Causal Machine Learning for Comparative Effectiveness of GLP-1 RA versus SGLT2i in Heart Failure Using Real-World EHR Data

Using causal machine learning on real-world electronic health records, this study found that GLP-1 receptor agonists are associated with a lower risk of all-cause mortality or heart failure hospitalization compared to SGLT2 inhibitors in heart failure patients, though evidence for individualized treatment selection remains limited.

The Gist

Imagine you are a doctor standing in a pharmacy aisle, but instead of choosing between two brands of cereal, you have to choose between two powerful heart medications for a patient: GLP-1 RA and SGLT2i. Both are like high-performance engines designed to keep a failing heart (Heart Failure) running smoothly, but doctors have struggled to know which one is the "perfect fit" for each specific person.

This paper is like a team of data detectives using a super-smart computer brain (called Causal Machine Learning) to solve that mystery using real-life patient records from Stony Brook University Hospital.

Here is the breakdown of what they found, using some everyday analogies:

1. The Problem: The "One-Size-Fits-All" Guess

Traditionally, doctors have had to guess which drug works best based on broad averages. It's like saying, "This umbrella is great for everyone," without checking if it's actually raining on your specific street. The goal of this study was to move from guessing to precision medicine—finding the right tool for the right person.

2. The Tool: The "Time-Traveling" Computer

The researchers used a special kind of AI that acts like a simulated time machine.

• Normally, you can't go back in time to see what would have happened to a patient if they had taken the other drug.
• This "Causal Machine Learning" looks at thousands of real patient records and uses complex math to simulate two parallel universes for every patient: one where they took Drug A, and one where they took Drug B.
• By comparing these two imaginary timelines, the computer can estimate the true cause-and-effect of the drugs, filtering out other factors that might confuse the results (like age or other illnesses).

3. The Big Discovery: The "Average Winner"

When they looked at the whole group of patients as a single crowd, the computer found a clear winner.

• The Result: Patients taking GLP-1 RA were less likely to die or be hospitalized for heart failure within a year compared to those taking SGLT2i.
• The Analogy: Imagine a race where, on average, the runners wearing the Red Shoes (GLP-1 RA) finished faster and got fewer injuries than those wearing the Blue Shoes (SGLT2i). For the general population, the Red Shoes seem like the safer bet.

4. The Twist: The "Personalized Fit" is Tricky

The researchers then asked a harder question: "Is there a specific type of runner who should wear the Blue Shoes instead?"

• They looked for "effect modifiers"—specific traits that might make one drug work better for a specific person.
• The Findings: They found a few clues (like how much water pills a patient takes, their body weight, or how well their kidneys work), but the evidence wasn't strong enough to say, "If you have trait X, you must take Drug Y."
• The Analogy: It's like trying to find a specific key that opens a special lock. They found a few keys that might fit, but they aren't 100% sure yet. The "Red Shoes" still seem like the best choice for almost everyone, even if a few people might have a slightly different need.

5. The Bottom Line: Promise with Caution

The study shows that using real-world data with smart AI is a powerful way to help doctors make better choices. It's like upgrading from a paper map to a live GPS system.

However, the authors warn that just because the GPS says "turn left," you still need to check the road conditions before you drive. These models are promising, but they need more testing and careful checking of their assumptions before doctors can start using them to make life-or-death decisions for every single patient.

In short: The computer says GLP-1 RA is generally the stronger shield against heart failure problems right now, but we are still figuring out exactly which specific patients might benefit more from the other option.

Technical Summary

Technical Summary: Causal Machine Learning for Comparative Effectiveness of GLP-1 RA versus SGLT2i in Heart Failure

1. Problem Statement

Clinicians currently lack robust precision medicine tools capable of estimating individualized treatment effects (ITE) for patients with heart failure (HF). While observational data from electronic health records (EHR) is abundant, traditional statistical methods often fail to adequately capture complex treatment heterogeneity or adjust for high-dimensional confounding factors. There is a critical need to move beyond population-average estimates to determine if specific subgroups of HF patients derive significantly different benefits from Glucagon-like peptide-1 receptor agonists (GLP-1 RA) versus Sodium-glucose cotransporter 2 inhibitors (SGLT2i).

2. Methodology

The study employed a Causal Machine Learning framework applied to real-world EHR data from Stony Brook University Hospital. Key methodological components include:

• Data Source: Retrospective analysis of patient records containing clinical variables, medication history, and outcomes.
• Causal Framework: The authors utilized a Doubly Robust (DR) framework. This approach combines outcome modeling (predicting the result given covariates) and propensity score modeling (predicting treatment assignment given covariates). The DR estimator ensures consistency if either the outcome model or the propensity model is correctly specified, thereby reducing bias from model misspecification.
• Comparative Analysis: The study compared GLP-1 RA against SGLT2i.
• Outcome Definition: The primary endpoint was a 1-year composite outcome consisting of all-cause mortality or HF-related hospitalization.
• Heterogeneity Analysis: Advanced machine learning algorithms were used to estimate individualized treatment effects and identify potential effect modifiers (variables that alter the treatment effect).

3. Key Contributions

• Application of Causal ML in HF: The paper demonstrates the feasibility of using causal machine learning to extract actionable insights from noisy, real-world EHR data for complex cardiovascular conditions.
• Dual Estimation Capability: The methodology successfully estimates both Average Treatment Effects (ATE) for the general population and Individualized Treatment Effects (ITE) to explore precision medicine potential.
• Identification of Effect Modifiers: The study systematically evaluated clinical variables to determine if they influence the comparative effectiveness of the two drug classes, moving beyond simple association to causal inference.

4. Key Results

• Population-Average Effect: Under the doubly robust framework, the analysis revealed a stable population-average effect. Specifically, GLP-1 RA was associated with a lower risk of the 1-year composite outcome (all-cause mortality or HF hospitalization) compared to SGLT2i.
• Individualized Treatment Selection: The heterogeneity analysis provided limited evidence to support distinct individualized treatment selection strategies based on the current data. In other words, the models did not strongly identify specific patient profiles where one drug was definitively superior to the other for all individuals in a way that would drastically change clinical decision-making for every patient.
• Potential Effect Modifiers: Despite the limited evidence for broad individualized selection, subgroup tests identified three specific variables as potential effect modifiers:
  1. Loop diuretic use
  2. Body Mass Index (BMI)
  3. Estimated Glomerular Filtration Rate (eGFR)

5. Significance and Future Directions

This research highlights the transformative potential of causal machine learning in translating observational data into actionable precision care. By demonstrating that GLP-1 RA may offer superior outcomes on average compared to SGLT2i in this specific HF cohort, the study provides a data-driven basis for clinical consideration.

However, the authors emphasize a crucial caveat: while the models show promise, rigorous validation and a careful assessment of causal assumptions (such as unmeasured confounding and positivity) are prerequisites before these tools can be safely implemented in clinical practice. The identification of specific effect modifiers (like eGFR and BMI) suggests that future studies with larger datasets may be able to refine these models to achieve true precision medicine goals.