One Configurable Engine for Risk and Prognostic Biomarkers

Two questions, one engine

Risk discovery asks who will develop a disease: cases are the first matching ICD outcome after recruitment. Prognostic discovery asks who will die from a disease they already have: cases are assigned from cause of death. Both run on the same filtering engine and the same time-to-event model, so results stay directly comparable.

Configurable, reproducible, auditable cohorts

• Per-period filtering: inclusion and exclusion are evaluated against named time windows, so a comorbidity can be required before recruitment yet excluded before the outcome.
• Audit by design: the can_use_in_cohort flag records why each participant is kept or dropped, giving transparent filtering statistics for every run.
• Sensitivity built in: stricter or looser cohort variants (for example a full comorbidity-exclusion block) are one config edit away, with no code changes.

Supported period keys: before_recruitment, all_period, all_period_before_outcome, and after_recruitment_before_outcome (where outcome_end = min(censor_date, date_of_death)).

Proof point: validated on a real cohort

On a heart failure to cardiovascular-death cohort (2,870 CV deaths, 13,506 controls), the engine reached a best multimodal C-index of 0.76, with a multi-protein hazard ratio of 3.50 per SD and clinically coherent top features (NT-proBNP, renin, fibrosis and remodelling proteins). The framework is statistically useful and clinically interpretable on the same run.

Generalises to any disease

Pointing Aurora at a new indication means editing the cohort JSON and enabling prognostic mode in the pipeline config. The same engine then supports trial enrichment, patient stratification, and companion-diagnostic development across therapeutic areas, with auditable cohorts and comparable risk and prognostic readouts throughout.