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
TRANSFORM-HF was a randomized phase 3, parallel-group trial comparing torsemide with furosemide in patients hospitalized for heart failure. The registry reports an enrollment of 2,973 participants, two treatment arms, an open-label design, and a primary endpoint of all-cause mortality measured by follow-up phone call or the National Death Index.
| Feature | TRANSFORM-HF |
|---|---|
| Phase | Phase 3 |
| Condition | Heart Failure |
| Design | Randomized, parallel-group, open-label |
| Allocation | Randomized |
| Primary purpose | Treatment |
| Enrollment | 2,973 |
| Arms | 2 |
| Interventions | Torsemide and furosemide |
| Primary endpoint | All-cause mortality, as measured by follow-up phone call or National Death Index (NDI) database. |
| Primary endpoint time frame | Up to 30 months |
| Hypothesis type | Superiority |
| Lead sponsor | Duke University |
| Study start | 2018-07-11 |
| Primary completion | 2022-07-29 |
2. Clinical Question
The central question was whether torsemide was superior to furosemide with respect to all-cause mortality among patients hospitalized for heart failure.
Population
Patients hospitalized for heart failure.
Intervention
Torsemide.
Comparator
Furosemide.
Primary question
Does torsemide provide a superior all-cause mortality outcome compared with furosemide through the registry-defined follow-up period?
3. Trial Design
Torsemide
- Torsemide
- Randomized treatment assignment
- Compared with furosemide for heart-failure management
Furosemide
- Furosemide
- Randomized treatment assignment
- Reference group in the reported regression analyses
4. Trial Timeline
Study start
The registry identifies July 11, 2018 as the study start date.
Primary mortality follow-up
The registered primary endpoint measured all-cause mortality through follow-up phone calls and National Death Index searches.
Primary completion
The registry identifies July 29, 2022 as the primary completion date.
5. Primary Endpoint
| Endpoint | Registry definition | Time frame | Analysis |
|---|---|---|---|
| All-cause mortality | All-cause mortality, as measured by follow-up phone call or NDI | Up to 30 months | Cox proportional-hazards model |
The registry defines death ascertainment through the Call Center and through searches of the National Death Index. Follow-up duration varied according to enrollment sequence: death information was obtained through 30 months for patients 1-500, through 24 months for patients 501-1000, through 18 months for patients 1001-1500, and through 12 months for the remaining patients, except those randomized less than 12 months before the study was stopped by the DSMB.
The primary analysis used a Cox proportional-hazards regression model comparing torsemide with furosemide. The model included treatment assignment together with age, sex, baseline ejection fraction category, and loop diuretic treatment before the index hospital admission.
6. Primary Result: All-Cause Mortality
Hazard ratio for all-cause mortality
95% CI: 0.885–1.180 · P = 0.765
Torsemide vs furosemide; two-sided superiority analysis.
| Measure | Reported result |
|---|---|
| Effect measure | Hazard ratio |
| Torsemide vs furosemide HR | 1.022 |
| 95% CI | 0.885–1.180 |
| P-value | 0.765 |
| Model | Cox proportional-hazards regression |
| Covariates | Age, sex, baseline ejection fraction, and loop diuretic treatment prior to index hospital admission |
The estimated hazard ratio of 1.022 means that the fitted model estimated the instantaneous rate of death to be approximately 2.2% higher with torsemide than with furosemide over the analyzed follow-up. This is a relative model-based hazard comparison; it is not a statement that 2.2% more patients died, nor does it describe an individual's probability of death.
The 95% confidence interval of 0.885–1.180 spans 1.00. Thus, the data are compatible with a range of relative hazard differences in either direction under the model, rather than providing a precise estimate of a treatment effect away from the null.
The P = 0.765 value addresses the statistical evidence against the null hypothesis specified for the superiority comparison. It does not measure the size or clinical importance of the estimated effect. The estimate itself and its confidence interval are therefore essential for interpretation.
Because this is a Cox model, the hazard ratio is also subject to the model's proportional-hazards interpretation. A single HR summarizes a relative hazard relationship across follow-up and should not be read as a constant individual-level risk ratio.
7. Secondary Time-to-Event Endpoint: Mortality or Hospitalization Through 12 Months
| Endpoint | Time frame | Effect | 95% CI | P-value |
|---|---|---|---|---|
| All-cause mortality or all-cause hospitalization | Up to 12 months | HR 0.918 | 0.826–1.020 | 0.113 |
| All-cause mortality or first all-cause hospitalization | 30 days | HR 0.943 | 0.753–1.181 | 0.610 |
Mortality or hospitalization through month 12
Hazard ratio
95% CI: 0.826–1.020 · P = 0.113
Cox proportional-hazards model with covariate adjustment.
An HR of 0.918 corresponds to an estimated instantaneous event rate approximately 8.2% lower with torsemide than with furosemide under the fitted model. That interpretation concerns the combined time-to-first-event endpoint, not mortality alone.
The 95% CI of 0.826–1.020 includes 1.00, so the confidence interval includes both a modest reduction and values close to no relative difference. The P-value of 0.113 describes the statistical evidence under the superiority hypothesis; it is not a measure of how clinically meaningful an HR of 0.918 would be.
Mortality or first hospitalization at 30 days
Hazard ratio
95% CI: 0.753–1.181 · P = 0.610
Cox proportional-hazards model with covariate adjustment.
The 30-day HR of 0.943 corresponds to an estimated instantaneous rate of the composite event approximately 5.7% lower with torsemide under the fitted Cox model. The confidence interval is substantially wider than the interval for the longer follow-up composite endpoint, extending from 0.753 to 1.181.
That interval reflects the statistical precision of the estimate rather than the range of individual patient outcomes. The P = 0.610 value does not quantify the magnitude of the treatment difference; the HR and confidence interval provide that information.
8. Secondary Hospitalization Analysis
The registry also reports total re-hospitalizations through month 12. The analysis population is defined as participants who had at least one re-hospitalization.
Reported risk ratio
95% CI: 0.836–1.068 · P = 0.366
Torsemide vs furosemide.
| Feature | Registry analysis |
|---|---|
| Endpoint | Total re-hospitalizations |
| Time frame | Up to 12 months |
| Analysis population | Participants who had at least one re-hospitalization |
| Reported method | Linear regression |
| Statistical model described in analysis notes | Negative binomial regression |
| Effect measure | Risk ratio |
| Estimate | 0.945 |
| 95% CI | 0.836–1.068 |
| P-value | 0.366 |
The analysis notes describe a negative binomial regression of the frequency of re-hospitalizations, using an offset equal to the log of each participant's months to the last re-hospitalization assessment through month 12. Re-hospitalizations greater than or equal to 5 were combined and analyzed as one category, with the offset for participants reaching that category based on the months to the fifth re-hospitalization.
The reported risk ratio of 0.945 corresponds to an estimated relative value approximately 5.5% lower with torsemide than with furosemide. The 95% CI of 0.836–1.068 crosses 1.00, so the interval includes both a relative reduction and a relative increase in the modeled quantity.
An important statistical detail is that this is a recurrent-event/frequency outcome, not simply whether a participant was hospitalized at least once. The analysis notes describe an exposure-time offset, which allows the frequency of re-hospitalizations to be related to the amount of follow-up available for assessment.
9. Quality-of-Life Results
Health-related quality of life was evaluated using change in KCCQ-CSS from baseline at 1 month, 6 months, and 12 months. The analysis population consisted of participants with both baseline and monthly KCCQ-CSS scores available so that change from baseline could be analyzed.
| Assessment | Mean difference | 95% CI | P-value |
|---|---|---|---|
| 1 month | 1.30 | -0.66 to 3.26 | 0.194 |
| 6 months | -0.39 | -2.53 to 1.74 | 0.718 |
| 12 months | 0.02 | -2.26 to 2.30 | 0.987 |
Why a mixed-effects model was used
The registry reports mixed-model repeated-measures analyses of KCCQ score change. This is appropriate for a longitudinal endpoint because the same participants can contribute observations at multiple visits. Treating those observations as completely independent would ignore the within-participant correlation that naturally arises when repeated measurements are taken from the same person.
The registry states that least-squares means from the mixed-model repeated-measures analysis were used, with treatment, visit, treatment-by-visit interaction, baseline KCCQ-CSS, age, sex, baseline ejection fraction, and prior loop diuretic treatment included as terms.
At 1 month, the estimated mean difference was 1.30, with a 95% CI of -0.66 to 3.26. At 6 months, it was -0.39, with a 95% CI of -2.53 to 1.74. At 12 months, it was 0.02, with a 95% CI of -2.26 to 2.30.
Each confidence interval includes zero, the null value for a mean difference. The P-values of 0.194, 0.718, and 0.987 correspond to the respective time-point analyses. They should not be interpreted as measures of the magnitude of the quality-of-life difference.
Because the analysis is longitudinal and covariate-adjusted, the reported estimates are model-based comparisons rather than simple unadjusted differences in observed means. The analysis population also requires availability of baseline and monthly KCCQ-CSS measurements, which is distinct from the full randomized population.
10. Depression Symptom Results
Symptoms of depression were assessed using the Patient Health Questionnaire-2 (PHQ-2) at month 1, month 6, and month 12. The analysis population consisted of participants with data available at each time point.
| Assessment | Method | P-value |
|---|---|---|
| Month 1 | Wilcoxon / Mann-Whitney | 0.904 |
| Month 6 | Wilcoxon / Mann-Whitney | 0.829 |
| Month 12 | Wilcoxon / Mann-Whitney | 0.342 |
The registry reports no effect estimate or confidence interval for these analyses. The posted statistical test is the Wilcoxon (Mann-Whitney) test, a nonparametric method for comparing distributions between two groups.
The P-values of 0.904, 0.829, and 0.342 describe the statistical evidence from the respective Wilcoxon / Mann-Whitney comparisons at months 1, 6, and 12. A P-value alone does not indicate the size, direction, or clinical importance of a difference.
Because no effect estimate or confidence interval is reported for these analyses, the statistical results cannot be reduced to a quantitative treatment-effect estimate from the registry entry alone. The choice of a nonparametric test also means that the analysis is not naturally expressed as a difference in arithmetic means.
11. Statistical Methodology
Cox proportional-hazards model
The primary mortality analysis and the two composite time-to-event analyses used Cox proportional-hazards regression. Cox regression is designed for outcomes defined by both whether an event occurs and when it occurs. This is particularly useful when participants have different lengths of follow-up or when some participants have not experienced the event by the end of observation.
The treatment hazard ratio is obtained by exponentiating the treatment coefficient. In this trial, the registry describes additional adjustment for age, sex, baseline ejection fraction, and prior loop diuretic treatment.
Hazard ratio
A hazard ratio compares the modeled instantaneous event rates between treatment groups. An HR of 1 corresponds to equal hazards under the model, while values below or above 1 indicate lower or higher estimated hazards, respectively.
| Endpoint | HR | Interpretation of point estimate |
|---|---|---|
| All-cause mortality | 1.022 | Approximately 2.2% higher estimated hazard with torsemide |
| Mortality or hospitalization through 12 months | 0.918 | Approximately 8.2% lower estimated hazard with torsemide |
| Mortality or first hospitalization at 30 days | 0.943 | Approximately 5.7% lower estimated hazard with torsemide |
These are transformations of the reported HR point estimates and describe relative hazards, not absolute risk reductions. An HR of 0.918, for example, does not mean that 8.2% fewer patients experienced the event.
Covariate adjustment
The reported Cox analyses adjusted for age, sex, baseline ejection fraction category, and loop diuretic treatment before the index hospital admission. Covariate adjustment can improve precision and account for prespecified prognostic factors, but it does not change the fundamental randomized comparison.
Baseline ejection fraction was represented using the registry-defined categories <40%, 41-49%, >50%, and unknown.
Negative binomial regression for re-hospitalization frequency
The re-hospitalization analysis is distinct from a simple binary hospitalization endpoint because participants can experience multiple hospitalizations. The registry's analysis notes describe negative binomial regression with an offset for the time available for assessment. This structure accounts for differences in observation time and accommodates frequency data that may be more dispersed than a simple Poisson model assumes.
Mixed-effects model for repeated KCCQ measurements
The quality-of-life analysis used a mixed-model repeated-measures approach. Its treatment-by-time interaction allows the estimated treatment difference to vary across visits rather than forcing a single treatment effect to apply identically at 1, 6, and 12 months.
Wilcoxon / Mann-Whitney test
The PHQ-2 analyses used the Wilcoxon (Mann-Whitney) test at each reported time point. This nonparametric two-group comparison is useful when the outcome distribution does not satisfy assumptions commonly associated with parametric mean-based comparisons.
12. Statistical Methods Explained
Why was a Cox model used for mortality?
Mortality is a time-to-event endpoint: both the occurrence of death and the timing of death matter. Cox regression can use the available follow-up time while accounting for participants who are censored before experiencing the event. It also allows the registry's specified covariates to be included in the analysis.
What does the primary HR of 1.022 mean?
It is the estimated ratio of the instantaneous mortality hazards for torsemide versus furosemide under the fitted Cox model. Because 1.022 is close to 1, the point estimate represents a small relative difference in the modeled hazard. The confidence interval is more informative about uncertainty than the point estimate alone.
Why does the confidence interval matter?
The 95% CI provides an interval representation of statistical uncertainty around the estimated treatment effect. For the primary HR, the interval is 0.885–1.180. Because it includes 1, the interval is compatible with both a lower and a higher mortality hazard for torsemide relative to furosemide under the model.
Why use a mixed-effects model for KCCQ?
KCCQ measurements were collected repeatedly at baseline, 1 month, 6 months, and 12 months. Repeated measurements from the same participant are correlated. A mixed-effects framework can account for this longitudinal structure while estimating treatment differences over time and adjusting for baseline KCCQ-CSS and other prespecified covariates.
Why is the re-hospitalization analysis different from the mortality analysis?
Death is a time-to-first-event outcome, whereas re-hospitalization can occur repeatedly. A frequency model can therefore represent the number of re-hospitalizations and incorporate the amount of time available for observation through an offset. The registry describes this approach using negative binomial regression.
What does a P-value tell us here?
A P-value measures the compatibility of the observed statistical result with a specified null hypothesis under the assumptions of the analysis. It does not quantify the size of a treatment effect, the probability that a treatment works, or the probability that the null hypothesis is true. Effect estimates and confidence intervals are needed to describe magnitude and precision.
13. Endpoint-by-Endpoint Statistical Summary
| Endpoint | Time frame | Method | Effect | 95% CI | P-value |
|---|---|---|---|---|---|
| All-cause mortality | Up to 30 months | Cox proportional-hazards model | HR 1.022 | 0.885–1.180 | 0.765 |
| All-cause mortality or hospitalization | Up to 12 months | Cox proportional-hazards model | HR 0.918 | 0.826–1.020 | 0.113 |
| Total re-hospitalizations | Up to 12 months | Negative binomial regression described in analysis notes | RR 0.945 | 0.836–1.068 | 0.366 |
| All-cause mortality or hospitalization | 30 days | Cox proportional-hazards model | HR 0.943 | 0.753–1.181 | 0.610 |
| KCCQ-CSS change | 1 month | Mixed-effects model | Mean difference 1.30 | -0.66 to 3.26 | 0.194 |
| KCCQ-CSS change | 6 months | Mixed-effects model | Mean difference -0.39 | -2.53 to 1.74 | 0.718 |
| KCCQ-CSS change | 12 months | Mixed-effects model | Mean difference 0.02 | -2.26 to 2.30 | 0.987 |
| PHQ-2 depression symptoms | Month 1 | Wilcoxon / Mann-Whitney | Not reported | Not reported | 0.904 |
| PHQ-2 depression symptoms | Month 6 | Wilcoxon / Mann-Whitney | Not reported | Not reported | 0.829 |
| PHQ-2 depression symptoms | Month 12 | Wilcoxon / Mann-Whitney | Not reported | Not reported | 0.342 |
14. Understanding the Primary Confidence Interval
The primary HR of 1.022 is the single best-fitting treatment-effect estimate produced by the reported Cox model. It should not be interpreted in isolation because any estimated treatment effect is subject to sampling uncertainty.
The 95% CI of 0.885–1.180 shows the uncertainty surrounding the primary HR. The interval spans the null value of 1.00 and therefore includes values corresponding to both lower and higher estimated hazards with torsemide.
A hazard ratio is relative. It does not tell us the absolute probability of death in either group. Absolute event risks would require the corresponding event counts or survival estimates, which are not reported in the statistical results summarized here.
15. Superiority Testing and the Null Value
The registry identifies the analyses as superiority hypotheses. For the primary hazard ratio, the null value is 1.00 because the effect measure is a ratio. A value below 1 indicates a lower modeled hazard with torsemide; a value above 1 indicates a higher modeled hazard.
| Effect measure | Null value | Direction below null | Direction above null |
|---|---|---|---|
| Hazard ratio | 1.00 | Lower modeled hazard | Higher modeled hazard |
| Risk ratio | 1.00 | Lower relative risk | Higher relative risk |
| Mean difference | 0 | Lower mean in treatment group | Higher mean in treatment group |
This distinction is important because the null value depends on the effect measure. A confidence interval containing 1.00 has a different interpretation from one containing 0, even though both represent inclusion of the relevant null value.
16. Censoring and Follow-up in the Mortality Endpoint
The primary endpoint is explicitly a time-to-event outcome. Participants contribute information until death or until their available follow-up ends. The registry describes different maximum follow-up periods according to enrollment sequence, with mortality information obtained through phone follow-up and National Death Index searches.
In survival analysis, censoring does not mean that a participant is treated as having survived indefinitely. Instead, the participant contributes information over the period for which the event status is known. The Cox model then uses the observed event and follow-up structure to estimate relative hazards.
17. Analysis Populations and Missing Measurements
The registry identifies specific analysis populations for several secondary endpoints. For total re-hospitalizations, the analysis population consists of participants who had at least one re-hospitalization. For KCCQ-CSS change, participants needed both baseline and monthly KCCQ-CSS scores available. For PHQ-2, the analysis population consists of participants with data available at each time point.
| Endpoint | Analysis population described in the registry |
|---|---|
| All-cause mortality | Patients followed through phone contact and/or National Death Index ascertainment |
| Mortality or hospitalization | Patients evaluated for the composite time-to-event endpoint |
| Total re-hospitalizations | Participants with at least one re-hospitalization |
| KCCQ-CSS change | Participants with both baseline and monthly KCCQ-CSS scores available |
| PHQ-2 | Participants with data available at each time point |
These distinctions matter because a secondary endpoint analyzed among participants with available measurements is not necessarily estimating exactly the same randomized-population quantity as the primary ITT-style treatment comparison would. The interpretation should therefore remain tied to the population specified for each analysis.
18. Statistical Interpretation Across Endpoint Types
Time-to-event outcomes
The mortality and composite endpoints incorporate both whether an event occurred and when it occurred. Cox regression expresses the comparison as a hazard ratio.
Recurrent hospitalization
The re-hospitalization analysis concerns event frequency rather than simply the first occurrence of hospitalization and uses a time-related offset.
Quality of life
KCCQ-CSS is measured repeatedly. The mixed-effects model accounts for longitudinal measurements and includes treatment-by-time interaction.
Depression symptoms
PHQ-2 analyses use Wilcoxon / Mann-Whitney tests at individual follow-up time points, with P-values reported without corresponding effect estimates or confidence intervals.
19. What the Results Do — and Do Not — Establish
The primary mortality analysis produced an HR of 1.022 with a 95% CI of 0.885–1.180 and P = 0.765. Statistically, the point estimate is close to the null value of 1, while the confidence interval spans values on both sides of the null.
The 12-month mortality-or-hospitalization analysis produced an HR of 0.918, with a 95% CI of 0.826–1.020 and P = 0.113. The 30-day composite analysis produced an HR of 0.943, with a 95% CI of 0.753–1.181 and P = 0.610.
The quality-of-life estimates were close to zero at each reported assessment: 1.30 at 1 month, -0.39 at 6 months, and 0.02 at 12 months. Their corresponding confidence intervals all included zero.
These statistical results describe the reported treatment comparisons. They do not by themselves establish how an individual patient will respond, nor do they convert relative effect measures into absolute individual risks.
20. Limitations
- Open-label design: the registry identifies the trial as unmasked. This is particularly relevant to patient-reported outcomes such as KCCQ-CSS and PHQ-2, where knowledge of treatment assignment can potentially influence responses.
- Endpoint-specific analysis populations: several secondary analyses are restricted to participants with available measurements or qualifying re-hospitalization histories, so their populations differ from the overall enrolled population.
- Repeated secondary analyses: quality-of-life outcomes were assessed at multiple time points, while depression symptoms were also evaluated at multiple time points. The reported P-values should be interpreted in the context of these multiple assessments rather than as isolated tests without regard to the broader analysis program.
- Hazard-ratio assumptions: Cox regression produces a model-based hazard ratio whose interpretation depends on the proportional-hazards framework. A single HR does not necessarily describe every point in time if the relative hazards vary substantially.
- Composite endpoints: mortality-or-hospitalization endpoints combine distinct clinical events. A composite HR does not show whether its components contribute equally or whether an observed difference is driven predominantly by one component.
- Quality-of-life availability: the KCCQ analysis requires baseline and monthly scores, so missing measurements affect which participants contribute to those estimates.
- Limited effect reporting for PHQ-2: the registry reports P-values for the Wilcoxon / Mann-Whitney analyses but no corresponding effect estimates or confidence intervals.
21. Why This Trial Matters Statistically
TRANSFORM-HF is a useful statistical teaching case because it combines several common clinical-trial analysis structures within one randomized comparison. The primary endpoint is a time-to-event outcome analyzed with covariate-adjusted Cox regression, while secondary endpoints require different statistical frameworks for composite time-to-event outcomes, recurrent hospitalization, repeated quality-of-life measurements, and nonparametric comparisons.
| Concept | How it appears in TRANSFORM-HF |
|---|---|
| Randomization | Randomized parallel-group comparison of torsemide and furosemide |
| Time-to-event analysis | All-cause mortality and mortality-or-hospitalization endpoints |
| Cox proportional-hazards model | Primary mortality and composite endpoint analyses |
| Hazard ratio | Primary and secondary time-to-event effect measure |
| Covariate adjustment | Age, sex, baseline ejection fraction, and prior loop diuretic treatment |
| Recurrent events | Total re-hospitalization frequency through month 12 |
| Negative binomial regression | Frequency model described for re-hospitalizations |
| Mixed-effects model | Repeated KCCQ-CSS change measurements |
| Treatment-by-time interaction | Included in the repeated-measures KCCQ model |
| Nonparametric testing | Wilcoxon / Mann-Whitney analyses for PHQ-2 |
| Confidence intervals | Reported for HRs, risk ratio, and KCCQ mean differences |
| P-values | Reported across primary and secondary analyses |
22. Related Tutorials
Learn more about the methods used in this trial:
23. Related Calculators
24. Sources
- ClinicalTrials.gov: NCT03296813.
- PubMed: PubMed record 42419438.
- PubMed: PubMed record 39196651.
- PubMed: PubMed record 38745502.
- PubMed: PubMed record 38436075.
- PubMed: PubMed record 37212600.
Continue through the Clinical Biostats statistical library
Explore the methods underlying randomized trials, survival analysis, longitudinal models, and treatment-effect estimation.
25. Record Summary
TRANSFORM-HF provides a compact example of how a randomized clinical trial can require several distinct statistical frameworks. Its primary all-cause mortality endpoint was analyzed using a covariate-adjusted Cox proportional-hazards model, producing an HR of 1.022 with a 95% CI of 0.885–1.180 and P = 0.765. Secondary mortality-or-hospitalization endpoints were also analyzed using Cox regression, while total re-hospitalizations were evaluated with a frequency model described in the registry analysis notes and quality-of-life change was evaluated using mixed-model repeated measures.
The quality-of-life estimates were 1.30, -0.39, and 0.02 at 1, 6, and 12 months, respectively, with confidence intervals spanning zero at each time point. PHQ-2 analyses used Wilcoxon / Mann-Whitney tests, with P-values of 0.904, 0.829, and 0.342 at months 1, 6, and 12.