Modeling chronic kidney disease progression using a continuous-time multistate Markov model: Transition dynamics and covariate effects Authors: Tolesa Futasa Begna, Akalu Banbeta Tereda, Jaleta Abdisa Fufa, Tadele Akeba Diriba, Legesse Kassa Debusho Published: 2026 ABSTRACT Chronic kidney disease is a progressive condition that impairs renal function and can ultimately lead to kidney failure. Its progression involves multiple intermediate stages, making conventional survival models inadequate for capturing the complexity of disease transitions. This study aimed to model chronic kidney disease progression using a continuous-time multistate Markov process to evaluate stage-specific transition dynamics and the effects of clinical and demographic factors. A retrospective cohort study of 194 patients with chronic kidney disease, comprising 1506 clinic visits at Jimma Medical Center between February 2019 and February 2024, was conducted. The multistate Markov framework was used to estimate transition intensities, transition probabilities, mean sojourn times, and next-state probabilities and to assess the effects of clinical and demographic characteristics on transitions between chronic kidney disease stages. Transition probabilities indicated an increasing likelihood of progression to more advanced stages over time, accompanied by decreasing probabilities of remaining in the same stage. The mean sojourn times in stages 1, 2, 3, and 4 were 4.33, 4.44, 5.79, and 4.57 months, respectively. Patients with hypertension were significantly more likely to transition from stage 4 to stage 5 than those without hypertension (hazard ratio = 2.62, 95% confidence interval: 1.79–3.84). Diabetes, hypertension, and cardiovascular disease were the primary factors associated with chronic kidney disease progression. Mean sojourn times and next-state probabilities complemented transition intensities by describing the expected duration of each chronic kidney disease stage and the likelihood of subsequent transitions. These findings provide insights into chronic kidney disease progression and may support stage-specific management and intervention strategies. Background and motivation Chronic kidney disease (CKD) is a major global public health concern that contributes substantially to morbidity and mortality.1,2 The prevalence of CKD continues to increase, affecting approximately 850 million people worldwide, with most cases occurring in low- and middle-income countries (LMICs). 3 Among adults aged 20–50 years, CKD is highly prevalent in Sub-Saharan Africa (SSA), primarily because of glomerular disorders, diabetes mellitus, and hypertension (HTN).4,5 Recent studies from different regions of Ethiopia have reported a prevalence ranging from 7.7% to 17.3%, with increasing rates paralleling the rising burden of diabetes and HTN.6,7,8,9 Limited awareness of the disease, inadequate access to life-saving dialysis and kidney transplantation, and the high prevalence of associated risk factors contribute to the growing burden of CKD in LMICs, including Ethiopia.3,9 The progression of CKD is a complex process involving multiple intermediate disease stages.10,11 The disease progresses through five stages, from mild renal impairment to severe kidney dysfunction, ultimately leading to end-stage renal disease (ESRD). 12 Given the high prevalence and clinical burden of CKD in Ethiopia, a comprehensive analysis of disease progression and its relationship with treatment outcomes is needed.6,13 Understanding transitions between disease stages is essential for informing timely interventions and improving patient management. This study aimed to estimate transition intensities and evaluate the effects of clinical and demographic factors on CKD progression using a continuous-time multistate Markov process model, 14 thereby improving the identification of patients at high risk of disease progression and supporting evidence-based disease management. From a methodological perspective, conventional approaches for analyzing CKD progression, such as regression models and survival analysis techniques, may not adequately capture the complex transitions between multiple disease states.15,16,17 In contrast, continuous-time multistate Markov process models provide a robust framework for representing multiple health states and the transitions between them over time. 18 These models account for the stochastic nature of disease progression and enable a more comprehensive evaluation of treatment effects while incorporating time-varying covariates that influence transition probabilities. 19 The continuous-time multistate Markov process model is a well-established method for describing transitions between a series of health states over continuous time. When evaluating progression through consecutive stages of a chronic disease, this model offers several advantages over conventional approaches. First, it enables the simultaneous estimation of multiple transition intensities between disease states, such as transitions from stage 1 to stage 2 or from stage 3 to ESRD, thereby providing a comprehensive description of disease progression, including transition probabilities and mean sojourn times. 19 Second, the model can incorporate individual-level covariates that influence transition intensities, allowing researchers and healthcare providers to identify patients at high risk of disease progression and develop targeted intervention strategies. 20 Recently, there has been growing interest in applying advanced statistical methods to improve the understanding of CKD progression. However, to the authors’ knowledge, most studies conducted in Ethiopia have primarily focused on identifying risk factors, with limited attention to the complex multistate transitions required to estimate transition intensities.21,22 For example, studies by Lintu et al. and Milik and Hrynkiewicz have applied reversible transition models to CKD, although disease progression is characterized by irreversible deterioration of renal function.23,24 Conversely, other studies have reported only the mean sojourn time for each disease stage,23,25 which, without corresponding transition probabilities, provides an incomplete description of disease progression. The present study addressed this gap by applying a robust statistical framework to estimate transition intensities and mean sojourn times while accounting for relevant clinical and demographic covariates. The study had two objectives: (a) to estimate the transition intensities between different stages of CKD and (b) to assess the effects of covariates on these transition intensities. By achieving these objectives, this study seeks to identify potential intervention points along the continuum of care for patients with CKD, thereby informing clinical practice and public health strategies aimed at reducing disease progression and its associated burden. The remainder of this paper is organized as follows. The methodology section describes the study data and provides an overview of the continuous-time multistate Markov process model used to analyze longitudinal data, estimate transition intensities, and calculate mean sojourn times, together with the associated statistical inference. The results section presents the study findings. The discussion and conclusion sections provide the discussion and conclusions, respectively, along with recommendations for future research. Methodology Study design and data collection A hospital-based retrospective cohort study was conducted using medical records of patients with CKD at Jimma Medical Center (JMC). The study population comprised all patients diagnosed with CKD who received treatment at JMC between February 2019 and February 2024. The data collection tool was developed through a review of the relevant literature and consultation with healthcare professionals specializing in CKD. Patient eligibility was determined by reviewing medical records according to the predefined inclusion criteria. Subsequently, sociodemographic, behavioral, and clinical data were extracted from patient registration logs and medical records based on regular follow-up visits during the study period. Diagnosis and treatment of CKD at JMC Healthcare professionals in Ethiopia follow the national guidelines for the clinical and programmatic management of major non-communicable diseases, including CKD. These guidelines were developed by the Ethiopian Federal Ministry of Health (EFMoH) and are based on the World Health Organization (WHO) priority package for non-communicable diseases. 26 JMC, a major referral hospital in southwest Ethiopia, diagnoses and manages CKD in accordance with national guidelines and current international recommendations. 27 At JMC, the diagnosis and management of CKD are integral components of patient care. The diagnostic process begins with a comprehensive clinical assessment, including a detailed medical history and evaluation of established risk factors such as a family history of kidney disease, diabetes mellitus, HTN, and lifestyle-related factors. The diagnosis is supported by laboratory and imaging investigations, including measurement of serum creatinine, estimation of glomerular filtration rate (GFR), urinalysis, and renal imaging studies. Furthermore, treatment is tailored to the stage of CKD and includes several key components. First, lifestyle modifications are emphasized, with patients receiving counseling on dietary changes, including reduced sodium and protein intake, weight management, and smoking cessation. Second, pharmacological management includes medications such as angiotensin-converting enzyme (ACE) inhibitors or angiotensin receptor blockers (ARBs) to control HTN and reduce proteinuria, whereas diuretics are prescribed to manage fluid overload. Third, the management of comorbid conditions, particularly diabetes mellitus and cardiovascular disease, is an essential component of CKD care. Fourth, regular follow-up visits are conducted to monitor renal function and adjust treatment plans as needed. For patients with advanced CKD (stages 4 and 5), JMC provides renal replacement therapy, including dialysis, or referrals for kidney transplantation when appropriate. Inclusion and exclusion criteria Adults aged 18 years or older with CKD, defined as an estimated GFR (eGFR) of <60 mL/min/1.73 m2 persisting for more than 3 months, were eligible for inclusion. Eligible participants were those who received treatment at JMC during the study period and had complete clinical and follow-up information in their medical records. Patients with acute kidney disease (AKD) without a confirmed diagnosis of CKD were excluded. In addition, patients who had been newly diagnosed with CKD but had not yet initiated treatment, those with incomplete clinical documentation, those receiving renal replacement therapy at baseline, kidney transplant recipients, patients with incomplete baseline or follow-up data, those with less than 3 months of follow-up, and those who transferred into or out of the hospital during the study period were excluded. Overall, 290 patients were initially identified, of whom 194 met the eligibility criteria. The remaining 96 patients were excluded because they did not meet the inclusion criteria. Study variables The dependent variable was the change in eGFR stage over a 3-month period, determined by comparing eGFR values at baseline and follow-up. The independent variables included age (years), sex (female or male), residence (urban or rural), urinary protein concentration (mg/dL), serum creatinine (mg/dL), hemoglobin (g/dL), sodium (mEq/L), potassium (mEq/L), diabetes mellitus (no or yes), HTN (no or yes), cardiovascular disease (no or yes), smoking status (no or yes), heart disease (no or yes), human immunodeficiency virus/acquired immunodeficiency syndrome (HIV/AIDS) (no or yes), anemia (no or yes), and coronavirus disease 2019 (COVID-19) (no or yes). Progression of CKD CKD is characterized by pathological abnormalities or markers of kidney damage indicated by an eGFR of <60 mL/min/1.73 m2 persisting for more than 3 months. eGFR is a widely used measure of kidney function that estimates the volume of blood filtered by the kidneys over time. CKD stages were classified according to the Kidney Disease Outcomes Quality Initiative (KDOQI) guidelines. 10 Changes in kidney function, primarily reflected by the eGFR and other indicators of kidney damage, form the basis for classifying disease progression. These stages provide a framework for assessing disease severity, monitoring progression, and guiding clinical management. A normal eGFR is approximately 90–120 mL/min/1.73 m2 in healthy adults. Figure 1 illustrates the possible transitions between disease stages. Stages 1–4 were treated as transient states, whereas stage 5 was specified as an absorbing state, from which no further transitions were permitted in the multistate model. ESRD corresponds to the final stage of CKD and is associated with kidney failure, cardiovascular complications, and an increased risk of mortality. 10 Transition states of CKD progression diagram. Operational definitions and measurements In this study, urea concentration in serum or plasma was measured using enzymatic or colorimetric methods, whereas protein levels in urine were assessed via dipstick or quantitative methods. Serum creatinine levels were determined using methods such as the Jaffe reaction or enzymatic assays to evaluate kidney function. Hemoglobin concentration were measured using automated hematology analyzers or spectrophotometric methods to assess anemia. Serum sodium and potassium concentrations were measured to evaluate electrolyte status through ion-selective electrode methods or flame photometry. Smoking status was classified as yes or no based on the information recorded in the medical records. Patients were classified as smokers if they had a documented history of cigarette smoking and as non-smokers if they had no documented history of cigarette smoking. Comorbidities, including HTN, diabetes mellitus, cardiovascular disease, heart disease, HIV/AIDS, anemia, and COVID-19, were identified based on the documented clinical diagnosis in the patients’ medical records. HTN was defined as a systolic blood pressure (SBP) of ≥140 mmHg, a diastolic blood pressure (DBP) of ≥90 mmHg, or both. Blood glucose levels were used to classify diabetes. Patients with fasting blood sugar (FBS) levels <125 mg/dL and random blood sugar (RBS) levels <200 mg/dL were classified as nondiabetic, whereas those with FBS ≥126 mg/dL, RBS >200 mg/dL, and a glycated hemoglobin (HbA1C) ≥6.5 were classified as diabetic. Cardiovascular disease was defined as the presence of disorders affecting the heart or blood vessels, diagnosed through clinical evaluation, imaging, and relevant biomarkers. Heart disease referred specifically to clinically diagnosed cardiac conditions, such as coronary artery disease or heart failure, diagnosed based on clinical criteria, imaging studies, and laboratory tests. HIV/AIDS diagnosis was confirmed through serological tests (e.g. enzyme-linked immunosorbent assay (ELISA)) for HIV infection and specific clinical criteria for AIDS with opportunistic infections. Anemia was defined as hemoglobin levels below the normal range depending on age and sex, typically assessed through a complete blood count (CBC). COVID-19 diagnosis was confirmed through polymerase chain reaction (PCR) or rapid antigen tests for severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), along with a clinical evaluation of symptoms. These definitions align with the protocols established at JMC. Statistical methods A continuous-time multistate Markov process model was used to analyze transitions between CKD stages over time. The model comprised five discrete disease states corresponding to the CKD stages as shown in Figure 1. Although disease progression occurs continuously over time, the state space is discrete, making the continuous-time Markov process (CTMP) framework appropriate for modeling transitions between disease stages. Under the Markov assumption, the probability of a future transition depends only on the patient’s current state. This framework enables the analysis of longitudinal time-to-event data by estimating transitions between transient states and the absorbing state. Compared with conventional statistical approaches, the continuous multistate Markov process model simultaneously estimates multiple transition intensities between stages, such as from stage 1 to stage 2 CKD or stage 3 to ESRD stages while quantifying transition probabilities and mean sojourn times.19,28 In addition, the model incorporates individual-level covariates to evaluate their effects on transition intensities, thereby facilitating the identification of factors associated with CKD progression and supporting the identification of patients at high risk of disease progression. 20 Transition probability.Let S(t) denote the state occupied by a patient at time t. The transition probability, Pij(s, t), is defined as the probability that a patient is in state i at time t, given that the patient was in state j at time s. 29 The transition probability matrix, P(s, t), with transition elements Pij(s, t), i, j = 1, 2, … ,5, is used to specify these probabilities. Given that CKD progresses irreversibly into various states, Pij(s, t) = 0 for all j > i. In contrast, the P55(s, t) = 1 in Stage 5 of CKD is an absorbing condition. Each row of the transition probability matrix sums to one, and all of the probabilities must be greater than or equal to zero; that is, pij(s, t) ≥ 0, ∀i, j ∈ 1, 2, …, 5. The transition probability matrix is time-dependent and is therefore denoted as P(t) instead of P. In time-homogeneous Markov models, the dependency of t is omitted. Transition intensity Assuming that a patient is in state (t) = i at time t, the transition intensity from state i to state j over a small time interval (t + σt) is defined as follows: 28 (1) λij=limσt→0P(S(t+σt)=j)∣S(t)=i)σt;i,j=1,2,…,5. Q(λ)=[−(λ12+λ13+λ14+λ15)λ12λ13λ14λ150−(λ23+λ24+λ25)λ23λ24λ2500−(λ34+λ35)λ34λ35000−(λ45)λ4500000] where λij are the transition intensities from state i to j. 28 The elements of Q matrix, λij; i, j = 1, 2, …,5 are model parameters, which control the movement to next possible states. Also, specification Q requires the sum of each row as 0 and the diagonal elements of Q become λii = −Σi ≠j λij(t). Sojourn time The mean sojourn time represents the average length of time a patient remains in a given transient state before transitioning to the next disease state. In a continuous-time multistate Markov process, the mean sojourn time can be estimated from the transition intensities or, equivalently, from the transition intensity matrix. Because CKD progression is assumed to be irreversible, the mean sojourn time corresponds to the expected duration of a patient’s stay in each transient disease stage before progressing to a more advanced stage. Under the Markov assumption, future transitions depend only on the patient’s current state and are independent of previous transitions. For a time-homogeneous CTMP model, the sojourn time in state i follows an exponential distribution with rate −λii. The probability of transitioning from the current state i to the next state j, referred to as P-next probability, is determined by the corresponding transition intensities −λijλii , and the remaining transition probabilities in the ith row of the transition intensity matrix are proportional to −λijλii . 30 Model estimation and diagnostics Model parameters were estimated using the maximum likelihood method. The likelihood function was constructed from the observed transitions between disease states over time and is expressed as follows: (2) L(Q)=∏i,jLi,j=∏i=1nPS(tij)S(ti,j+1)(ti,j+1−tij),i,j=1,2,…,5, where Li,j denotes the element corresponding to the S(tij)th row and the S(ti,j + 1)th column of the transition probability matrix P(t) evaluated over the interval ti,j + 1 − tij. The observed transition data were used to estimate the transition intensities λij. The likelihood L(Q) was maximized in terms of log(λij), therefore the estimates and standard error of λij were obtained from log (λij) using optimization technique. 30 The estimated transition probability matrix was derived from the transition intensity matrix using the matrix exponential, P(t) = exp(Q(t)), where “exp” denotes the matrix exponential, 30 Q is the transition intensity matrix, and t is the elapsed time. This is established using the forward Kolmogorov differential equation and canonical decomposition, which is expressed as follows: ddtP(s,t)=P(s,t)×Q, with the initial condition P(s, s) = I, where I denotes the identity matrix. Solving the system of forward differential equations yields the transition probability matrix in terms of the transition intensity matrix. To evaluate the effects of covariates, transition intensities were modeled using a proportional hazards framework. Under the multistate Markov model, covariates modify the transition intensities between disease states. The proportional hazards model expresses the transition intensity for an individual r as the product of a baseline transition intensity and an exponential function of the covariates. Accordingly, the transition intensity matrix Q was incorporated into the likelihood function to estimate transition intensities. Subsequently, the covariate effects on the transition intensity, for instance xrp for an individual r at time p, were investigated by replacing the transition intensity element λij as follows: λij(t∣X(t))=λij(0)⋅eβij′X(t), where X(t) is a vector of covariates at time t, β is the vector of regression coefficients associated with those covariates, and λij(0) is the baseline hazard function that represents the hazard rate for a subject with all covariates equal to zero. Model fit was evaluated using the Akaike information criterion (AIC), which was calculated as AIC=2p−2ln(L^) , where p is the number of estimated model parameters and L^ is the maximized likelihood defined in Equation (2). 20 Among competing models, the model with the smallest AIC value was considered to provide the best balance between model fit and complexity. The likelihood ratio test (LRT) was used to compare nested models by evaluating whether the more complex model provided a significantly better fit than the reduced model.13,20 Following data preparation, all statistical analyses were performed using R statistical software (version 4.2.1) with the ‘mstate’ and ‘msm’ packages. Results Demographic and clinical characteristics of patients with CKD A total of 194 patients with CKD were included in this study, of whom 106 (54.6%) were male. The mean (SD) age was 45.9 (14.9) years. The most prevalent comorbidities were HTN (92.7%), diabetes mellitus (90.2%), and anemia (85.1%), whereas COVID-19 was the least common comorbidity, affecting 25.8% of the study population (Table 1). Table 1.Summaries of the categorical variables for CKD patients at JMC between February 2019 and February 2024.VariablesCategoriesn (%)SexMale106 (54.6)Female88 (45.4)ResidenceUrban80 (41.2)Rural114 (58.8)DiabetesYes175 (90.2)No19 (9.8)HypertensionYes180 (92.7)No14 (7.3)Heart diseaseYes73 (37.6)No121 (62.4)SmokingYes52 (26.8)No142 (73.2)Cardiovascular diseaseYes108 (55.6)No86 (44.4)HIV/AIDSYes67 (34.5)No127 (65.5)AnemiaYes165 (85.1)No29 (14.9)COVID-19Yes50 (25.8)No144 (74.2)CKD: chronic kidney disease; JMC: Jimma Medical Center; HIV/AIDS: human immunodeficiency virus/acquired immunodeficiency syndrome; COVID-19: coronavirus disease 2019. Summaries of the categorical variables for CKD patients at JMC between February 2019 and February 2024. CKD: chronic kidney disease; JMC: Jimma Medical Center; HIV/AIDS: human immunodeficiency virus/acquired immunodeficiency syndrome; COVID-19: coronavirus disease 2019. The mean laboratory values for serum urea, potassium, hemoglobin, and sodium were 176.4 mg/dL, 4.6 mmol/L, 9.3 g/dL, and 137.1 mmol/L, respectively (Table 2). These findings suggested that the patients had varying degrees of renal impairment, uremia, electrolyte imbalance, and anemia, which are common features of CKD. Table 2.Summaries of the continuous covariates (unit of measurements: mg/dL for urinary protein concentration; mmol/L for potassium and sodium; g/dL for hemoglobin).VariablesMean95% CIAge in years45.9(45.1–46.6)Urinary protein concentration176.4(170.6–182.2)Potassium4.6(4.5–4.70)Hemoglobin9.3(9.2–9.4)Sodium137.1(134.8–137.6)CI: confidence interval. Summaries of the continuous covariates (unit of measurements: mg/dL for urinary protein concentration; mmol/L for potassium and sodium; g/dL for hemoglobin). CI: confidence interval. Transition states of CKD progression The observed transitions of patients with CKD among the five disease states during follow-up were summarized using transition counts, transition probabilities, and transition intensities, as presented in Table 3. Table 3.Counts of the transition states, estimated transition probability, and intensity for CKD progression at one month time.TransitionTostatesFromState 1State 2State 3State 4State 5TransitionState 112268201911CountsState 20119831813State 30020010024State 4000168120State 50000227TransitionState 10.7940.1430.0180.0320.012ProbabilityState 20.0000.7980.1670.0170.018State 30.0000.0000.8410.1380.021State 40.0000.0000.0000.8030.197State 50.0000.0000.0000.0001.000TransitionState 1−0.2310.1790.0040.0390.008IntensityState 20.000−0.2250.2040.0040.018State 30.0000.000−0.1730.1680.005State 40.0000.0000.000−0.2190.219State 50.0000.0000.0000.0000.000CKD: chronic kidney disease. Counts of the transition states, estimated transition probability, and intensity for CKD progression at one month time. CKD: chronic kidney disease. Transition counts. The transition counts presented in Table 3 summarized the observed movements of patients between the five CKD stages during follow-up. The entries represent the number of observed transitions from state i to state j, where i, j = 1, 2, …,5. State 1 had the highest number of self-transitions (122), indicating that patients most frequently remained in the same disease stage between consecutive visits. Forward transitions from state 1 to states 2, 3, 4, and 5 were observed in 68, 20, 19, and 11 instances, respectively. Similarly, patients in state 2 most frequently transitioned to state 3 (83 transitions), followed by transitions to states 4 (18 transitions) and 5 (13 transitions). In state 2, 200 self-transitions were recorded, with approximately 50% and 13% observed transitions to states 4 (100 ) and 5 (24 ), respectively. This may indicate a high likelihood of progression to more severe stages of CKD. Transitions from state 4 to state 5 were notably observed 168 times, highlighting the severity of the condition and the likelihood of further deterioration. State 5 represents end-stage CKD, with 227 self-transitions, indicating irreversible progression. No transitions from state 5 to earlier disease stages were observed, confirming the chronic and progressive nature of the disease at this stage (Table 3). Transition probabilities The transition probabilities presented in Table 3 represent the estimated probabilities of moving between CKD stages over a 1-month interval. The estimated probabilities of remaining in the same disease stage after 1 month were 0.794, 0.798, 0.841, and 0.803 for stages 1, 2, 3, and 4, respectively. These findings indicate that patients were more likely to remain in their current stage than to transition to another stage during the 1-month period. The estimated probabilities of transitioning from stages 1, 2, 3, and 4 to the subsequent stage were 0.143, 0.167, 0.138, and 0.197, respectively. Direct transitions from stages 1, 2, and 3 to stage 5 were uncommon, with estimated probabilities of 0.012, 0.018, and 0.021, respectively, whereas patients in stage 4 had a substantially higher probability (0.197) of progressing to stage 5. Overall, the estimated transition probabilities indicate that disease progression occurred gradually, with patients tending to remain in their current stage over a 1-month period while exhibiting an increasing likelihood of progression to more advanced stages. Following model fitting, the estimated transition probability matrices for 12, 24, and 36 months are presented in Table 4. For patients initially in state 1, the estimated probability of progressing to state 5 increased over time, reaching 0.959 at 36 months. Correspondingly, the probabilities of remaining in state 1 or transitioning to intermediate states (states 2, 3, and 4) decreased over time. Similar patterns were observed for patients initially in states 2, 3, and 4, with progressively higher probabilities of reaching stage 5 as the follow-up period increased. For example, the estimated probability of transition from state 2 to state 5 increased from 0.478 at 12 months to a value approaching 1.0 at 36 months. Across all initial disease stages, the probability of remaining in the same state decreased with increasing follow-up time, whereas the probability of progressing to state 5 increased. Overall, the estimated transition probability matrices demonstrate progressive movement toward the absorbing state over time. Table 4.Estimated P-next transition probabilities over different time periods (months).TransitionYear 1Year 2Year 31 → 10.0630.0040.0011 → 20.1380.0180.0021 → 30.2210.0720.0121 → 40.1830.0970.0261 → 50.3950.8090.9592 → 20.0680.0050.0012 → 30.2310.0430.0062 → 40.2230.0750.0142 → 50.4780.8770.9793 → 30.1210.0150.0023 → 40.1910.0360.0053 → 50.6880.9490.9934 → 40.0700.0050.0014 → 50.9300.9950.999 Estimated P-next transition probabilities over different time periods (months). Transition intensities. The estimated transition intensities between CKD stages are presented in Table 3. The diagonal elements of the transition intensity matrix represent the total rate of leaving each state and are therefore negative, whereas the off-diagonal elements represent the transition intensities between disease stages. For state 1, the estimated transition intensity was 0.179 to stage 2, 0.039 to stage 4, and 0.008 to stage 5, with a diagonal element of −0.231. For stage 2, the estimated transition intensities were 0.204 to stage 3, 0.004 to stage 4, and 0.018 to stage 5, with a diagonal element of −0.225. Patients in stage 3 had transition intensities of 0.168 to stage 4 and 0.005 to stage 5, with a diagonal element of −0.173. For stage 4, the transition intensity to stage 5 was 0.219, with a corresponding diagonal element of −0.219. As expected for an absorbing state, stage 5 had no outgoing transition intensities. Overall, the estimated transition intensity matrix indicates forward progression between disease stages, with state 5 functioning as the absorbing state in the model. Estimation of mean sojourn times and P-next probabilities The mean sojourn time represents the average duration that a patient is expected to remain in a given CKD state before transitioning to the next disease state. Mean sojourn times were estimated from the transition intensity matrix and are presented in Table 5. Patients in states 1, 2, 3, and 4 remained in these stages for an estimated mean of 4.33, 4.44, 5.79, and 4.57 months, respectively, before progressing to a more advanced state. Among the transient states, state 3 had the longest estimated mean sojourn time (5.79 months), indicating that patients remained in this stage longer than in the other disease stages before progression. The standard errors (SEs) quantify the precision of the estimated mean sojourn times, whereas the 95% confidence intervals (CIs) indicate the range of values within which the true mean sojourn times are expected to lie. Table 5.Mean sojourn times in transient states.StatesMean sojourn time (in months)SE95% CIState 14.330.537(3.39–5.52)State 24.440.385(3.75–5.26)State 35.790.456(4.96–6.76)State 44.570.530(3.64–5.74)CI: confidence interval; SE: standard error. Mean sojourn times in transient states. CI: confidence interval; SE: standard error. The estimated P-next probabilities are presented in Table 6. Each value represents the probability that the next observed transition from a given state will be to a particular subsequent state. They probabilities differ from the transition probabilities because they describe the distribution of the next transition, rather than the probability of occupying a given state after a specified period. Together, the mean sojourn times and P-next probabilities complement the transition intensities by summarizing the expected time spent in each disease state and the probabilities of subsequent transitions. Table 6.Probability of the transition to the next higher health state, conditional on leaving the current state (P-next probabilities).State 1State 2State 3State 4State 5State 100.8970.0190.01680.0672State 2000.9040.0170.079State 30000.9720.028State 400001State 500000The values in Table 6 represent the P-next probabilities, where each probability denotes the probability of transition from state k to the next state l, conditional on leaving the state k. Probability of the transition to the next higher health state, conditional on leaving the current state (P-next probabilities). The values in Table 6 represent the P-next probabilities, where each probability denotes the probability of transition from state k to the next state l, conditional on leaving the state k. For example, the estimated P-next probabilities of progressing from states 1, 2, and 3 to states 2, 3, and 4 were 89.7%, 90.4%, and 97.2%, respectively (Table 6). The survival curves for patients in the different CKD stages are presented in Figure 2. Patients in states 1 and 2 exhibited higher survival probabilities over time than those in the more advanced stages. In contrast, patients in state 4 showed a marked decline in survival probability during the third month, indicating a significant decline in kidney function. Furthermore, the 12-month estimated survival probabilities were approximately 0.1, compared to 0.4 for moderate CKD, 0.6 for mild CKD, and 0.7 for normal CKD. Plot of fitted survival probability for each state. Comparison of multistate Markov models for CKD progression Several continuous-time multistate Markov process models incorporating different combinations of covariates were fitted to identify the model that best described CKD progression. Model building was performed sequentially by first fitting models containing a single covariate and subsequently fitting multivariable models comprising clinically relevant covariates. Model performance was evaluated using the AIC and the LRT. The comparison of the fitted models is presented in Table 7. Table 7.Comparison of candidate multistate Markov models using likelihood ratio tests, AIC values, and P-values.CovariatesAICLRTDfP-valueModels with No covariate2184.5---Models with Sex2200.73.82100.955 Residence2196.18.41100.589 Age22032.5100.991 Potassium2266.12.44100.655 Urinary protein concentration22024310<0.001a Hemoglobin218342.910<0.001a Sodium217134.210<0.001a Diabetes215550.910<0.001a HTN21485710<0.001a Heart disease22031.99100.996 Smoking21996.14100.804 CVD218619.7100.032 HIV/AIDS2186.617.97100.055 Anemic2124.180.3910<0.001a COVID-192189.814.67100.145 Urinary protein concentration + potassium2140106.46400.002 Diabetes + HTN + CVDa2135110.0630<0.001a Diabetes + HTN + CVD + anemic2141104.840<0.001a All significant variables as a covariates215810180<0.001a All variables as a covariates2176113.61580.051aThe best-fitting model based on the model selection criteria.All significant variables contains seven covariates: urinary protein, hemoglobin, sodium, diabetes, hypertension, CVD, and anemia.AIC: Akaike information criterion; CVD: cardiovascular disease; Df: degrees of freedom; HTN: hypertension; LRT: likelihood ratio test; COVID-19: coronavirus disease 2019; HIV/AIDS: human immunodeficiency virus/acquired immunodeficiency syndrome. Comparison of candidate multistate Markov models using likelihood ratio tests, AIC values, and P-values. The best-fitting model based on the model selection criteria. All significant variables contains seven covariates: urinary protein, hemoglobin, sodium, diabetes, hypertension, CVD, and anemia. AIC: Akaike information criterion; CVD: cardiovascular disease; Df: degrees of freedom; HTN: hypertension; LRT: likelihood ratio test; COVID-19: coronavirus disease 2019; HIV/AIDS: human immunodeficiency virus/acquired immunodeficiency syndrome. Among the single-covariate models, urinary protein concentration, hemoglobin, sodium, diabetes mellitus, HTN, and anemia were significantly associated with improved model fit compared with the model without covariates (P < 0.05). The single-covariate model including HTN had the lowest AIC value (2148), indicating the best fit among the single-covariate models. Among the multivariable models, the model including diabetes mellitus, HTN, and cardiovascular disease showed statistically significant covariate effects (P < 0.05) and the lowest AIC value (2135), indicating the best fit among the multivariable models evaluated. The comparison of candidate models demonstrated that the multivariable model including diabetes mellitus, HTN, and cardiovascular disease (CVD) provided the best fit to the data according to the model selection criteria (Table 7). Consequently, this model was selected as the final model for evaluating CKD progression. The results indicate that these comorbidities were associated with differences in transition intensities between disease stages. Covariate effects on transition intensities during CKD progressions The estimated hazard ratios (HRs) and corresponding 95% CIs for the effects of diabetes mellitus, HTN, and CVD on the transition intensities between CKD stages are presented in Table 8. An HR >1 indicates an increased hazard of transitioning from one disease state to another associated with the covariate, whereas a 95% CI that does not include 1 indicates statistical significance. Table 8.HRs and 95% CIs for the best model.StatesCovariatesHypertensionDiabetesCVD1 → 21.53 (0.62–3.77) 1.73 (1.67–3.49) 1.08 (0.59–1.97)1→32.41(0.02–6.10)2.10 (0.001–5.5)1.13 (0.006–8.9)1→41.41 (0.03–5.20)1.41 (0.03–7.10)1.59 (0.26–9.85)1→51.41 (0.02–9.60)1.42 (0.02–9.80)1.46 (0.16–4.80)2→31.22 (0.60–2.45)1.16 (0.57–2.35)1.27 (0.87–1.86)2→42.01 (0.01–6.90)1.63 (0.03–5.90)1.33 (0.11–5.50)2→52.04 (0.08–8.50)1.76 (0.06–5.50)1.52 (0.31–7.55)3→41.50 (0.83–2.71)0.93 (0.52–1.67)1.67 (0.99–2.65)3→51.43 (0.02–4.70)1.14 (0.02–5.10)2.03 (0.06–5.27)4→5 2.62 (1.79–3.84) 1.34 (0.93–1.93) 2.01 (1.44–2.81) Boldface values indicate hazard ratios with 95% CIs that do not include 1.CI: confidence interval; HR: hazard ratio; CVD: cardiovascular disease. HRs and 95% CIs for the best model. Boldface values indicate hazard ratios with 95% CIs that do not include 1. CI: confidence interval; HR: hazard ratio; CVD: cardiovascular disease. For the transition from stage 1 to stage 2, diabetes mellitus was associated with a significantly higher hazard of progression (HR = 1.73; 95% CI: 1.67–3.49) than the absence of diabetes mellitus. Although the estimated HRs varied across transitions and covariates, most transition-specific estimates had 95% CIs that included 1, indicating insufficient statistical evidence of an association. In contrast, HTN was significantly associated with progression from state 4 to state 5 (HR = 2.62, 95% CI: 1.79–3.84). Similarly, CVD was significantly associated with the transition from stage 4 to stage 5 (HR = 2.01, 95% CI: 1.44–2.81) (Table 8). Discussion In this study, a multistate Markov model was used to characterize the progression of CKD among patients followed at JMC between February 2019 and February 2024, to aid in understanding CKD progression and identify factors that influence kidney deterioration. Men accounted for a slightly higher proportion of the study population, with 106 (54.6%) of the 194 patients being men. This finding is consistent with previous studies reporting that men have a greater risk of CKD progression, kidney failure, and albuminuria than women. These differences have been attributed to biological mechanisms, including the potential protective effects of estrogen in women as well as behavioral and lifestyle factors such as smoking, alcohol consumption, and delayed healthcare-seeking behavior among men. 31 However, other studies have reported a higher prevalence of CKD among women. These discrepancies may reflect differences in study populations, healthcare access, risk factor distributions, or the distinction between disease prevalence and disease progression. 32 The multistate Markov analysis showed that patients experienced more self-transitions than transitions to more advanced disease stages during follow-up, indicating that most patients remained in the same CKD state between consecutive visits. This finding is consistent with the relatively slow progression of CKD and may reflect the effects of regular clinical follow-up and management of comorbidities such as HTN and diabetes mellitus. It may also indicate that sufficient decline in eGFR to meet the threshold for progression to the next disease stage often occurs over a prolonged period. However, because this was a retrospective study, the factors contributing to the observed stability could not be fully evaluated. These findings differ from those reported by Lintu et al. 23 but are consistent with previous studies by Grover, Sabharwal et al. 13 and Dalip et al. 33 The discrepancy may be attributable, at least in part, to differences in the modeling approach. Specifically, Lintu et al. 23 allowed reverse transitions between CKD stages, whereas the present study assumed irreversible disease progression in accordance with the clinical course of CKD. Furthermore, the present study demonstrated an increasing probability of progression to more advanced CKD stages over time, particularly to state 5. Patients in state 4 had the highest estimated probability of progressing to state 5 (0.197) over the 1-month interval, indicating an increased likelihood of reaching the absorbing state compared with patients in the earlier disease stages. This finding is consistent with the natural history of CKD, in which advanced disease stages are associated with a greater risk of progression to ESRD. The increased probability of progression from stage 4 to stage 5 is biologically plausible because stage 4 is characterized by severely reduced kidney function and markedly diminished renal reserve. Under these conditions, additional insults, such as uncontrolled HTN, diabetes mellitus, infection, or acute kidney injury, may accelerate progression to ESRD. Similar patterns have been reported in previous studies, which showed that advancing CKD states are associated with progressive nephron loss, compensatory hyperfiltration, renal fibrosis, and an increased risk of ESRD.13,23,20,25 Consistent with these findings, the estimated P-next probability matrices at 12, 24, and 36 months demonstrated an increasing probability of progression to state 5 with longer follow-up. Together, these results underscore the importance of early detection and timely management of CKD and its comorbidities to help delay progression to ESRD. The estimated mean sojourn times indicated that patients remained in state 3 for an average of 5.79 months before progressing to state 4, whereas the estimated mean sojourn time for state 4 was 4.57 months before progression to the absorbing state (state 5). A similar study conducted in India by Lintu et al. 23 reported a mean sojourn time of 1.40 years (16.8 months) for state 4, which is substantially longer than that observed in the present study. This difference may reflect variations in the study populations, healthcare settings, follow-up schedules, and modeling approaches. In addition, patients included in this hospital-based cohort may have presented with more advanced disease, resulting in shorter durations within each CKD stage before further progression. Differences in healthcare resources and long-term follow-up practices between the two settings may also have contributed to the observed variation in sojourn times. The estimated mean sojourn times and P-next probabilities complement the transition intensities by summarizing the expected duration spent in each disease stage and the probabilities of subsequent transitions, thereby facilitating the clinical interpretation of disease progression. To assess the influence of clinical factors on CKD progression, several continuous-time multistate Markov models incorporating different combinations of covariates were compared. The model including diabetes mellitus, HTN, and CVD provided the best fit according to the model selection criteria. Covariates were selected using a sequential approach based on the AIC and LRTs; however, this strategy may introduce model selection uncertainty, particularly in studies with relatively small sample sizes. The superior performance of this multivariable model is biologically plausible because these three comorbidities are closely interconnected and may collectively reflect cumulative systemic damage. Diabetes mellitus contributes to glomerular injury, HTN accelerates nephron loss by increasing intraglomerular pressure, and CVD reflects vascular dysfunction that may impair renal perfusion. Their combined effects may therefore explain the improved performance of the multivariable model. Although several demographic, clinical, and biochemical variables were evaluated during model development, only diabetes mellitus, HTN, and CVD remained significantly associated with the transition intensities in the final model. Several studies have reported that diabetes mellitus, HTN, and CVD are important determinants of CKD progression. Grover et al. 13 demonstrated that diabetes mellitus and HTN accelerate disease progression, supporting the inclusion of these comorbidities in predictive models. Similarly, associations between CVD and CKD progression have been reported by Uddin et al. 34 and Webster et al. 35 Other factors, including smoking, sex, place of residence, HIV/AIDS, COVID-19, and laboratory variables such as hemoglobin, serum creatinine, sodium, urea, and anemia, have also been identified as potential predictors of CKD progression.10,11,36,37,38 However, these variables were not retained in the final multivariable model, suggesting that their independent effects may have been attenuated in this hospitalized cohort or explained by other clinical risk factors. The estimated transition-specific HRs from the multistate Markov model further demonstrated that diabetes mellitus, HTN, and CVD influenced progression between CKD stages in this cohort. According to the estimated HRs, patients with diabetes mellitus had a significantly higher hazard of progressing from stage 1 to stage 2 than those without diabetes mellitus (HR = 1.73, 95% CI: 1.67–3.49). This finding suggests that diabetes mellitus may contribute to disease progression during the early stages of CKD and highlights the importance of optimal glycemic management to help delay CKD progression. Similar findings have been reported by Taha and Mohammed (2023), 25 who analyzed data from a military hospital in Iraq, as well as by other studies using multistate or hidden Markov models.13,20,34 The association is biologically plausible because chronic hyperglycemia promotes glomerular and vascular injury, renal fibrosis, and increased intraglomerular pressure, ultimately contributing to progressive loss of kidney function. HTN was significantly associated with progression from stage 4 to stage 5 (HR = 2.62, 95% CI: 1.79–3.84), indicating an increased hazard of progression to ESRD. This finding is consistent with previous reports demonstrating that uncontrolled HTN accelerates CKD progression.13,20,25,23 Persistent HTN contributes to nephron loss by damaging the renal vasculature, increasing intraglomerular pressure, and promoting nephrosclerosis, thereby facilitating progression to ESRD. CVD was significantly associated with progression from stage 4 to stage 5 (HR = 2.01, 95% CI: 1.44–2.81). This association is biologically plausible because CVD and CKD share similar pathogenic mechanisms, including endothelial dysfunction, chronic inflammation, oxidative stress, and vascular calcification. These processes may impair renal perfusion, promote nephron injury, and accelerate progression ESRD. However, because these associations were transition-specific, they should not be interpreted as applying uniformly across all stages of CKD progression. Further studies are warranted to investigate the biological mechanisms linking these comorbidities with disease progression. This study has several limitations. First, incomplete patient records and missing follow-up information for key clinical variables resulted in the exclusion of some patients, which may have introduced selection bias. Second, stages 3A and 3B were combined into a single stage because of the limited sample size, potentially reducing the ability to detect differences in progression within stage 3. Finally, because albuminuria data were not routinely available, the full Kidney Disease: Improving Global Outcomes (KDIGO) eGFR–albuminuria risk classification could not be applied, and disease staging was based solely on eGFR. Additionally, detailed longitudinal clinical information, including treatment modality, medication use, blood pressure and glycemic control, proteinuria progression, dialysis initiation, kidney transplant referral, and mortality, was not consistently available in the retrospective medical records. Consequently, these variables and potential competing risks could not be incorporated into the analyses. External validation, internal cross-validation, and comparisons with alternative modeling approaches, such as competing-risk models and joint longitudinal–survival models, were beyond the scope of this single-center retrospective study and should be explored in future research. Moreover, the assumptions underlying the continuous-time Markov model were not formally assessed. Future studies should also consider hidden Markov or misclassification models to account for potential measurement error and temporary fluctuations in eGFR that may result in apparent reversals of CKD stage. Conclusion This study demonstrated that the probability of progression to more advanced stages of CKD increased over time, with patients in the later disease stages having the greatest likelihood of progressing to the absorbing state (stage 5). The estimated transition probabilities, transition intensities, mean sojourn times, and P-next probabilities provided complementary information on disease progression and characterized the dynamics of transitions between CKD stages. Diabetes mellitus, HTN, and CVD were identified as the principal clinical factors associated with transition intensities in the final multistate model, highlighting their potential importance in CKD progression. These findings support the value of early detection, close monitoring, and appropriate management of patients with CKD, particularly those with these comorbidities. Future multicenter studies with external validation and more comprehensive longitudinal clinical data are warranted to further validate and extend these findings.