GLP-1RA Cardiovascular Benefits: AI Analysis of Patient Data
- this text details the statistical methods used to analyze data from the
- * Training Set: 70% of participants from both LEADER and SUSTAIN-6 were combined to train the models.
- * PRISM (Patient Response Identifiers for Stratified Medicine): A multi-step machine learning (ML) approach was used.
Summary of the Statistical Analysis Methods Used in the study
this text details the statistical methods used to analyze data from the LEADER and SUSTAIN-6 trials to understand treatment response to GLP-1 receptor agonists in type 2 diabetes. Here’s a breakdown:
1. Data Partitioning:
* Training Set: 70% of participants from both LEADER and SUSTAIN-6 were combined to train the models.
* Test Set: The remaining 30% from each trial were combined to test the models.
2. Identifying Treatment Response Subgroups (PRISM Framework):
* PRISM (Patient Response Identifiers for Stratified Medicine): A multi-step machine learning (ML) approach was used.
* Step 1: Developed a multivariable risk model for the outcome without treatment (control conditions).
* Step 2: Used interaction modeling with regularized regression (elastic net) to identify baseline factors that modify treatment benefit.
3. Overall Clinical Benefit Assessment:
* Cox Proportional Hazard Model: Used to test the overall clinical benefit of treatment across the combined LEADER + SUSTAIN-6 cohort. This model included a variable to account for which trial the participant was from.
* Sensitivity Analyses: Performed to address potential imbalances in randomization within identified subgroups (based on Standardized Mean Difference (SMD) > 0.1 and p < 0.05).
4. Estimating Treatment Affect:
* Predicted Survival Probability & NNT-ARR: Calculated at 3.6 years (median follow-up of the original trials).
* NNT-ARR (Number Needed to Treat based on Absolute Risk Reduction): Estimated using methods by Austin et al. (references 19 & 20).
* Bootstrapping: Used to calculate confidence intervals and standard errors for ARR (1000 samples with replacement).
* Quantitative Scale Interactions: Tested using Gail and Simon methods (reference 21).
5. Variables Considered for Transferability & Heterogeneity:
A wide range of baseline characteristics were examined to understand how treatment response might vary:
* Demographics: Age, sex, BMI
* Diabetes History: Duration of diabetes, baseline HbA1c
* Cardiovascular History: History of CVD, HF, MI, stroke, peripheral artery disease, hypertension
* Kidney Function: eGFR
* Medications: Use of metformin, sulphonylureas, thiazolidinediones, DPP4 inhibitors, RAS blockers, calcium channel blockers, beta blockers, diuretics, antiplatelet treatment, and statins.
* Urinary Albumin/creatinine Ratio (UACR): Used in LEADER analysis only (due to missing data in SUSTAIN-6).
6. External Validation:
* The same analytical approach was used for external validation, referencing original study populations (reference 10).
In essence, the study employed a elegant machine learning approach (PRISM) combined with customary survival analysis techniques to identify subgroups of patients who might benefit differently from GLP-1 receptor agonist treatment. They also focused on quantifying the magnitude of these differences and assessing the robustness of thier findings.