Estimating Chronic Kidney Disease Stage Transitions from Irregular Electronic Health Record Data Using an Expectation-Maximization Framework

This study demonstrates that an Expectation-Maximization framework effectively estimates clinically plausible chronic kidney disease stage transition probabilities from irregular electronic health record data in patients with small renal masses, offering a robust alternative to naive estimators for use in decision-analytic and health economic models.

The Gist

Imagine you are trying to map the journey of a hiker climbing a mountain, but you only get to see them at random times. Sometimes you see them at the base camp, and three days later you see them at the summit. Other times, you see them at the summit, and six months later you see them back at the base camp.

If you just drew a straight line between those two points, you might think the hiker teleported up the mountain instantly or fell down a cliff. But in reality, they probably took a winding path, maybe rested at a mid-point, or had a bad day where they slipped a little before recovering.

This is exactly the problem researchers faced with Chronic Kidney Disease (CKD).

The Problem: The "Missing Pages" in the Medical Story

Kidney disease is like a slow, steady decline in a car's engine performance. Doctors measure this using a score called eGFR (how well the kidneys filter blood). Based on this score, patients are put into "Stages" (Stage 1 is healthy, Stage 5 is kidney failure).

In the real world, patients don't visit the doctor on a perfect schedule.

• The Naïve Approach: If a doctor sees a patient in Stage 3 today and Stage 2 next month, a simple computer program might say, "Aha! The patient got better!" and record a transition from Stage 3 to Stage 2.
• The Reality: The patient likely didn't get better. They probably had a temporary illness, took a new medication, or had a bad lab test. The "Stage 2" reading was just a glitch or a temporary dip. The patient actually stayed in Stage 3 or got slightly worse, but the doctor missed the middle steps.

When researchers tried to predict the future of kidney disease using these "naïve" methods, the data was full of noise. It looked like patients were jumping up and down the stages randomly, making it impossible to build a reliable map for the future.

The Solution: The "Sherlock Holmes" Algorithm

The researchers (led by Wendy Qi and colleagues) used a clever mathematical tool called the Expectation-Maximization (EM) algorithm.

Think of the EM algorithm as a Sherlock Holmes detective.

1. The Clues (E-Step): The detective looks at the two points where the patient was seen (e.g., "Stage 3 in January" and "Stage 4 in July").
2. The Guess (M-Step): Instead of assuming the patient jumped straight from 3 to 4, the detective asks: "What is the most likely path they took in those six months?"
   • Did they stay at Stage 3 for five months and then slip to 4?
   • Did they slowly drift from 3 to 3.5 to 4?
   • Did they have a temporary "Stage 2" blip that wasn't real?

The algorithm runs this simulation thousands of times, filling in the "missing pages" of the medical story with the most probable scenarios. It essentially says, "We didn't see the patient in February, March, April, May, and June, but based on the laws of kidney disease, here is what probably happened."

What They Found

By using this "detective" method on data from 527 patients with small kidney tumors, they found a much clearer picture:

• The "Stay Put" Reality: Most patients don't jump stages. They tend to stay in the same stage for a long time. The "naïve" method made it look like patients were bouncing around wildly, but the EM method showed that stability is the norm.
• No Magic Reversals: The simple method saw many patients "improving" (going from Stage 4 back to Stage 3). The EM method realized these were likely just temporary fluctuations. It smoothed out the noise, showing that once kidneys get worse, they rarely get significantly better on their own.
• Age Matters, Gender Doesn't: The detective found that older hikers (patients) were slightly more likely to slide down the mountain faster than younger ones. However, being a man or a woman didn't really change the path.

Why This Matters

This study is like upgrading from a crumpled, hand-drawn sketch to a high-definition GPS map.

Before, doctors and health planners had to guess how kidney disease would progress because the data was too messy. Now, they have a reliable set of "transition probabilities."

• For Doctors: It helps them predict how a patient's kidneys will likely fare over the next 5 or 10 years, helping them decide whether to operate on a tumor now or wait and watch.
• For Policymakers: It helps them calculate the cost of care and plan for the future, knowing exactly how many people will likely move from "mild" to "severe" kidney disease.

The Bottom Line

The researchers didn't just count the dots on the graph; they connected the dots with logic and math to fill in the gaps. By using this "Expectation-Maximization" framework, they turned messy, irregular hospital records into a clear, trustworthy story about how kidney disease actually progresses, helping everyone make better decisions for patient care.

Technical Summary

1. Problem Statement

The study addresses the challenge of estimating Chronic Kidney Disease (CKD) stage transition probabilities using real-world Electronic Health Record (EHR) data. Key methodological hurdles include:

• Irregular Observation Intervals: Clinical data is collected at patient-specific, non-uniform time points, violating the fixed-interval assumption of standard discrete-time Markov models.
• Interval Censoring: The exact timing of stage transitions between visits is unknown. Patients may transition through multiple unobserved states between two measurements.
• Bias in Naïve Estimators: Simple counting methods (naïve estimators) that only count observed transitions between consecutive visits often conflate short-term laboratory fluctuations (e.g., transient eGFR drops due to acute illness) with true disease progression. This leads to:
  • Overestimation of "backward" transitions (improvement) that are actually noise.
  • Underestimation of true progression rates.
  • Selection bias when excluding patients with sparse data.
• Specific Population Context: The study focuses on patients with Small Renal Masses (SRMs), a group where preserving kidney function is critical. Existing general CKD models may not accurately reflect the progression dynamics of this specific subpopulation, which often involves older patients with high comorbidity burdens.

2. Methodology

The authors propose a Discrete-Time Markov Model estimated via an Expectation–Maximization (EM) algorithm to handle incomplete and irregularly spaced data.

A. Data Source and Cohort

• Source: University of Virginia (UVA) Small Renal Mass (SRM) registry (2006–Jan 2026).
• Cohort: 527 patients with at least two valid outpatient eGFR measurements prior to definitive treatment (or throughout follow-up for untreated patients).
• Exclusions: Inpatient measurements were excluded to avoid transient fluctuations; patients with fewer than two measurements were excluded.
• Staging: CKD stages (1–5) were defined using KDIGO eGFR thresholds.

B. The EM Algorithm Framework

The core innovation is adapting the EM algorithm (specifically following Sherlaw-Johnson et al.) to reconstruct unobserved intermediate transitions.

1. Model Definition: A discrete-time Markov chain with 6 states (CKD 1, 2, 3a, 3b, 4, 5). The goal is to estimate the one-step transition probability matrix $P$.
2. E-Step (Expectation):
   • Treats unobserved transitions between measurement times as latent variables.
   • Computes the expected number of one-step transitions ($S_{ij}$) between stages $i$ and $j$, conditional on the observed start and end states and the time elapsed ($t$).
   • This involves summing over all possible latent paths weighted by their probabilities:
     $$S_{ij}(P^{(p)}) = \sum_{m,n,t} O_{mnt} \sum_{k=0}^{t-1} \frac{(P^{(p)})_{mi}^k P_{ij}^{(p)} (P^{(p)})_{jn}^{t-k-1}}{(P^{(p)})_{mn}^t}$$
3. M-Step (Maximization):
   • Updates the transition matrix $P$ by normalizing the expected transition counts:
     $$P_{ij}^{(p+1)} = \frac{S_{ij}(P^{(p)})}{\sum_k S_{ik}(P^{(p)})}$$
4. Initialization & Convergence:
   • Initialized with a naïve counting estimator but with small positive values ($10^{-5}$) for clinically unlikely transitions to prevent "locking" (where unobserved transitions remain zero forever).
   • Iterates until the change in probabilities is $< 10^{-5}$.
   • Implemented in Python using NumPy; computationally efficient (~21–37 seconds).

C. Validation

• Likelihood Ratio Test: Compared the log-likelihood of the EM-estimated model against a naïve empirical estimator.
• Subgroup Analysis: Stratified results by Age (<65 vs. ≥65) and Sex.
• Cycle Lengths: Tested robustness using 3-month and 6-month cycle lengths.

3. Key Results

• Transition Structure: The EM framework produced clinically plausible matrices dominated by self-transitions (staying in the same stage) and adjacent-stage progression.
• Reduction of Spurious Backward Transitions:
  • The naïve estimator showed significant "backward" transitions (e.g., Stage 3a $\to$ 2), attributed to short-term eGFR noise.
  • The EM framework significantly reduced these spurious improvements by averaging over the irregular intervals, effectively filtering out transient noise.
• Robustness to Cycle Length: Transition patterns were consistent between 3-month and 6-month cycle lengths. The 6-month matrix appeared slightly smoother, further dampening short-term variability.
• Subgroup Differences:
  • Age: Patients $\ge$ 65 years showed slightly higher probabilities of progressing to advanced stages compared to younger patients.
  • Sex: Minimal differences were observed between males and females.
• Model Fit: Likelihood ratio tests yielded p-values of 0.999, indicating the EM-based models fit the observed data as well as (or better than) the naïve empirical models, without introducing structural bias.

4. Key Contributions

1. Methodological Advancement: Provides a principled, computationally efficient method to estimate discrete-time transition matrices from irregularly observed, interval-censored EHR data without excluding patients with sparse data.
2. Noise Filtering: Demonstrates that EM algorithms can effectively distinguish between true disease progression and short-term biological/measurement variability (transient eGFR fluctuations), reducing the "backward transition" artifact common in naïve counting.
3. Population-Specific Evidence: Generates the first empirically derived, model-ready CKD transition matrices specifically for patients with Small Renal Masses (SRMs), a critical population for active surveillance vs. intervention decision-making.
4. Practical Utility: The resulting transition matrices are directly applicable to Markov cohort models, microsimulations, and Markov Decision Processes (MDPs) for health economic evaluations and clinical decision support.

5. Significance and Implications

• Clinical Decision Making: Accurate transition probabilities are vital for evaluating the trade-offs between cancer control (surgery) and kidney preservation (active surveillance) in SRM patients. This study provides the necessary inputs to model long-term renal outcomes more accurately.
• Real-World Data Utilization: The framework offers a generalizable solution for leveraging "messy" real-world EHR data (irregular visits, missing data) for disease progression modeling, moving beyond the limitations of strictly controlled clinical trial data or rigid survival models.
• Future Research: The study establishes a foundation for extending these methods to other chronic diseases with irregular follow-up and suggests that future work could integrate covariates (e.g., diabetes, comorbidities) directly into the transition probability estimation.

Conclusion: The paper successfully demonstrates that the EM algorithm is a superior alternative to naïve counting for estimating CKD progression from EHR data, yielding stable, clinically realistic transition probabilities that account for the inherent irregularity of real-world clinical practice.