← Clinical Trials
Acute Myocardial Infarction Phase 3 Superiority · Time-to-Event NCT02924727

PARADISE-MI: Complete Statistical Analysis of Sacubitril/Valsartan in Acute Myocardial Infarction

An independent statistical review of the randomized, quadruple-masked phase 3 PARADISE-MI trial comparing LCZ696 (sacubitril/valsartan) with ramipril after acute myocardial infarction, focusing on the Cox-model hazard ratios for the adjudicated primary composite of cardiovascular death, heart failure hospitalization or outpatient heart failure, its components, the secondary endpoints and the negative binomial analysis of total events.

ClinicalTrials.gov NCT02924727  ·  Study start: 2016-12-09  ·  Primary completion: 2021-02-26  ·  Sponsor: Novartis Pharmaceuticals
About this page

This page provides an independent statistical analysis and educational interpretation of publicly reported results. ClinicalTrials.gov provides the official trial registry record.

1. Trial at a Glance

PARADISE-MI (Prospective ARNI vs ACE Inhibitor Trial to DetermIne Superiority in Reducing Heart Failure Events After MI) was a randomized, parallel-group, quadruple-masked phase 3 trial testing whether the angiotensin receptor–neprilysin inhibitor LCZ696 (sacubitril/valsartan) was superior to the ACE inhibitor ramipril in reducing a composite of cardiovascular death and heart failure events in patients with acute myocardial infarction.

5669
Enrolled
2 randomized arms
0.9012
Primary HR
95% CI 0.7778–1.0441
0.1659
Primary P-value
Superiority, Cox model
~43 mo
Max. follow-up
Up to approximately 43 months
FeaturePARADISE-MI
PhasePhase 3
ConditionAcute myocardial infarction
DesignRandomized, parallel-group, quadruple-masked, active-controlled
Primary purposeTreatment
Enrollment5669 participants
ComparisonLCZ696 (sacubitril/valsartan) vs ramipril
Primary endpointTime to first CEC-confirmed composite of CV death, HF hospitalization or outpatient HF
Primary analysisCox proportional hazards model, hazard ratio, superiority hypothesis, Full Analysis Set
StatusCompleted; results posted on ClinicalTrials.gov
Key datesStart 2016-12-09; primary completion 2021-02-26
ClinicalTrials.govNCT02924727
SponsorNovartis Pharmaceuticals (industry)

2. Clinical Question

The trial asked whether replacing ACE inhibition with combined angiotensin receptor blockade and neprilysin inhibition would lower the rate of cardiovascular death or heart failure events after an acute myocardial infarction. Because the comparator was an active drug rather than placebo, the question is one of relative superiority between two active strategies, not whether renin–angiotensin system blockade is beneficial at all.

Population

Patients with acute myocardial infarction. The ClinicalTrials.gov record lists acute myocardial infarction as the condition studied; the detailed eligibility criteria are described in the registry entry.

Intervention

LCZ696 (sacubitril/valsartan), given with placebo of ramipril to maintain blinding.

Comparator

Ramipril, given with placebo of LCZ696 to maintain blinding.

Primary question

Does LCZ696 reduce the hazard of the first CEC-confirmed CV death, HF hospitalization or outpatient HF event compared with ramipril?

3. Trial Design

01
Acute MIEligible patients enrolled
02
Randomize5669 enrolled, 2 arms
03
Masked treatmentLCZ696 or ramipril with matching placebos
04
AdjudicationClinical Endpoint Committee confirms events
05
AnalysisCox model, up to ~43 months
ARM 1 · EXPERIMENTAL

LCZ696 (sacubitril/valsartan)

  • Active LCZ696
  • Placebo of ramipril
  • Safety set at risk for serious adverse events: 2820
ARM 2 · ACTIVE COMPARATOR

Ramipril

  • Active ramipril
  • Placebo of LCZ696
  • Safety set at risk for serious adverse events: 2816
Double-dummy masking. The registry lists both active drugs and a matching placebo for each (placebo of LCZ696 and placebo of ramipril). This is the standard double-dummy arrangement: every participant takes one active drug and one placebo, so participants, investigators, outcome assessors and the care team (quadruple masking) cannot infer the assignment from the tablets. The registry also lists valsartan and placebo of valsartan among the interventions; how these were used within the protocol is detailed in the registry entry rather than in its posted results.
Allocation
Randomized, two parallel arms
Masking
Quadruple (participant, care provider, investigator, outcomes assessor)
Hypothesis
Superiority of LCZ696 over ramipril for every posted analysis
Event confirmation
Endpoint events confirmed by a Clinical Endpoint Committee (CEC)

4. Analysis Populations

Efficacy analyses posted on ClinicalTrials.gov use the Full Analysis Set for every endpoint. Safety is summarized in a separate safety set restricted to participants who actually took study treatment.

Analysis populationDefinition / role
Full Analysis SetRandomized participants analyzed by assigned treatment; used for the primary and all secondary efficacy analyses.
Safety setParticipants who took study treatment; used for adverse-event summaries (2820 LCZ696, 2816 ramipril at risk). The registry notes that 25 patients did not take study treatment after randomization and were therefore excluded from the safety set.

The distinction matters. The Full Analysis Set follows the intention-to-treat principle closely: patients stay in the arm they were randomized to regardless of adherence, preserving the comparability that randomization creates. The safety set answers a different question — what happened to people exposed to each drug — so it deliberately excludes patients who never took a dose.

5. Endpoints

Primary endpoint

EndpointRegistry definitionTime frame
Number of participants with first CEC-confirmed primary composite endpointA confirmed composite endpoint includes cardiovascular (CV) death, heart failure (HF) hospitalization, or outpatient heart failureFrom randomization to first occurrence (up to approximately 43 months)

Although the registry titles the measure as a count of participants, the posted analysis is a time-to-first-event comparison: each participant contributes time from randomization until the first adjudicated component event or censoring, and the treatment effect is expressed as a hazard ratio from a Cox proportional hazards model. The registry also posts separate Cox analyses for each component (CV death, first HF hospitalization, first outpatient HF) under the primary outcome.

Secondary endpoints with posted analyses

EndpointTime frameMethod / effect measure
Confirmed composite of CV death or HF hospitalizationTime from randomization to first occurrence (up to approximately 43 months)Cox model / HR
Confirmed composite of HF hospitalization or outpatient HFTime from randomization to first occurrence (approximately up to 43 months)Cox model / HR
Confirmed composite of CV death, non-fatal spontaneous MI or non-fatal strokeTime from randomization to first occurrence (approximately up to 43 months)Cox model / HR
Total number of confirmed composite endpointsTime from randomization to end of study (approximately up to 43 months)Negative binomial regression / rate ratio
All-cause mortality (Full Analysis Set)Time from randomization to death (approximately up to 43 months)Cox model / HR

6. Primary Endpoint Results

The primary analysis compared LCZ696 with ramipril in the Full Analysis Set using a Cox proportional hazards model under a superiority hypothesis, with a two-sided 95% confidence interval.

Hazard ratio for first CEC-confirmed CV death, HF hospitalization or outpatient HF

HR 0.9012

95% CI: 0.7778–1.0441   ·   P = 0.1659

LCZ696 (sacubitril/valsartan) vs ramipril · Full Analysis Set · Cox proportional hazards model

Clinical Biostats interpretation

What the estimate means. An HR of 0.9012 means that, under the fitted Cox model and averaged over follow-up, the estimated instantaneous rate of a first primary event in the LCZ696 group was about 90% of that in the ramipril group — roughly a 10% lower estimated hazard.

What it does not mean. It does not mean that 10% fewer patients had an event, that 10% of patients benefited, or that each patient's risk fell by 10%. It is a relative, model-based summary of event rates, not an absolute risk difference; the registry analysis does not translate it into absolute event proportions.

Precision. The 95% CI of 0.7778 to 1.0441 spans 1. The data are compatible with a hazard about 22% lower with LCZ696 and also with a hazard about 4% higher. The trial therefore did not establish superiority, but neither does the interval establish that the two drugs are equivalent: equivalence would require a prespecified margin and an interval lying entirely within it.

The p-value. P = 0.1659 indicates that data at least this extreme would not be unusual if the true HR were 1. It is not a measure of how large the effect is, and a non-significant p-value is not evidence that the effect is zero. "Absence of evidence" for superiority is the correct reading.

Cautions. A single HR assumes that the ratio of hazards is roughly constant over the up to approximately 43 months of follow-up; the registry does not report a check of proportional hazards. Censoring (for example, at end of study) is assumed to be non-informative. Because the primary comparison did not reach statistical significance, all subsequent comparisons on this page should be read as supportive or exploratory rather than confirmatory.

Components of the primary composite

The registry posts a separate Cox analysis for each component of the primary composite, each analysing the time to the first occurrence of that specific event type.

AnalysisHR (LCZ696 vs ramipril)95% CIP-value
Primary composite0.90120.7778–1.04410.1659
CV death0.87400.7104–1.07540.2031
First HF hospitalization0.86530.7044–1.06290.1679
First outpatient HF0.68310.4546–1.02660.0667
Clinical Biostats interpretation

All three component HRs lie below 1 and point in the same direction as the composite, and every 95% CI includes 1. The outpatient HF component has the lowest point estimate (0.6831) but also the widest interval (0.4546 to 1.0266), which is the typical signature of a component with fewer events: less information produces a less stable estimate.

Component analyses are not independent tests of efficacy. They describe which parts of the composite drive the overall result and whether the components behave consistently. Selecting the component with the smallest p-value and presenting it as a finding would inflate the false-positive rate. In addition, when CV death is analysed alongside non-fatal events, death acts as a competing event for the non-fatal components: a patient who dies can no longer be hospitalized, which complicates the interpretation of cause-specific hazard ratios for HF hospitalization and outpatient HF.

Posted hazard / rate ratios with 95% CIs (vertical line = 1.0; values <1 favour LCZ696)
Primary composite
0.9012
CV death
0.8740
First HF hospitalization
0.8653
First outpatient HF
0.6831
CV death or HF hospitalization
0.9133
HF hospitalization or outpatient HF
0.8422
CV death, non-fatal MI or stroke
0.9007
Total composite events (rate ratio)
0.8361
All-cause death
0.8751
0.40.60.81.01.2

7. Secondary Endpoint Results

All secondary analyses were posted for the Full Analysis Set under a superiority hypothesis with two-sided 95% confidence intervals.

Secondary endpointMethodEstimate95% CIP-value
CV death or HF hospitalizationCox modelHR 0.91330.7824–1.06620.2507
HF hospitalization or outpatient HFCox modelHR 0.84220.6979–1.01630.0732
CV death, non-fatal spontaneous MI or non-fatal strokeCox modelHR 0.90070.7734–1.04890.1785
Total number of confirmed composite endpointsNegative binomial regressionRate ratio 0.83610.7018–0.99610.0452
All-cause deathCox modelHR 0.87510.7279–1.05200.1556

Total (first and recurrent) composite events

Rate ratio for total confirmed composite endpoints

RR 0.8361

95% CI: 0.7018–0.9961   ·   P = 0.0452

Negative binomial regression · Full Analysis Set · randomization to end of study

This is the only posted analysis whose 95% CI excludes 1. A rate ratio of 0.8361 means that the estimated rate of confirmed composite events (counting every event, not only the first) was about 16% lower with LCZ696 than with ramipril. The upper confidence limit, 0.9961, lies just below 1, and the p-value of 0.0452 lies just below 0.05, so the result is close to the conventional threshold.

Why this result should not be read as confirmatory: the primary endpoint did not reach statistical significance. In a trial with a prespecified primary hypothesis, secondary endpoints are usually interpreted within a testing strategy that controls the overall type I error, and under a typical fixed-sequence approach testing would stop once the primary endpoint fails. The ClinicalTrials.gov record does not describe a multiplicity-adjustment strategy, and the posted p-value of 0.0452 is unadjusted. Among the nine posted comparisons, one nominal p-value below 0.05 is the kind of finding that can arise from multiple testing, so it is best viewed as hypothesis-generating.

The remaining secondary HRs are all below 1 and all have intervals that include 1. The consistency of direction across endpoints is informative descriptively, but these composites share many of the same events (CV death appears in several of them), so they are strongly correlated and do not represent independent confirmations of one another.

8. Safety: Serious Adverse Events

Serious adverse events were summarized in the safety set, which excludes the 25 patients who did not take study treatment after randomization.

Safety measureLCZ696 (sacubitril/valsartan)Ramipril
Participants with serious adverse events (affected / at risk)1146 / 28201126 / 2816

The counts of participants with at least one serious adverse event are very similar in the two arms relative to arms of nearly identical size. No formal statistical comparison of serious adverse events was posted to ClinicalTrials.gov; adverse-event tables are typically descriptive, and many categories are compared informally because the trial was not powered to test safety differences. Serious adverse events here are a participant-level count ("any SAE"), so they do not convey severity, recurrence or specific event types.

9. Statistical Methodology

Time-to-first-event analysis with the Cox proportional hazards model

The primary endpoint, its components and four of the five secondary endpoints were analysed with Cox's proportional hazards model. The model compares the instantaneous event rate (the hazard) between the two arms without specifying the shape of the underlying baseline hazard, and summarizes the treatment effect as a single hazard ratio.

Conceptual form
h(t | treatment) = h0(t) · exp(β · treatment),    HR = exp(β)

where h0(t) is the unspecified baseline hazard in the ramipril arm and β is the log hazard ratio for LCZ696. The key assumption is that the ratio of hazards is constant over time. The registry does not specify stratification factors or covariates for the model.

Negative binomial regression for total events

Time-to-first-event analyses discard every event after a patient's first one. To use all adjudicated events, the total number of confirmed composite endpoints was analysed with a negative binomial regression model, which estimates the ratio of event rates between arms.

Conceptual form
log E[Yi] = log(follow-upi) + α + β · treatmenti,    Var(Yi) = μi + kμi2,    RR = exp(β)

where Yi is a patient's number of events, follow-up time is an offset and k is the dispersion parameter. The extra variance term allows for the fact that some patients have many events and most have none, which a Poisson model would ignore and which would make its confidence intervals too narrow.

Superiority framework and two-sided testing

Every posted analysis is labelled as a superiority hypothesis with a two-sided 95% confidence interval. The null hypothesis is HR = 1 (or rate ratio = 1); rejection requires the interval to exclude 1. A superiority design that fails to reject does not license a claim of non-inferiority or equivalence, because no non-inferiority margin is reported in the registry.

Adjudicated endpoints

All efficacy events were confirmed by a Clinical Endpoint Committee. Blinded central adjudication applies uniform definitions across sites and reduces misclassification, which otherwise tends to bias hazard ratios toward 1. Combined with quadruple masking, it limits the risk that knowledge of treatment influences event ascertainment.

Composite endpoints

Combining CV death, HF hospitalization and outpatient HF increases the number of events and therefore the statistical power. The trade-off is that components of different clinical severity are weighted equally in a time-to-first-event analysis: an outpatient HF episode counts the same as a CV death if it occurs first.

Multiplicity

The registry posts nine comparisons (the primary composite, three components and five secondary analyses). The record does not describe how the familywise type I error was controlled across them, so the p-values shown should be treated as nominal. Only the primary endpoint carries a clear confirmatory interpretation.

AnalysisRoleInterpretation
Primary compositePrimary endpointConfirmatory superiority test; not statistically significant
Individual componentsPosted under the primary outcomeDescriptive; show consistency of direction across components
Secondary composites and all-cause deathSecondary endpointsSupportive; nominal p-values
Total composite eventsSecondary endpointNominally significant; hypothesis-generating given the non-significant primary result

10. Statistical Methods Explained

Why was a Cox proportional hazards model used for the primary endpoint?

The primary endpoint is the time to a first event, and patients were followed for different lengths of time (up to approximately 43 months). The Cox model handles right-censored follow-up naturally and gives a single relative effect, the hazard ratio, without requiring any assumption about the shape of the event-rate curve over time. It is the standard analysis for time-to-first-event composites in cardiovascular outcome trials.

What does an HR of 0.9012 with a 95% CI of 0.7778 to 1.0441 mean in practice?

The best single estimate is a hazard about 10% lower with LCZ696, but the interval runs from about 22% lower to about 4% higher. Because it includes 1, the trial cannot exclude no difference between the drugs. It also cannot exclude a moderate benefit, which is why the result is described as "not statistically significant" rather than "no effect."

Why was the total-events analysis done with negative binomial regression rather than a Cox model?

Patients after a myocardial infarction can have repeated heart failure events. A Cox time-to-first-event model stops counting at the first event. Negative binomial regression models the count of all events per patient relative to follow-up time and allows for overdispersion, meaning that events cluster in a subset of high-risk patients. Its effect measure is a rate ratio (0.8361 here) rather than a hazard ratio.

Why can the total-events result be nominally significant when the first-event result is not?

Counting recurrent events adds information, so the rate ratio estimate is more precise and slightly further from 1 (0.8361 versus 0.9012). This can push the interval just below 1. However, the two analyses answer different questions, and when the primary endpoint fails, a secondary endpoint with P = 0.0452 cannot rescue the trial's confirmatory conclusion; without a prespecified multiplicity strategy it remains supportive evidence.

Why do the component hazard ratios look more favourable than the composite, yet none is significant?

Each component has fewer events than the composite, which widens its confidence interval. The first outpatient HF analysis illustrates this: HR 0.6831 but a 95% CI of 0.4546 to 1.0266. A lower point estimate from a less precise analysis is not stronger evidence. It reflects greater uncertainty.

Why does the safety set contain fewer patients than were enrolled?

Efficacy was analysed in the Full Analysis Set according to randomized assignment, whereas safety was analysed only among patients who took study treatment. The registry reports that 25 patients did not take study treatment after randomization and were excluded from the safety set, leaving 2820 at risk in the LCZ696 arm and 2816 in the ramipril arm.

11. Limitations

12. Why This Trial Matters Statistically

PARADISE-MI is a clear teaching example of how to read a well-conducted trial whose primary endpoint did not reach statistical significance, and of how first-event and recurrent-event analyses of the same outcomes can lead to different nominal conclusions.

ConceptHow it appears in PARADISE-MI
Active-comparator superiorityLCZ696 tested against ramipril rather than placebo
Double-dummy maskingMatching placebos for each active drug support quadruple masking
Full Analysis SetEfficacy analysed by randomized assignment
Composite time-to-event endpointCV death, HF hospitalization or outpatient HF, first occurrence
Cox proportional hazards modelHazard ratios for the primary endpoint, components and most secondary endpoints
Confidence intervalsPrimary interval 0.7778–1.0441 spans 1: compatible with modest benefit or slight harm
P-values vs effect sizeP = 0.1659 for the primary endpoint does not indicate "no effect"
Negative binomial regressionRate ratio for total (recurrent) composite events
MultiplicityA nominal P = 0.0452 secondary result after a non-significant primary endpoint
Competing risksCV death precludes later non-fatal HF events in component analyses

Statistical interpretation

Every posted estimate lies below 1, but only the negative binomial analysis of total events has a 95% CI excluding 1, and that result is nominal. The prespecified primary superiority test was not significant.

Clinical interpretation

The trial did not demonstrate that LCZ696 reduces the first occurrence of CV death or heart failure events compared with ramipril after acute MI. Serious adverse event counts were similar between arms. The findings do not establish equivalence either.

13. Related Tutorials

Learn more about the methods used in this trial:

14. Related Calculators

15. Sources

Explore the methods behind the results

Connect this trial's endpoints and analyses to step-by-step statistical tutorials, calculators and other trial analyses.

16. Record Summary

PARADISE-MI randomized patients with acute myocardial infarction to LCZ696 (sacubitril/valsartan) or ramipril under quadruple masking, with 5669 enrolled. The primary Cox-model analysis of the first CEC-confirmed CV death, HF hospitalization or outpatient HF gave an HR of 0.9012 (95% CI 0.7778–1.0441; P = 0.1659), so superiority was not established. Component and secondary hazard ratios all fell below 1 with intervals including 1, and the negative binomial analysis of total composite events gave a nominally significant rate ratio of 0.8361 (95% CI 0.7018–0.9961; P = 0.0452). Serious adverse events occurred in 1146 of 2820 and 1126 of 2816 participants, respectively. The most useful reading combines the relative effect estimates, their confidence intervals, the distinction between first-event and total-event analyses and the role of the primary endpoint in controlling false-positive conclusions.

Clinical Biostats methodology: Each trial page separates reported results from educational interpretation, explains the statistical methods the trial actually used, and highlights the design features that determine how far the results can be generalized.