Personalised prediction of maintenance dialysis initiation in patients with chronic kidney disease stages 3–5: a multicentre study using the machine learning approach

Introduction

Chronic kidney disease (CKD) and end-stage renal disease (ESRD) are significant global health problems that burden healthcare systems worldwide. In Taiwan, the overall prevalence of CKD was 8.2%, and the incidence of treated ESRD was 529 per million population, according to the US Renal Data System annual report published in 2021.1 International guidelines recommend referring patients with CKD to nephrology for pre-ESRD care upon reaching an advanced stage to improve the quality of care and reduce costs.2 3 One critical component of pre-ESRD care is counselling patients on choosing kidney replacement therapy (KRT) following shared decision-making. It may involve preparing vascular access for haemodialysis (HD) at least 6 months before HD initiation, placing a peritoneal dialysis (PD) catheter at least 2 weeks before PD initiation, or identifying suitable donors for pre-emptive kidney transplantation before dialysis is required to replace failing kidney function. Therefore, personalising the timing of referral for maintenance dialysis initiation is essential for each patient. However, early personalised estimation of the optimal timing for maintenance dialysis initiation presents a significant challenge for patients with CKD. This decision relies not only on the glomerular filtration rate (GFR) level but also on symptoms of uraemia syndrome and the ability to manage complications such as electrolyte imbalance, acid–base disturbances and fluid overload through medical treatment.3–5

Numerous traditional and artificial intelligence (AI) models have been developed and evaluated to estimate the duration until the initiation of KRT among patients with CKD.6–10 Among these models, machine learning (ML) approaches that use complex computer algorithms are effective in identifying the most critical factors and developing predictive models with superior performance.11 12 However, existing models rely primarily on laboratory data and comorbidities to inform their analyses and predictions, neglecting important indicators of CKD progression such as an annual decline in GFR, proteinuria and medication use. Furthermore, the timing (index date) chosen for these models is heterogeneous due to the diverse stages of CKD represented in the patient cohorts, resulting in relatively low predictive power. Therefore, developing an accurate and reliable predictive model that incorporates comprehensive clinical data and standardises index date selection is crucial to improving personalised decision-making for patients with CKD.

In this study, we aim to develop and validate an ML model for early personalised prediction of maintenance dialysis initiation within 1-year and 3-year timeframes among patients with CKD stages 3–5 using a multicentre longitudinal cohort.

Materials and methods

Study population

We identified patients diagnosed with CKD between 1 January 2008 to 31 December 2017, based on the International Classification of Disease, ninth revision (ICD-9) code 585 and 10th revision (ICD-10) code N18. First-time patients diagnosed with CKD stages 3–5 and who had not undergone KRT, defined as the index date, were included in the study. We confirmed the diagnosis based on the G-stages, in which the estimated glomerular filtration rate (eGFR) was calculated using the CKD-EPI creatinine equation.13 To define correctly the stage of CKD, only patients who had been diagnosed with CKD prior to the index dates and had their eGFRs observed within 3–6 months after the index date were included. Patients who were younger than 20 years at the time of diagnosis of CKD stages 3–5 or underwent pre-emptive kidney transplantation were excluded from the study (figure 1).

Request permission

Figure 1

Enrollment process of the first study. The index date is the point of time at which patients with CKD were first diagnosed with stages 3–5. CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; KRT, kidney replacement therapy.

Results

The performances of different prediction models

The performances of different prediction models are shown in table 2. For a 1-year prediction of successful dialysis treatment, the highest AUC value of 0.96 was observed for the ANN model (ie, sensitivity, 0.88; specificity, 0.75; precision, 0.39 and F1 score, 0.6), followed by the GBM and RF models with an AUC value of 0.89. Likewise, the ANN model was performed with a better AUC value of 0.92 (ie, sensitivity, 0.87; specificity, 0.79; precision, 0.63 and F1 score, 0.73) than other ML models in the 3-year prediction of receiving maintenance dialysis. The ROC curves of varying prediction models for 1-year and 3-year successful dialysis treatment are shown in figure 2.

Table 2

View inline|View popup

Summary of different classification models

Request permission

Figure 2

The performance of the prediction models in the testing dataset. (A and C) 1-year prediction with machine learning and ANN models; (B and C) 3-year prediction with machine learning and ANN models. ANN, artificial neural network; AUC, area under the receiver operating characteristic curve; LGBM, light gradient boosting machine; XGB, extreme gradient boosting.

Features importance

The lists of the top 20 important features that might impact the prediction model’s performance for 1-year and 3-year successful dialysis are shown in figure 3. The essential features of the 1-year and 3-year follow-up models were baseline eGFR, BUN, creatinine, triglyceride, age, gender, Hgb, CHOL, PLTs, albuminuria, diabetes disease, hypertension and related medications (eg, diuretics, insulin, dipeptidyl peptidase-4 and calcium channel blockers).

Request permission

Figure 3

Feature importance of the ANN prediction model. (A) 1-year prediction model and (B) 3-year prediction model. ANN, artificial neural network; eGFR, estimated glomerular filtration rate; DPP-4, dipeptidyl peptidase

Discussion

Our study findings demonstrated that ML classification models are well suited for a meaningful prediction of the initiation of maintenance dialysis in patients with CKD stages 3–5. The ANN method showed a better performance level for 1-year and 3-year prediction of dialysis commencement with a higher AUC (0.96 and 0.92), good sensitivity (0.88 and 0.87) and specificity (0.75 and 0.79).

In previous studies, AI models have been applied to predict CKD progression and start KRT. In 2015, Jamshid Norouzi et al used an adaptive neuro-fuzzy inference system to predict renal failure progression. Their model could accurately (>95%) predict the GFR for 6-month to 18-month intervals. However, only 465 patients with CKD were included in their study, and it was noted that proteinuria was not an important feature in their model.8 In 2019, Jing Xiao et al developed ML models to predict CKD progression. Their model used only the patient’s demographics and biochemical blood features, not features derived from a urinalysis. Besides, the predictive power of the model was not high (AUC: 0.873, sensitivity: 0.83 and specificity: 0.82).17 Another model was performed using only comorbidity data from 8492 patients to predict the onset of KRT, and their results were even lower (AUC, sensitivity and specificity were only 0.773, 0.623 and 0.781, respectively).7 Recently, Qiong Bai et al also conducted an ML model to predict the risk of ESRD. This model selected many important factors associated with the progression of CKD, including demographics, blood tests and comorbidities, but not proteinuria. However, the predictability has not improved compared with previous models.18 It could be explained by the fact that many patients in that study were in the early stages of CKD, resulting in a low percentage of those who progressed to ESRD when followed for a short period of time. The imbalance in the outcome can significantly affect the model’s predictive power.

In this study, we only focused on patients with CKD stages 3–5, and their risk of progression to ESRD is high. Hence, predicting the time of their dialysis commencement is very practical in our daily clinical care. Moreover, we carefully identified the model features associated with CKD progression and KRT based on the clinical setting and traditional logistic regression analysis. Forty-five significant prognostic factors were selected, including patient demographics, comorbidities, routine blood and urine tests and commonly used medications. Therefore, the predictive ability of our model has higher accuracy.

In further analysis of the ANN model, we also ranked all predictors according to their influence on the 1-year and 3-year models using SHAP values.19 Notably, several distinct features have been identified, respectively. For example, age, comorbidity (CCI score), PLT counts and WBC counts were important contributing factors in the 1-year prediction model, whereas gender and other medications such as proton pump inhibitors, beta-lactam antibacterial agents, organic nitrates and H2-receptor antagonists were relevant factors in the 3-year model. Common important factors identified in both models included eGFR at baseline, blood urea, serum creatinine and albuminuria (see figure 3). These are also key determinants for the risk classification of CKD according to the 2012 KDIGO guidelines.13 20 Other contributing factors in both models included serum Hgb level, TG or CHOL levels, hypertension, diabetes mellitus, diuretic use, antihypertensive agents and medications for controlling blood glucose levels (see figure 3). Anaemia typically develops during the course of CKD; a decrease in serum Hgb is significantly associated with the progression of CKD.21 22 Diabetic nephropathy is the leading cause of ESRD in adults.23 24 In patients with diabetic CKD, blood glucose levels are associated with poor outcomes such as serum creatinine doubling, ESRD and mortality, and intensive glycaemic control could reduce these risks.25–29 Additionally, several studies have demonstrated that certain levels of dyslipidaemia is independently associated with rapid renal progression, KRT, all-cause mortality and cardiovascular death in predialysis patients.30–33 Hypertension may occur early during the course of CKD and is related to a more rapid decline of kidney function, the development of cardiovascular disease and death in patients with CKD.34 35 Early intervention and tight control of blood pressure could lessen the risk of CVD and all-cause death in patients with and without CKD.36 37 Diuretics are an important part of guideline-directed medical therapy for patients with CKD with hypertension, oedema and hyperkalaemia.38 In terms of adverse effects, whether diuretics are an independent risk factor for CKD progression remains controversial. However, these medicines played important roles in both our models.39–41 Therefore, diuretics should be used with caution in patients with CKD stages 3–5. Finally, the GFR decline rate is also influenced by some immutable patient factors. The Kidney Disease Outcomes Quality Initiative guideline has provided ample evidence that African-American race (not justified in this study), male gender and older age are related to a more rapid GFR reduction.20 In summary, our models take advantage of the important factors involved in the progression of CKD, are consistent with current clinical practice guidelines and are highly applicable. They could be a good screening tool to determine the likelihood of initiating long-term dialysis by using the available clinical data on the patient. Several limitations need to be addressed. First, due to the patient’s lack of weight and height, the body surface area was not adjusted for eGFR. As a result, the determination of G-stage using unadjusted eGFR may be inaccurate for oversized patients. Second, using the decline in eGFR between baseline eGFR and previous eGFR may not accurately capture the progression of CKD when compared with the annual decline in eGFR during the follow-up period. Consequently, this factor did not significantly contribute to our model. Third, we only used retrospective data from three hospitals in Taipei to create our models, and it is widely recognised that racial and regional variables also influence CKD progression. Further work should involve training and validating the models through multinational and multiracial data before the clinical application is generalised. Fourth, we incorporated all important features into the prediction model, acknowledging that this approach might not be practical for clinical implementation. Nevertheless, these features underwent meticulous screening and hold varying degrees of significance in relation to CKD progression. Additionally, we assessed the model using only the top 10 important features and obtained comparable results (online supplemental tables S4 and S5, online supplemental figures S1 and S2).

Conclusion

We have shown that using the machine learning approach can develop a highly predictive model for estimating the timing of maintenance dialysis initiation in patients with CKD stages 3–5, which provides a further step towards personalised treatment in this population.