A lifetime risk model for HFpEF, and an honest C-statistic

Design

prognostic model development with external validation in two trials and three registries

Population

20,332 patients with ejection fraction 50% or more (derivation); 28,062 externally

Primary outcome

heart failure hospitalisation or cardiovascular death

Effect

pooled C-statistic 0.714 (95% CI 0.652-0.775) in trials, 0.658 (95% CI 0.599-0.717) in registries, calibration adequate

LIFE-Preserved was derived in 20,332 patients aged 40 to 90 with an ejection fraction of 50% or more from the Swedish Heart Failure Registry, using 14 routinely available predictors in cause- and sex-specific Cox models. Using age as the timescale lets it project beyond the observed follow-up, so it produces a lifetime estimate as well as a short-term one, adjusted for competing risks. External validation drew on two trials and three registries totalling 28,062 patients.

Performance was pooled C-statistic 0.714 (95% CI 0.652-0.775) in the trials and 0.658 (95% CI 0.599-0.717) in the registries, with adequate calibration everywhere and similar results in men and women. That is moderate discrimination, and it should be read as a statement about HFpEF rather than about the model: a syndrome this heterogeneous does not sort neatly by 14 variables.

Where it earns its place is the consultation rather than the algorithm. Calibration - the model getting the average right - is what matters when the purpose is to tell a patient what the next few years look like, and that held across five external sources. Use it to frame shared decisions about how hard to push preventive treatment, not to decide who gets it.

• Use the model where calibration matters - explaining prognosis - rather than as a gate for treatment.
• Check that the 14 predictors you would enter are actually recorded in your clinic before relying on it.
• Say 'moderate accuracy' to patients: this ranks risk imperfectly, it does not predict an individual's course.
• Note the derivation population is Swedish registry-based; Indian HFpEF cohorts are younger with more diabetes and rheumatic disease, so calibration locally is unverified.
• Recheck the estimate after a hospitalisation rather than treating the first number as fixed.

Why it matters

It gives HFpEF something it has lacked: a defensible individual prognosis to put in front of a patient who asks what the next ten years hold.

Don't overread it

Moderate discrimination means this ranks groups better than it predicts individuals; it has not been shown to improve outcomes when used.

The statistics, in plain English

A C-statistic of 0.66 to 0.71 means that if you pick one patient who will have an event and one who will not, the model ranks them correctly about two-thirds of the time. That is well short of the 0.8 threshold people associate with a good diagnostic test, but risk models for heterogeneous syndromes rarely beat it. Calibration - whether predicted risk matches observed risk on average - is the property that matters for counselling, and that was adequate in every external dataset.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.