A model for predicting the progression of CKD to advanced stages and the need for RRT was developed and validated in this study. Risk at 4.5 years and time-to-event (survival analysis) in patients with CKD stages 3–5 were assessed based on data analysis using artificial intelligence tools. The proposed model showed adequate performance, allowing its systematic implementation in CKD clinical management programs, and guaranteeing its usefulness as part of the strategies that guide clinical decision making.

Publications of prediction models for chronic kidney disease have increased in recent years, and this has been analyzed in four recently published systematic reviews [14,15,16,17].However, skepticism still remains among clinicians regarding the performance and applicability of these models, apart from the fact that clinical practice standards are generic when it comes to defining a precise recommendation on this subject. For example, KDIGO recommends using prediction models for timely referral to RRT. However, it does not define how and when to use these tools [18].

In this context, it is important to have prediction tools that guide decision making, the management of prevention programs, and timely multidisciplinary intervention strategies. However, several of the published models include populations with different levels of disease severity and do not precisely define the outcomes and the time of disease progression in which they should be used [14]. Our study developed a prediction model that can be applied in patients with CKD stages 3–5, with a 5-year follow-up, determining the following as main outcomes: (i) progression of CKD stage based on the eGFR (ii) reduction greater than 25% in the eGFR compared to baseline, and (iii) onset of RRT (dialysis for more than 3 months or kidney transplant).

To minimize the risk of bias, a cohort of 2,143 patients distributed in three arms was included for each outcome. For the incidence of RRT, the entire cohort of 2,143 patients was included. For progression of CKD stage and significant decrease in eGFR, there were 2,060 patients, of which 1,035 and 802 presented the respective outcomes. Furthermore, the external validation process included a cohort of 648 patients. As it became evident during the cohort selection process, when working with real data, this study was contingent upon missing information, as is the case of alkaline phosphatase, where no data was found for 92.6% of the patients. Because these are precise clinical data, the use of imputation strategies was ruled out. On the other hand, as demonstrated by the performance metrics, the models that predict RRT could be showing data overfitting due to imbalance in the classes; there were very few patients who presented the outcome. This was especially evident in the external validation for RRT, where a C-index of 0.9518 was obtained, but this cohort only had two patients who showed this outcome.

The inclusion of a validation cohort increases the reliability of the models. This is how the kidney failure risk equation (KFRE) developed by Tangri et al. [7] has become the standard for comparison, given that the model has shown consistently good performance in patients with CKD stages 3–5 in several external validation studies with a low risk of bias. One of these validations included 31 multinational cohorts with a mean baseline eGFR of 46 mL/min/1.73 m2 and showed that the KFRE model has a high discrimination capacity and adequate calibration [15]. Another model validated with an external cohort and a low risk of bias is the Kaiser Permanente Northwest (KPNW) model [9].

Although some clear prognostic factors of this and other prediction models are the main ingredients of our model, the context of our country requires particular sociodemographic factors that are included here. Our cohort is consistent with the country’s reality, with an absence of relevant variables that limit the sample, and with greater participation of patients from the central region. Upon admission to the cohort, the largest participation included stage 3 patients, and the frequency RRT initiation decreased over time.

The prediction of CKD progression to advanced stages or admission to dialysis allows for the implementation of strategies for individualized treatment, control of risk factors, evaluation of population indicators, establishment of education management strategies, selection of renal support therapies, and preemptive renal transplantation [19].

Conventionally, risk factors related to CKD progression included in most prediction models have been demographic variables such as age, sex, and geographic origin, and clinical variables such as comorbidities, eGFR or its deterioration in the previous year, albuminuria, serum bicarbonate, albumin, calcium, hemoglobin, and phosphorus, among others [6].In recent publications describing other models proposed concomitantly with our own model, we found unconventional predictors such as biomarkers (CXCL12, NT-proBNP, NGAL, and troponin T) and the application of artificial intelligence methods [20,21,22].

Although CKD management programs in Colombia seem to be very well structured, they do not offer sub-specialized management programs for the entire population with CKD in terms of health policies. This is due to several reasons, in particular, a disproportion between the number of patients with CKD and the current number of nephrologists in the country. In their daily roles, these nephrologists cover different areas of work, such as critical care, hospital nephrology, kidney units, transplant groups, and outpatient and preventive services. This reveals an insufficient human resource, far from international standards (a nephrologist–patient ratio of approximately 1:2000 patients), requiring the support of non-nephrologists, trained and familiar with this group of patients. In general, these programs are configured based on an initial snapshot of the stage of the disease exclusively on the eGFR. It is possible to optimize the care of these patients with the implementation of the prediction models. Thus, a possible scenario can be proposed in which the care of stage 3 patients with a low risk of progression could continue in first level centers, improving the window of opportunity for the highly specialized care of patients classified showing a high risk of progression. This would lead to an improvement in cost-effectiveness indicators.

Among the strengths of our model, we can highlight the following: (1) the creation of a machine learning-based technological tool that makes it possible for non- specialized healthcare personnel to estimate the risk for patients. This is useful in prevention programs in the context of a limited specialized human resource; (2) the estimation of the risk of a significant decrease in eGFR as an additional outcome to stage progression benefits the early implementation of prevention strategies to avoid the deterioration of renal function, even within the same risk category; (3) the predictors were obtained from frequent registration data in the clinical follow-up records of this group of patients. The use of additional resources was not necessary for their application; (4) to date, no models have been developed in a population with sociodemographic characteristics like ours. This represents an opportunity to use a tool based on the context of our country and others in the Latin American region; (5) an external validation cohort was included, and the tool is designed to be implemented in a specific population of patients with stage 3–5 CKD; and (6) the translation of a mathematical model into a digital tool in the form of a calculator with a graphical, easy-to-use interface facilitates its use in these patient care scenarios.

This model has certain limitations as follows: (1) the follow-up period was 5 years, which limits the analysis of disease progression in some groups of patients; (2) there was no differentiation into racial groups, which affects the analysis of subgroups and progression according to this predictor, which is part of the tools frequently used to calculate the eGFR; (3) model development was based on a retrospective cohort, which reduced the availability of complete data in relation to the predictors of interest for the final analysis; (4) the population subgroup belonging to the middle-income socioeconomic stratum was overrepresented (90%), which is well above the percentages recorded in national evaluation surveys (45.5%) and which may imply more favorable socioeconomic contexts and determinants than those of the average Colombian population; 6) the percentage of representation of women was greater than 50%, which may not be related to the proportion of the population on dialysis or subjected to kidney transplant; and 7) the geographical distribution of the patients included in the sample was linked to the distribution of affiliates of the selected insurer, which does not have members in all regions of the country. This prevents the generalization of estimates to the entire national territory.

In conclusion, the developed model constitutes a tool to help manage the progression of CKD in terms of early intervention and optimization of available human resources.