, Liu C, et al. Prediction of 3-year risk of diabetic kidney disease using machine learning based on electronic medical records. Journal of Translational Medicine. 2022;20(1):1-10.
Low S, Lim SC, Zhang X, Zhou S, Yeoh LY, Liu YL, et al. Development and validation of a predictive model for chronic kidney disease progression in type 2 diabetes mellitus based on a 13-year study in Singapore. Diabetes research and clinical practice. 2017;123:49-54.
Lin C-C, Niu MJ, Li C-I, Liu C-S, Lin C-H, Yang S-Y, et al. Development and validation of a risk prediction model for chronic kidney disease among individuals with type 2 diabetes. Scientific Reports. 2022;12(1):4794.
Allen A, Iqbal Z, Green-Saxena A, Hurtado M, Hoffman J, Mao Q, et al. Prediction of diabetic kidney disease with machine learning algorithms, upon the initial diagnosis of type 2 diabetes mellitus. BMJ Open Diabetes Research and Care. 2022;10(1):e002560.
Ou S-M, Tsai M-T, Lee K-H, Tseng W-C, Yang C-Y, Chen T-H, et al. Prediction of the risk of developing end-stage renal diseases in newly diagnosed type 2 diabetes mellitus using artificial intelligence algorithms. BioData Mining. 2023;16(1):8.
Hui M, Ma J, Yang H, Gao B, Wang F, Wang J, et al. ESKD Risk Prediction Model in a Multicenter Chronic Kidney Disease Cohort in China: A Derivation, Validation, and Comparison Study. Journal of Clinical Medicine. 2023;12(4):1504.
ElSayed NA, Aleppo G, Aroda VR, Bannuru RR, Brown FM, Bruemmer D, et al. 11. Chronic Kidney Disease and Risk Management: Standards of Care in Diabetes—2023. Diabetes Care. 2023;46(Supplement_1):S191-S202.
Liew A, Bavanandan S, Prasad N, Wong MG, Chang JM, Eiam‐Ong S, et al. Asian Pacific Society of Nephrology clinical practice guideline on diabetic kidney disease–an executive summary. Nephrology. 2020;25(11):809-17.
Collins GS, Omar O, Shanyinde M, Yu L-M. A systematic review finds prediction models for chronic kidney disease were poorly reported and often developed using inappropriate methods. Journal of clinical epidemiology. 2013;66(3):268-77.
Levey AS, Stevens LA, Schmid CH, Zhang Y, Castro III AF, Feldman HI, et al. A new equation to estimate glomerular filtration rate. Annals of internal medicine. 2009;150(9):604-12.
Van Buuren S, Groothuis-Oudshoorn K. mice: Multivariate imputation by chained equations in R. Journal of statistical software. 2011;45:1-67.
Tibshirani R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological). 1996;58(1):267-88.
Van der Laan MJ, Polley EC, Hubbard AE. Super learner. Statistical applications in genetics and molecular biology. 2007;6(1).
Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30.
Wang P, Liu C, Wei Z, Jiang W, Sun H, Wang Y, et al. Nomogram for Predicting Early Mortality after Umbilical Cord Blood Transplantation in Children with Inborn Errors of Immunity. Journal of Clinical Immunology. 2023:1-14.
Jager KJ, Van Dijk PC, Zoccali C, Dekker FW. The analysis of survival data: the Kaplan–Meier method. Kidney international. 2008;74(5):560-5.
Spanopoulos D, Okhai H, Zaccardi F, Tebboth A, Barrett B, Busse M, et al. Temporal variation of renal function in people with type 2 diabetes mellitus: a retrospective UK clinical practice research datalink cohort study. Diabetes, Obesity and Metabolism. 2019;21(8):1817-23.
Hussain S, Jamali MC, Habib A, Hussain MS, Akhtar M, Najmi AK. Diabetic kidney disease: An overview of prevalence, risk factors, and biomarkers. Clinical Epidemiology and Global Health. 2021;9:2-6.
Thomas MC, Brownlee M, Susztak K, Sharma K, Jandeleit-Dahm KA, Zoungas S, et al. Diabetic kidney disease. Nature reviews Disease primers. 2015;1(1):1-20.
Shaw J, Rudzicz F, Jamieson T, Goldfarb A. Artificial intelligence and the implementation challenge. Journal of medical Internet research. 2019;21(7):e13659.
Clark RR, Hou J. Three machine learning algorithms and their utility in exploring risk factors associated with primary cesarean section in low‐risk women: A methods paper. Research in nursing & health. 2021;44(3):559-70.
Wang D, Thunéll S, Lindberg U, Jiang L, Trygg J, Tysklind M. Towards better process management in wastewater treatment plants: Process analytics based on SHAP values for tree-based machine learning methods. Journal of Environmental Management. 2022;301:113941.
Saputro SA, Pattanaprateep O, Pattanateepapon A, Karmacharya S, Thakkinstian A. Prognostic models of diabetic microvascular complications: a systematic review and meta-analysis. Systematic reviews. 2021;10(1):1-11.
Khanam PA, Hoque S, Begum T, Habib SH, Latif ZA. Microvascular complications and their associated risk factors in type 2 diabetes mellitus. Diabetes & Metabolic Syndrome: Clinical Research & Reviews. 2017;11:S577-S81.
Rayego-Mateos S, Rodrigues-Diez R, Fernandez-Fernandez B, Mora-Fernández C, Marchant V, Donate-Correa J, et al. Targeting inflammation to treat diabetic kidney disease: The road to 2030. Kidney International. 2022.
Winter L, Wong LA, Jerums G, Seah J-m, Clarke M, Tan SM, et al. Use of readily accessible inflammatory markers to predict diabetic kidney disease. Frontiers in endocrinology. 2018;9:225.
Ryu S, Shin JW, Kwon S, Lee J, Kim YC, Bae Y-S, et al. Siglec-F–expressing neutrophils are essential for creating a profibrotic microenvironment in renal fibrosis. The Journal of Clinical Investigation. 2022;132(12).
Guo W, Song Y, Sun Y, Du H, Cai Y, You Q, et al. Systemic immune-inflammation index is associated with diabetic kidney disease in Type 2 diabetes mellitus patients: Evidence from NHANES 2011-2018. Frontiers in Endocrinology. 2022;13:1071465.
Liu L, Gao B, Wang J, Yang C, Wu S, Wu Y, et al. Clinical significance of single and persistent elevation of serum high-sensitivity C-reactive protein levels for prediction of kidney outcomes in patients with impaired fasting glucose or diabetes mellitus. Journal of Nephrology. 2021;34:1179-88.
Schei J, Stefansson VTN, Eriksen BO, Jenssen TG, Solbu MD, Wilsgaard T, et al. Association of TNF receptor 2 and CRP with GFR decline in the general nondiabetic population. Clinical Journal of the American Society of Nephrology: CJASN. 2017;12(4):624.
Yamanouchi M, Furuichi K, Shimizu M, Toyama T, Yamamura Y, Oshima M, et al. Serum hemoglobin concentration and risk of renal function decline in early stages of diabetic kidney disease: a nationwide, biopsy-based cohort study. Nephrology Dialysis Transplantation. 2022;37(3):489-97.
Keane WF, Zhang Z, Lyle PA, Cooper ME, de Zeeuw D, Grunfeld J-P, et al. Risk scores for predicting outcomes in patients with type 2 diabetes and nephropathy: the RENAAL study. Clinical journal of the American Society of Nephrology. 2006;1(4):761-7.
Sahay M, Kalra S, Badani R, Bantwal G, Bhoraskar A, Das A, et al. Diabetes and Anemia: International Diabetes Federation (IDF)–Southeast Asian Region (SEAR) position statement. Diabetes & Metabolic Syndrome: Clinical Research & Reviews. 2017;11:S685-S95.
Liu J, Wei Q, Guo C, Dong G, Liu Y, Tang C, et al. Hypoxia, HIF, and associated signaling networks in chronic kidney disease. International journal of molecular sciences. 2017;18(5):950.
Xin S, Zhao X, Ding J, Zhang X. Association between hemoglobin glycation index and diabetic kidney disease in type 2 diabetes mellitus in China: A cross-sectional inpatient study. Frontiers in Endocrinology. 2023;14:1108061.
Siddiqui K, George TP, Nawaz SS, Yaslam M, Almogbel E, Al-Rubeaan K. Significance of glycated LDL in different stages of diabetic nephropathy. Diabetes & Metabolic Syndrome: Clinical Research & Reviews. 2019;13(1):548-52.
Dake AW, Sora ND. Diabetic dyslipidemia review: an update on current concepts and management guidelines of diabetic dyslipidemia. The American journal of the medical sciences. 2016;351(4):361-5.
Jialal I. Management of diabetic dyslipidemia: Navigating the new American and European Guidelines. Diabetes & Metabolic Syndrome. 2020;14(5):877-9.
Miki A, Hashimoto Y, Tanaka M, Kobayashi Y, Wada S, Kuwahata M, et al. Urinary pH reflects dietary acid load in patients with type 2 diabetes. Journal of Clinical Biochemistry and Nutrition. 2017;61(1):74-7.

FigureS1.pdf
Fig S1. Shapley additive explanationdependence plot of SuperLearner (the remaining 16 from the top 20), depicting how a single variable affects the prediction. (Footnotes: The abbreviations of all analytical variables are detailed in Table S1)
SupplementalMaterials3.docx
Supplemental Materials Table S1. A complete list of the variables in the analysis Table S2. Demographic, clinical, and biological characteristics of patients in the training and validation cohorts. Footnotes: Continuous variables are presented as medians (interquartile ranges), and categorical variables are presented as counts (percentages).

Download PDF

Version 1

posted

You are reading this latest preprint version

A SuperLearner approach to predicting diabetic kidney disease upon the initial diagnosis of T2DM in hospital

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Introduction

Materials and Methods

Study design and population

Candidate predictor variables and outcomes

Modelling

Model evaluation and validation

Statistical analysis

Results

Patient characteristics

Predictor variables and outcomes

development, evaluation, and validation

Explanation of risk factors

Discussion

Conclusions

Abbreviations

Declarations

References

Supplementary Files

Status:

Version 1