💬 NLP

A Method for Characterizing Disease Progression from Acute Kidney Injury to Chronic Kidney Disease

This study utilizes electronic health record data and multi-state modeling to identify fifteen distinct post-acute kidney injury clinical states, revealing how both established and novel risk factors differentially influence the progression to chronic kidney disease across these trajectories.

Original authors: Yilu Fang, Jordan G. Nestor, Casey N. Ta, Jerard Z. Kneifati-Hayek, Chunhua Weng

Published 2026-04-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yilu Fang, Jordan G. Nestor, Casey N. Ta, Jerard Z. Kneifati-Hayek, Chunhua Weng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your body is a complex city, and your kidneys are the vital water treatment plants keeping everything clean. Sometimes, a sudden disaster hits the plant—a pipe bursts, a power outage occurs, or a toxic spill happens. In medical terms, this is Acute Kidney Injury (AKI). It's a sudden crisis.

Most of the time, the city repairs the plant, and things go back to normal. But for some people, the damage doesn't fully heal. Instead, the plant slowly starts to rust and degrade over years, turning into a permanent problem called Chronic Kidney Disease (CKD).

The big challenge for doctors has always been: How do we know which patients will recover fully and which ones are on a slow, slippery slope toward permanent damage?

This paper is like a team of detectives using a massive, high-tech map (Electronic Health Records) to solve that mystery. Here is how they did it, explained simply:

1. The Old Way vs. The New Way

The Old Way: Imagine trying to predict the weather by looking at a single photo of the sky taken at 9:00 AM. You might see clouds, but you don't know if it's going to rain in an hour or clear up by noon. Previous studies looked at patients at just one moment in time (like their age or if they had diabetes) and guessed their future. This missed the dynamic changes happening day-by-day.

The New Way: This study treats a patient's health like a movie, not a photo. The researchers watched the "movie" of thousands of patients' lives after their kidney injury. They tracked every medication, every lab test, and every symptom over time to see how the story unfolded.

2. Building a "Health GPS"

To make sense of this massive amount of data, the researchers built a digital "GPS" for each patient.

The Map: They took all the medical codes (like "diabetes," "heart failure," or "antibiotics") and turned them into a digital path.
The Compass: They added the patient's creatinine levels (a blood test that measures kidney function) as a second layer of data.
The Result: This created a unique "fingerprint" for every day of a patient's recovery.

3. Discovering the "Neighborhoods" of Recovery

Using a smart computer algorithm (clustering), the researchers looked at all these digital fingerprints and realized that patients didn't all recover the same way. They naturally grouped into 15 distinct "neighborhoods" or "states."

Think of these like different recovery paths in a video game:

State A: The "Quick Recovery" path. The patient gets better fast and stays healthy.
State B: The "Stuck in Traffic" path. The patient has lingering issues like heart trouble or liver problems.
State C: The "Danger Zone." The patient is stable but has specific risk factors that make them likely to crash later.

The study found that most patients (75%) stayed in one neighborhood or made just one move. But for those who moved between neighborhoods, the destination mattered immensely.

4. The Crystal Ball: Predicting the Future

Once they mapped out these 15 neighborhoods, they used math to calculate the odds. They asked: "If a patient is in Neighborhood X today, what are the chances they will develop Chronic Kidney Disease in 5 years?"

The Findings:

The High-Risk Zones: Some neighborhoods had a very high chance of leading to chronic disease. For example, patients in a specific state with heart failure and high blood pressure were much more likely to develop CKD.
The Surprises: The study found some unexpected clues.
- In one group, having a slightly higher temperature on the first day was a warning sign.
- In another group, taking common painkillers (NSAIDs) within a year of the injury was a major risk factor.
- Interestingly, for some patients, having sepsis (a severe infection) at the start actually seemed to protect them from chronic kidney disease later on (a counter-intuitive finding that needs more study).

5. Why This Matters

This isn't just about numbers; it's about personalized care.

Imagine a doctor today. They might say, "You had kidney injury, come back in a year."
With this new method, the doctor could say: "Based on your specific recovery path and your current 'neighborhood,' you are in a high-risk zone. You need to come back in 3 months, and we need to watch your blood pressure closely."

The Bottom Line

This paper is a blueprint for a smart, data-driven early warning system. By treating patient recovery as a dynamic journey rather than a static snapshot, it helps doctors spot the "ticking time bombs" before they explode. It turns the chaotic story of kidney injury into a clear map, allowing doctors to guide the right patients to the right care at the right time, potentially saving their kidneys from permanent damage.

1. Problem Statement

Acute Kidney Injury (AKI) is a severe condition affecting up to 50% of ICU patients, with survivors facing a high risk of progressing to Chronic Kidney Disease (CKD). While the link between AKI and CKD is established, identifying which specific AKI patients are at the highest risk remains challenging.

Limitations of Current Approaches: Existing studies largely rely on static, single-timepoint features (e.g., demographics, baseline comorbidities) or focus on specific AKI subtypes (e.g., sepsis-associated). These methods fail to capture the dynamic, time-varying nature of disease evolution and the complex interplay between clinical trajectories and biomarker changes.
Goal: To develop a scalable, data-driven framework that dynamically tracks patient clinical evolution from AKI to CKD using longitudinal Electronic Health Record (EHR) data, enabling the identification of high-risk subpopulations and distinct progression pathways.

2. Methodology

The study employs a multi-stage pipeline combining Natural Language Processing (NLP), deep learning, clustering, and survival analysis on a retrospective cohort from the NewYork-Presbyterian/Columbia University Irving Medical Center (NYP/CUIMC).

A. Data Source and Cohort Construction

Dataset: 7+ million patients (1985–2023).
Inclusion: Patients with a first AKI-associated hospitalization (within 48 hours of admission) defined by a ≥50% rise in creatinine from baseline.
Exclusion: Patients with pre-existing or concurrent CKD (based on codes or baseline creatinine ≥1.2 mg/dL), or those with insufficient creatinine measurements.
Final Cohort: 20,699 patients; 3,491 (17%) developed CKD.

B. Patient Representation (Vector Embeddings)

The core innovation is the construction of time-indexed patient vectors ( $\mathbf{x}_i^{(t)}$ ) that concatenate two distinct embeddings:

Medical Coding Sequence Embeddings: Generated using Paragraph Vector Models trained on EHRs. These capture cumulative exposure to conditions, medications, and procedures over time.
Temporal Creatinine Embeddings: Generated using a Transformer-based framework (adapted from Zerveas et al.) for multivariate time-series learning.
- Pre-training: Masked language modeling on a separate population to learn temporal dependencies in kidney function.
- Fine-tuning: Optimized for CKD diagnosis classification.
- Output: A fixed-length vector representing the kidney-specific physiological trajectory up to time $t$ .

Concatenation: $\mathbf{x}_i^{(t)} = \text{concat}(\mathbf{e}_{\text{codes}}^{(t)}, \mathbf{e}_{\text{Cr}}^{(t)})$ .

C. Clinical State Identification

Clustering: Applied K-Means++ clustering on the set of all patient vectors across the cohort.
Optimization: The optimal number of clusters ( $K$ ) was determined using the elbow method (Within-Cluster Sum of Squares) and validated by a board-certified nephrologist for clinical plausibility.
Result: 15 distinct post-AKI clinical states were identified. Each state is characterized by organ-system specific disease prevalences (e.g., respiratory failure, heart failure) and demographic profiles.

D. Multi-State Modeling and Risk Analysis

Transition Modeling: Used a time-inhomogeneous Markovian multi-state model to estimate transition probabilities between the 15 clinical states and to terminal states (CKD development or death). The Aalen-Johansen estimator was used to handle competing risks.
Survival Analysis: Identified risk factors for CKD development within specific subpopulations (defined by initial states or state transitions) using weighted Cox proportional hazards models (to address violations of the proportional hazards assumption).
Variables: Included demographics, baseline eGFR, comorbidities (DM, HTN, CHF, etc.), medication use, and acute physiological measurements (vitals, labs) from the first 24 hours.

3. Key Results

A. Clinical States and Trajectories

15 Distinct States: The model identified 15 unique clinical phenotypes. For example:
- S0: High prevalence of gastrointestinal and genitourinary diseases.
- S4: Associated with cardiovascular and respiratory issues.
- S6: Dominated by respiratory failure (92% prevalence).
Patient Flow: 75% of patients (15,607) remained in a single state or made only one transition during the study period.
CKD Risk Variation: Transition probabilities to CKD varied significantly by state. State 3 (S3) had the highest risk (42.3% at 10 years), while others were lower.

B. Risk Factor Identification

The study validated known risk factors and discovered novel, state-specific predictors:

Validated Factors: AKI severity, Diabetes (DM), Hypertension (HTN), Congestive Heart Failure (CHF), and Liver Disease were confirmed as risk factors in specific states (e.g., S4 and S5).
Novel Findings:
- State 0 (S0): Higher Day 1 Potassium increased risk; higher Sodium and Sepsis were protective.
- State 4 (S4): Higher Day 1 Temperature and Systolic Blood Pressure (SBP) increased risk; RAS inhibitors and higher Calcium/Magnesium were protective.
- State 5 (S5): Higher SBP and NSAID use increased risk; higher Diastolic BP and Sodium were protective.
- Transitions (S0 $\to$ S12): NSAID use and higher temperature were linked to increased risk.
Counter-Intuitive Finding: In S4, younger patients with higher hemoglobin and albumin levels had a higher risk of CKD, contradicting some prior literature.

4. Key Contributions

Dynamic Representation: Unlike static models, this approach integrates temporal medical codes with biomarker trajectories (creatinine) to create a comprehensive, time-evolving patient vector.
Granular Subtyping: The identification of 15 distinct clinical states allows for a more nuanced understanding of AKI recovery and progression than broad "AKI vs. Non-AKI" or single-comorbidity classifications.
Multi-State Framework: Successfully applied multi-state modeling to EHR data to quantify transition probabilities and map common progression pathways, explicitly accounting for competing risks (death).
Actionable Insights: The discovery of state-specific risk factors (e.g., the protective nature of sepsis in S0 vs. risk factors in S4) suggests that "one-size-fits-all" follow-up strategies are insufficient.

5. Significance and Future Directions

Clinical Impact: This framework supports the development of automated decision-support tools within EHRs. Clinicians can use these tools to identify high-risk AKI patients based on their specific clinical trajectory and assign them to targeted monitoring or early nephrology referrals.
Scalability: The method is cost-effective and scalable, leveraging existing structured EHR data without requiring manual chart review.
Generalizability: While focused on AKI-to-CKD, the framework is adaptable to other acute-to-chronic disease progressions (e.g., acute liver injury to cirrhosis) by swapping specific biomarkers and coding ontologies.
Limitations: The study relied on structured data (excluding unstructured notes), had limited urine data availability, and was conducted at a single tertiary center, necessitating future multi-institutional validation.

In conclusion, this paper presents a robust, data-driven methodology for characterizing the complex, non-linear progression from acute to chronic kidney disease, offering a pathway toward precision medicine in nephrology.