Heterogeneous Treatment Effects in HFpEF: Distinguishing Drug-Specific Response from Prognostic Phenotypes Across Randomized Trials

Technical Summary: Heterogeneous Treatment Effects in HFpEF

Problem Statement
Heart failure with preserved ejection fraction (HFpEF) is a heterogeneous syndrome characterized by diverse pathophysiological phenotypes and significant comorbidity burdens. Despite this heterogeneity, randomized controlled trials (RCTs) in HFpEF have largely reported neutral average treatment effects. The authors argue that these null results may obscure drug-specific benefits within distinct subgroups. A critical methodological limitation in current research is the definition of "treatment response." Conventional approaches often dichotomize patients into "responders" and "non-responders" based on arbitrary thresholds (e.g., a specific improvement in the Kansas City Cardiomyopathy Questionnaire score). The paper posits that such machine learning models trained on these thresholds often capture prognostic signals (baseline disease severity and natural clinical trajectories common to both treatment and placebo arms) rather than true predictive signals (variables that specifically modify the treatment effect). Consequently, these models fail to distinguish which patients benefit from a specific drug versus those who simply have a favorable natural history.

Methodology
To disentangle prognostic signals from drug-specific treatment effects, the authors developed a framework utilizing data from four HFpEF clinical trials: TOPCAT (spironolactone), RELAX (sildenafil), NEAT-HFpEF (isosorbide mononitrate), and INDIE-HFpEF (inorganic nitrite). The study employed two complementary analytic approaches:

1. Prognostic Responder Model (Approach 1):

   • A regularized logistic regression (LASSO) model was trained on the treatment arm of the TOPCAT trial to predict a composite "responder" status (defined by improvements in NYHA class, KCCQ, 6-minute walk test, or peak VO₂).
   • This model was evaluated for its ability to predict outcomes in both treatment and placebo arms.
   • Validation: The authors tested whether the identified "responders" showed a differential treatment effect (interaction) compared to "non-responders." If the model captures only prognosis, the treatment effect should be similar across both groups.

2. Interaction-Based Individual Treatment Effect (ITE) Model (Approach 2):

   • This two-stage framework uses continuous outcomes (change in KCCQ, 6MWT, or VO₂) to directly estimate treatment-by-variable interactions.
   • Stage 1 (Prognostic Adjustment): A Lasso regression model predicts the outcome using baseline variables and a treatment indicator, capturing the average treatment effect and prognostic factors. The residuals ($\hat{\epsilon}_1$) represent the variation unexplained by baseline prognosis or the average treatment effect.
   • Stage 2 (Interaction Modeling): A Ridge regression (or Elastic Net) models the residuals from Stage 1 as a function of treatment-by-variable interaction terms ($T \times X$). This isolates baseline characteristics that specifically modify the treatment effect.
   • Validation: The authors performed permutation tests to assess the statistical significance of the interaction. Crucially, they conducted cross-trial validation: an ITE model trained on one trial was applied to the others. They hypothesized that true drug-specific modifiers would only show significant interactions within their own trial (diagonal significance) and not generalize to trials with different mechanisms of action.
   • Interpretability: Surrogate decision trees (depth=3) were trained on the continuous ITE predictions to translate complex interaction coefficients into clinically actionable rules using unstandardized variables.

Key Results

• Prognostic vs. Predictive Distinction: The Approach 1 prognostic model achieved good discrimination (AUROC 0.76 in TOPCAT) and predicted long-term survival outcomes. However, when applied to both treatment and placebo arms, it showed no significant treatment-by-subgroup interaction ($p=0.17$). This confirms that conventional "responder" models primarily identify patients with favorable natural trajectories (prognosis) rather than those who derive specific benefit from the drug.
• Drug-Specific Heterogeneity (Approach 2): The ITE models successfully identified distinct, mechanism-specific effect modifiers for each trial:
  • TOPCAT (Spironolactone): Identified a cardiorenal-inflammation phenotype. Variables such as creatinine clearance <30 mL/min and bundle branch block were associated with increased benefit, while diabetes and elevated BUN were associated with reduced benefit or harm. In the Americas-only cohort, the treatment effect separation between high- and low-ITE groups was substantial (+12.1 KCCQ points).
  • NEAT-HFpEF (Isosorbide Mononitrate): Identified a hemodynamic phenotype. Patients with high pulse pressure, rales, and elevated systolic BP (signs of congestion/stiffness) experienced reduced benefit or harm, whereas less hemodynamically compromised patients showed modest benefit.
  • INDIE-HFpEF (Inorganic Nitrite): Identified a volume-overload and metabolic phenotype. Patients with elevated BUN, jugular venous pressure, and metabolic derangement derived less benefit, likely due to impaired NO-cGMP-PKG signaling.
  • RELAX (Sildenafil): Identified inflammation and volume overload markers, though the interaction did not reach statistical significance, likely due to smaller sample size and endpoint limitations.
• Cross-Trial Specificity: The cross-validation matrix demonstrated a "diagonal" pattern of significance. Each ITE model was statistically significant only when validated on its own trial's drug. For instance, the TOPCAT model did not predict differential response to nitrates, and the NEAT model did not predict response to spironolactone. This confirms the models capture drug-specific mechanisms rather than a generic "treatment responsiveness."
• Clinical Interpretability: Surrogate decision trees successfully approximated the complex ITE models (e.g., TOPCAT tree $R^2=0.665$), providing clear rules such as "patients without diabetes and with bundle branch block benefit most from spironolactone."

Significance and Claims
The paper claims that its proposed framework successfully addresses the methodological challenge of distinguishing prognostic phenotypes from predictive treatment modifiers in HFpEF. By moving away from dichotomized "responder" definitions toward continuous interaction modeling, the study demonstrates that:

1. Neutral trials may mask heterogeneity: The overall null results of trials like TOPCAT and NEAT-HFpEF likely conceal significant, drug-specific benefits in identifiable subgroups and potential harm in others.
2. Mechanism-specificity is detectable: Treatment effect modifiers are aligned with the biological mechanisms of the drugs (e.g., cardiorenal for aldosterone antagonists, hemodynamic for nitrates).
3. Clinical Utility: The surrogate decision trees offer a pathway for phenotype-guided therapy using routinely collected clinical variables, potentially enabling more precise patient selection in future trials and clinical practice.
4. Trial Design Implications: The authors argue for "phenotype-informed trial design," suggesting that future HFpEF trials should consider pre-specified enrichment strategies based on predicted ITE to enhance the detection of treatment effects.

The authors maintain a modest tone, acknowledging that these findings are exploratory and require prospective validation in independent cohorts or enrichment-design trials. They also note limitations regarding the statistical power of smaller trials and the fact that baseline clinical variables explain only a portion of the treatment effect heterogeneity, suggesting a need for more granular data modalities (genomics, proteomics) in the future.