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

Figures

• 

  Figure 1.. Representation of the proposed model and validation pipeline. Stage 1 uses Lasso regression (a method that selects the most relevant predictors by shrinking less important ones to zero) to remove the effect of baseline disease severity on outcomes, so that Stage 2 can use Ridge regression (a method that retains all candidate predictors while preventing any single variable from dominating) to isolate which patient variables specifically modify treatment benefit. Each patient receives a predicted treatment effect score (ITE) and is classified as higher or lower predicted benefit. The model is validated through permutation testing to confirm that the subgroup separation is statistically significant, cross-trial analysis to confirm that the identified variables are specific to each drug's mechanism of action, and surrogate decision trees to translate the model into clinically interpretable rules using routinely available variables.

• 

  Figure 2.. Prognostic model discrimination and variable importance for predicting clinical deterioration in HFpEF. (A) Receiver operating characteristic (ROC) curves with 95% confidence intervals for the TOPCAT-trained prognostic model evaluated across four HFpEF clinical trials and a pooled external validation cohort. The model demonstrated satisfactory internal discrimination in TOPCAT (AUC = 0.76) and external generalizability to NEAT-HFpEF (AUC = 0.74), with more modest/poor performance in INDIE-HFpEF (AUC = 0.60) and RELAX (AUC = 0.53). (B) Mean absolute SHAP values indicating the relative contribution of each baseline covariate to model predictions. NYHA functional class and KCCQ Overall Summary Score were the dominant predictors.

• 

  Figure 3.. Kaplan-Meier event-free survival curves stratified by predicted treatment response in TOPCAT. Patients classified as predicted responders (blue) versus non-responders (purple) to spironolactone were compared across five clinical endpoints: (A) the primary composite of cardiovascular death or heart failure hospitalization, (B) all-cause death, (C) cardiovascular death, (D) all-cause hospitalization, and (E) heart failure hospitalization. Log-rank p-values are shown for each comparison. Predicted responders demonstrated significantly better event-free survival for the primary composite endpoint (P < 0.001) and heart failure hospitalization (P < 0.001).

• 

  Figure 4.. Prognostic model discrimination by treatment arm across HFpEF clinical trials. Area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals for the prognostic model applied separately to treatment and placebo arms in TOPCAT, RELAX, and INDIE-HFpEF. Comparable AUROC values between treatment and placebo arms within each trial indicate that model predictions reflect baseline prognostic risk rather than treatment-specific effects. The dashed line indicates chance-level discrimination (AUROC = 0.5).

• 

  Figure 5.. Interaction-based treatment effect model trained on TOPCAT (Spironolactone), comparing the full cohort with the Americas-only cohort. (A, C) Non-zero treatment–covariate interaction coefficients (β_T×X) from the penalized Ridge regression model, ranked by absolute magnitude. Positive coefficients (green) indicate variables associated with greater predicted treatment benefit, while negative coefficients (red) indicate variables associated with reduced benefit or predicted harm. Panel A shows the full TOPCAT cohort (N = 1,536; CV R2 = 0.266) and Panel C shows the Americas-only cohort after exclusion of Russia/Georgia sites (N = 719; CV R2 = 0.189). Core treatment-modifying variables (i.e., diabetes mellitus, creatinine clearance < 30, BUN value, and bundle branch block) are identified in both cohorts, though their relative ranking shifts with the restricted sample. (B, D) Differential treatment effect stratified by predicted individualized treatment effect (ITE). Patients predicted to have higher benefit (blue) versus lower benefit (orange) were compared based on the difference in outcome change between treatment and placebo arms (TE = Δ_Tx − Δ_Pb). Restricting to the Americas…

• 

  Figure 6.. Interaction-based treatment effect model trained on NEAT-HFpEF (Isosorbide Mononitrate). (A) Non-zero treatment–covariate interaction coefficients (β_T×X) from the penalized Ridge regression model. Pulse pressure, rales, and systolic blood pressure emerged as the strongest interaction variables, all with negative coefficients suggesting that higher values are associated with reduced treatment benefit. An additional 6 variables are not shown. (B) Differential treatment effect across validation studies by predicted ITE subgroup. The NEAT-trained model identified a significant differential treatment effect when validated on its own cohort (NEAT: high ITE +1.3 vs. low ITE −8.3), with consistent directionality across external studies.

• 

  Figure 7.. Interaction-based treatment effect model trained on INDIE-HFpEF (Inorganic Nitrite). (A) Non-zero treatment–covariate interaction coefficients (β_T×X). BUN value was the dominant interaction variable, with elevated BUN associated with reduced predicted treatment benefit. Pulse pressure value was the only variable with a positive interaction coefficient of notable magnitude. One additional variable is not shown. (B) Differential treatment effect across validation studies by predicted ITE subgroup. The INDIE-trained model identified patients with a predicted treatment benefit of +4.9 in its own cohort, compared to −6.3 in the predicted low-ITE subgroup.

• 

  Figure 8.. Interaction-based treatment effect model trained on RELAX (Sildenafil). (A) Non-zero treatment–covariate interaction coefficients (β_T×X). Chronic kidney disease stage was the strongest negative interaction variable, while aortic regurgitation showed the largest positive interaction, suggesting that patients with aortic regurgitation may derive relatively greater benefit from sildenafil. An additional variables not shown. (B) Differential treatment effect by predicted ITE subgroup. The RELAX model showed a +0.8 VO2-point differential between high-ITE (+0.3) and low-ITE (−0.5) subgroups, but this did not reach statistical significance (p=0.10), likely due to the small sample size (N=189) and the narrow dynamic range of peak VO2.

• 

  Figure 9.. Cross-validation matrix of interaction-based treatment effect models across HFpEF clinical trials. (A) Interaction statistic (TE_high − TE_low) for each combination of training study (rows) and validation study (columns). Positive values indicate that the predicted high-ITE subgroup experienced a greater treatment effect than the predicted low-ITE subgroup. Diagonal cells with bold borders represent internal validation. The largest interaction statistics consistently appear along the diagonal, indicating that each model's treatment-modifying signal is strongest for its own drug. (B) Permutation test p-values displayed as −log10(P) for each training–validation pair. Statistically significant interactions (P < 0.05, shown in bold red) were identified for TOPCAT validated on itself (P = 0.001), NEAT-HFpEF validated on itself (P = 0.003), INDIE-HFpEF validated on itself (P = 0.004), and the INDIE-trained model validated on RELAX (P = 0.036). This diagonal-dominant pattern supports the interpretation that the ITE models capture drug-specific rather than generic prognostic signals.

• 

  Figure 10.. Surrogate decision tree approximating the TOPCAT-trained interaction model for spironolactone treatment effect prediction. A depth-3 decision tree was fit to approximate the complex Ridge interaction model's individualized treatment effect (ITE) predictions using un-standardized baseline variables. Diabetes mellitus was the primary split: patients without diabetes were predicted to benefit from spironolactone (ITE range: +0.06 to +0.50), with the greatest benefit among those with bundle branch block (Any BBB) who were not on statins (+0.50). Patients with diabetes were predicted to experience harm, particularly those on oral hypoglycemics with low creatinine clearance (≤71.9 mL/min; ITE = −0.57). Leaf node colors indicate predicted treatment effect direction and magnitude (blue = benefit, red = harm). Surrogate fidelity: R2 = 0.665, Spearman ρ = 0.771, high/low agreement = 78.8%.

• 

  Figure 11.. Surrogate decision tree approximating the NEAT-HFpEF–trained interaction model for isosorbide mononitrate treatment effect prediction. Insulin use was the primary split, with insulin-dependent patients uniformly predicted to experience harm (ITE range: −0.06 to −0.11). Among non-insulin patients, pulse pressure further stratified outcomes: those with lower pulse pressure (≤63.5 mmHg) who were not on statins showed the only predicted benefit (+0.04). The narrow ITE range (−0.11 to +0.04) reflects the overall modest heterogeneity of treatment effect in this trial. Surrogate fidelity: R2 = 0.642, Spearman ρ = 0.815, high/low agreement = 82.3%.

• 

  Figure 12.. Surrogate decision tree approximating the INDIE-HFpEF–trained interaction model for inorganic nitrite treatment effect prediction. BUN value was the primary split: patients with elevated BUN (>37.5 mg/dL) were uniformly predicted to experience harm, with higher systolic blood pressure (>132 mmHg) further worsening the predicted effect (ITE = −0.21). Among patients with lower BUN, smaller body surface area (≤1.90 m2) and lower BMI (≤29.24 kg/m2) were associated with the greatest predicted benefit (ITE = +0.13). Surrogate fidelity: R2 = 0.587, Spearman ρ = 0.724, high/low agreement = 77.6%.