Dialysis Risk Prediction and Treatment Effect Estimation for AKI patients using Longitudinal Electronic Health Records

Technical Summary: Dialysis Risk Prediction and Treatment Effect Estimation for AKI Patients

1. Problem Statement

The progression from Acute Kidney Injury (AKI) to dialysis or End-Stage Renal Disease (ESRD) is a critical clinical outcome with high morbidity and mortality. Clinicians face two primary challenges in managing these patients:

• Risk Prediction: Identifying which AKI patients are at high risk of rapid progression to dialysis to allow for timely intervention.
• Causal Inference (Treatment Effects): Determining how specific medications influence kidney outcomes. This is complicated by polypharmacy (patients taking multiple drugs simultaneously) and confounding by indication (where drugs are prescribed because of the severity of the disease, making it hard to distinguish if the drug caused the outcome or the disease did).

2. Methodology

The researchers developed a unified framework that treats both prediction and treatment effect estimation as a causal inference problem using longitudinal Electronic Health Record (EHR) data.

• Cohort Construction: A fixed-window design was used with a cohort of 81,401 patients. The study utilized a 90-day observation window (to capture clinical history) followed by a 730-day prediction window.
• Data Representation: The model processes sequences of diagnoses, procedures, and medications. To capture physiological changes, they incorporated laboratory trend features (eGFR, creatinine, and BUN), calculating the mean, last observed value, and slope for each.
• Model Architecture (Transformer-Based Causal Model):
  • The core is a Transformer encoder that uses self-attention to learn long-range dependencies in clinical sequences.
  • Multi-head Design: The architecture features four specialized output heads:
    1. Outcome Head: Predicts the probability of dialysis/ESRD.
    2. Propensity Head: Models the likelihood of a patient receiving a specific treatment (to account for selection bias).
    3. Two Potential Outcome Heads: Estimates what the outcome would have been under treated vs. untreated conditions.
• Counterfactual Inference: To estimate the Average Treatment Effect (ATE), the model performs "intervention" during inference. It modifies the medication tokens in a patient's history (removing them to simulate "untreated" or adding them to simulate "treated") and measures the resulting shift in predicted dialysis risk.
• Clinical Validation Layer: To ensure the model's "causal" signals weren't just noise, they performed post-hoc statistical analyses (using IPTW, AIPW, and OLS) on actual kidney lab markers to see if the medication effects aligned with physiological reality.

3. Key Contributions

1. Unified Framework: A single transformer-based architecture that performs both risk prediction and counterfactual treatment estimation.
2. Ingredient-Level Granularity: Instead of looking at specific brand names, the model estimates effects at the active-ingredient level, reducing the noise caused by fragmented medication coding.
3. Imbalance-Aware Training: Developed strategies (weighted loss, focal loss) to handle the extreme rarity of the outcome (only 1.1% prevalence).
4. Multi-Modal Integration: Successfully combined discrete event sequences (codes) with continuous physiological trajectories (lab trends).

4. Results

• Predictive Performance: The Transformer model outperformed baseline Logistic Regression and Linear SVM models.
  • AUC: 0.694 (90-day window).
  • F1 Score: 0.201 at a high decision threshold (0.883).
  • Observation Window: Interestingly, a 90-day window performed better than a 365-day window, suggesting that longer windows may introduce irrelevant "noise" into the model.
• Treatment Effect Estimation (ATE):
  • ACE/ARB Inhibitors (e.g., Lisinopril): Showed a protective effect (lower predicted risk of dialysis), which aligns with clinical knowledge of their renoprotective properties.
  • Loop Diuretics (e.g., Furosemide): Showed a risk-increasing signal, though the authors note this likely reflects the severity of the patient's condition (confounding by indication) rather than the drug being inherently toxic.
• Validation: The model's signals for ACE/ARBs were partially supported by observed stability in eGFR and creatinine levels.

5. Significance

This work demonstrates that deep learning, specifically Transformer architectures, can move beyond simple correlation to provide hypothesis-generating causal insights from messy, real-world EHR data. While the authors caution that these are not definitive clinical proofs, the framework provides a scalable way for researchers to identify potential medication risks and benefits in large populations, potentially guiding more personalized and safer nephrology care.