Post-AKI Clinical States Analysis
- Post-AKI clinical states are distinct physiological and biochemical phenotypes emerging after acute kidney injury, enabling precise risk stratification.
- Data-driven clustering and temporal embedding techniques categorize patients into groups with varying risks for CKD progression, death, and readmission.
- Predictive models using multi-state Markov frameworks and GAM analyses offer actionable insights for targeted follow-up and optimal resource allocation.
Acute kidney injury (AKI) is a frequent and serious complication among hospitalized patients, characterized by abrupt and dynamic impairment of renal function. The concept of post-AKI clinical states encompasses the diverse physiological, biochemical, and risk-based phenotypes that emerge following an AKI episode. These states provide critical granularity for prognostication, management, and prevention of adverse outcomes such as chronic kidney disease (CKD), renal non-recovery, and mortality.
1. Stratification of Post-AKI Clinical States
Post-AKI clinical states are defined along several axes, reflecting heterogeneity in recovery, progression, and systemic involvement. Data-driven approaches have identified distinct clusters based on patient attributes, AKI trajectories, and organ-system codes.
Cluster-Based Phenotypes
In a cohort of approximately 20,000 AKI survivors, unsupervised clustering of nine features (demographics, renal indices, inflammatory markers, comorbidities) yielded three major strata (Silva et al., 2024):
| Group | Age/Comorbidity | Renal Function | Inflammation/Other |
|---|---|---|---|
| Group 1 | >75 years, high comorbidity (CHF, DM, cancer) | Worst (high discharge and peak creatinine, low eGFR) | High CRP, male predominance |
| Group 2 | ~70 years, moderate comorbidity | Intermediate | Mid-range CRP |
| Group 3 | ~65 years, lowest comorbidity | Best preserved eGFR, low creatinine | Lowest CRP, female predominance |
Assignment of new patients to a group is based on Euclidean distance to cluster centroids in the standardized feature space.
Longitudinal State Embeddings
A separate framework used temporal embeddings (from longitudinal EHR codes and creatinine) to define fifteen post-AKI states, each labeled by dominant medical codes and associated with distinct CKD progression risks (Fang et al., 18 Nov 2025). These states ranged from those dominated by respiratory, cardiovascular, or neurological failure to those linked primarily to metabolic derangements or hypertension/diabetes codes.
2. Dynamic Trajectories and Transition Models
The post-AKI period is marked by transitions between states, either towards renal recovery or adverse outcomes. Multi-state Markov models and trajectory-based classifications offer detailed quantitative frameworks (Adiyeke et al., 2023, Fang et al., 18 Nov 2025).
AKI Recovery Trajectories
Trajectories are typically classified as (Adiyeke et al., 2024):
- Rapidly Reversed AKI: Return of creatinine to within 25% of baseline within 48 hours.
- Persistent AKI with Renal Recovery: Prolonged elevation but recovery to baseline by discharge.
- Persistent AKI without Renal Recovery: Continued elevation at discharge or death.
Persistent AKI—especially without recovery—carries markedly higher risks for death, CKD, prolonged resource utilization, and post-discharge complications.
Markov State Models
Multistate models formalize the stochastic evolution between AKI stages, resolution, death, or discharge. For example, after Stage 1 AKI, the transition probabilities at 7 days are: resolution to No AKI 25%, persistent Stage 1 68%, progression to Stage 2 less than 2%, in-hospital death <0.2% (Adiyeke et al., 2023).
In embedding-based models, patients traverse a state-space {S₀ … S₁₄, CKD, Death}, with transition intensities estimated using cumulative hazards and Aalen–Johansen estimators. The majority of patients (75%) occupy a single state or transition only once post-AKI before CKD or censoring (Fang et al., 18 Nov 2025).
3. Predictive Modeling of Post-AKI Outcomes
Machine learning frameworks tailored to phenotype-specific risk yield improved prognostic performance. Generalized additive models (GAMs) trained within each cluster enable interpretation and calibration of mortality risk (Silva et al., 2024).
- For Group 1 (oldest, most comorbid), 90-day post-discharge mortality AUROC was 0.77.
- Global models (single GAM or logistic regression without stratification) performed modestly worse in discrimination (AUROC ≈ 0.72–0.73).
Decision-curve analyses confirm that group-specific GAMs dominate global models across clinically relevant thresholds. Notably, the effect of individual variables (such as C-reactive protein or discharge creatinine) on predicted risk is contingent upon cluster membership.
In CKD progression models using longitudinal state embeddings, the highest 5-year CKD risk was observed in the S₃ (“respiratory failure” cluster): 29% at 5 years, 42% at 10 years (Fang et al., 18 Nov 2025).
4. Risk Factors and Heterogeneity in Progression
Risk factors for adverse post-AKI outcomes display state dependency. Classical predictors (age, baseline eGFR, diabetes, hypertension, heart failure) remain robust but may interact with state membership. For example:
- In S₀ (gastrointestinal/genitourinary failure predominant), risk of CKD progression is increased by lower eGFR, higher potassium, DM, HTN, CHF, liver disease; sepsis at admission is paradoxically protective (HR 0.49, p=0.004).
- In S₅ (hypertension/diabetes codes), outpatient NSAID use is a strong risk factor for CKD progression (HR 1.47, p=0.004) (Fang et al., 18 Nov 2025).
Frailty (Charlson index ≥3) and prolonged ICU stay (≥48 h) independently reduce the likelihood of AKI resolution and increase death risk; for example, frailty reduces the hazard of resolution of Stage 2 AKI by ~40% and increases the hazard of death by 45–65% across AKI stages (Adiyeke et al., 2023).
5. Clinical and Operational Implications
Stratified post-AKI phenotypes and state trajectories allow precision triage and allocation of follow-up intensity.
- High-risk strata (e.g., Group 1 by clustering, S₃ or S₀→S₁₂ transitions): Recommendations include intensive nephrology follow-up, advanced monitoring, and early intervention.
- Intermediate-risk groups: Medium-intensity follow-up (primary care plus nurse specialist evaluations).
- Low-risk groups (youngest, low comorbidity, preserved renal markers): Safe for less frequent follow-up, primary care management (Silva et al., 2024).
Resource utilization, length of stay, and likelihood of readmission/discharge destination are all predicted by trajectory class and cluster/state assignment (Adiyeke et al., 2024). Post-AKI monitoring and interventions can be prioritized based on the distinct risk profiles identified by these frameworks.
Ethical considerations arise, as stratified models help avoid misclassification by “average” thresholds, thus enhancing fairness by aligning care with individual risk and phenotype.
6. Methodological Advances and Limitations
The application of unsupervised learning (k-means with clustering-aware refinement, Doc2Vec and transformer embedding), together with multi-state modeling, enables dynamic and high-dimensional patient characterization (Silva et al., 2024, Fang et al., 18 Nov 2025). The use of explainable models (GAMs with smooth cubic splines) facilitates interpretability and clinical acceptance.
A subset of studies use periodic re-clustering (“concept-drift adaptation”) to ensure updated relevance as patient populations and treatments evolve. However, the absence of reported calibration metrics (e.g., calibration slope, Brier score) in some models leaves room for further validation.
7. Synthesis and Outlook
Post-AKI clinical states represent a multiaxial stratification of the heterogeneous recovery landscape following acute kidney injury. Data-driven cluster phenotypes, state-transition models, and trajectory-informed risk assessment improve stratification granularity, allowing for targeted surveillance, resource allocation, and more nuanced clinical decision-making. Established and novel risk factors demonstrate state-dependent penetrance and inform further mechanistic investigations. Progress in this field will likely include integration of additional biological omics data, calibration of models for diverse populations, and development of adaptive follow-up protocols underpinned by these multistate frameworks (Silva et al., 2024, Fang et al., 18 Nov 2025, Adiyeke et al., 2024, Adiyeke et al., 2023).