Article · Biomarker Science

Prognostic Biomarker Discovery in Heart Failure

Predicting cardiovascular death among patients who already have heart failure, using multimodal machine learning on UK Biobank (N ≈ 500,000).

Aurora prognostic biomarker discovery: highlighted biomarker signals rising above screened candidates, heart failure cohort

From risk to prognosis

Earlier Hurdle work modelled incident cardiovascular outcomes across a broad population and found cardiovascular death to be a particularly informative endpoint. The prognostic pipeline keeps that same CV-death endpoint, so results stay comparable, but shifts the question from "who will develop CV disease?" to "among people who already have heart failure, who will die of a cardiovascular cause?"

	Risk biomarkers	Prognostic biomarkers
Population	Broad cohort by condition codes	Patients with I50.* heart failure
Cases	First ICD outcome after recruitment	Death from any CV cause (I00–I99)
Question	Who will get the disease?	Who will die from the disease?

Risk biomarkers vs prognostic biomarkers — how the question changes.

Cohort and approach

Inclusion: participants with an I50.* heart failure diagnosis (primary run requires HF at any time, with no diagnosis-based exclusions to preserve statistical power).
Outcome: cases are HF patients whose cause of death is cardiovascular (ICD I00–I99); controls are HF patients who survived, were censored, or died of a non-CV cause.
Model: penalised Cox time-to-event model across stacked data modalities, from routine clinical data through to proteomics and polygenic risk scores.
Engineering: refactored per-period inclusion/exclusion with staged filtering and a can_use_in_cohort audit flag; stricter variants are one config edit away for sensitivity analysis.

Result 1: discrimination improves with each data layer

Bar chart: C-index for 15-year CV-death prognosis in HF patients rises from 0.60 (age, sex baseline) to 0.76 for clinical + proteomics + PRS — C-index for 15-year CV-death prognosis in HF patients, by data modality (UK Biobank).

Adding clinical labs and proteomics lifts the C-index from ~0.60 (baseline demographics) to 0.76 for the best multimodal model. Genetics alone adds little; the clinical + proteomic stack carries most of the gain, consistent with Hurdle's clinical-enrichment experiments on CV death. The cohort comprised 2,870 cardiovascular deaths and 13,506 sampled controls (~1:4.7); 73.2% of cases were male.

Typology	C-index	ROC-AUC	Events / test n
Baseline + clinical + proteomics + PRS	0.758	0.743	95 / 475
Baseline + proteomics	0.744	0.716	96 / 483
Baseline + clinical	0.672	0.657	563 / 3,260
Baseline + standard PRS	0.611	0.599	548 / 3,146
Baseline (age, sex)	0.599	0.588	568 / 3,276

Table 1 — C-index for 15-year CV-death prognosis in HF patients, by data modality (UK Biobank).

Result 2: multi-protein signatures beat a single marker

Bar chart: hazard ratio for CV death is 1.77 for NT-proBNP alone vs 3.50 for a multi-protein Cox model — Per-SD hazard ratio for CV death, top vs bottom tertile (95% CI). UK Biobank HF cohort.

NT-proBNP is the canonical prognostic marker for heart failure, but on its own it gives only modest separation (HR 1.77 per SD). A multi-protein Cox signature roughly doubles the hazard ratio to 3.50 and widens 15-year survival separation between high- and low-risk tertiles from ~22 to ~46 percentage points. Prognosis is multi-dimensional; Hurdle's panel-level discovery captures signal that a single assay misses.

Discovered biomarkers tell a coherent clinical story

Marker	Direction	Why it makes biological sense
NT-proBNP	↑ risk	Canonical measure of cardiac dysfunction and congestion.
Renin	↑ risk	RAAS activation; reduced renal perfusion worsens HF.
SPON1	↑ risk	ECM protein; marker of advanced fibrosis and remodelling.
ANGPTL4 / EDN1	↑ risk	Metabolic stress and vasoconstriction in failing hearts.
GDF-15 / urate	↑ risk	Systemic stress and cardiorenal-metabolic burden.
HPGDS / diastolic BP	↓ risk	Higher levels track with better prognosis.

Table 2 — discovered biomarkers and their clinical rationale.

These are cardiovascular- and HF-specific signals (natriuretic peptides, RAAS, fibrosis and remodelling markers), not arbitrary noise. That coherence, together with C-index 0.76, makes the discovery both statistically useful and clinically interpretable.

Why it matters and where it goes next

The prognostic pipeline and per-period inclusion/exclusion filtering work together on a real UK Biobank cohort, turning the same Aurora engine that powers risk discovery into a prognostic-biomarker discovery platform. The framework generalises to any disease: a new cohort is defined by editing cohort JSON and enabling prognostic mode in the pipeline config, supporting trial enrichment, patient stratification, and companion-diagnostic development across therapeutic areas.