Datasets:

ClarusC64
/

ffr-anatomy-prediction-discordance-detection-v0.1

id string	case_context string	vessel_tortuosity float64	calcification_burden float64	lesion_length_mm int64	segmentation_confidence float64	image_artifact_score float64	ai_ffr_prediction float64	ai_ffr_run_variance float64	model_disagreement float64	invasive_ffr_ground_truth float64	abs_error float64	discordance_flag int64	discordance_type string	subgroup_risk_label string	reliability_drop_score float64	notes string	constraints string	gold_checklist string
APDD-001	clean vessel	0.2	0.1	12	0.95	0.06	0.84	0.01	0.02	0.83	0.01	0	none	none	0.08	aligned baseline	<=280 words	flag+type+subgroup+risk
APDD-002	moderate calcification stable	0.38	0.45	20	0.86	0.1	0.78	0.02	0.05	0.75	0.03	0	calcification-tolerant	moderate-calcified	0.22	still within band	<=280 words	flag+type+subgroup+risk
APDD-003	heavy calcification silent drift	0.4	0.75	26	0.78	0.14	0.76	0.03	0.06	0.67	0.09	1	calcification-driven overestimate	heavily-calcified	0.62	AI looks plausible but wrong	<=280 words	flag+type+subgroup+risk
APDD-004	high tortuosity segmentation wobble	0.7	0.2	22	0.72	0.12	0.74	0.05	0.1	0.66	0.08	1	tortuosity-driven instability	high-tortuosity	0.58	variance spikes	<=280 words	flag+type+subgroup+risk
APDD-005	mixed anatomy + artifacts	0.55	0.65	28	0.68	0.26	0.75	0.06	0.14	0.63	0.12	1	artifact-amplified discordance	calcified+artifact	0.74	artifact magnifies error	<=280 words	flag+type+subgroup+risk
APDD-006	long lesion low confidence	0.5	0.4	40	0.62	0.18	0.73	0.06	0.12	0.64	0.09	1	low-confidence under-call risk	long-lesion+low-conf	0.66	confidence collapse	<=280 words	flag+type+subgroup+risk

Goal

Detect discordance
between coronary anatomy complexity
and AI-derived FFR accuracy
against invasive FFR ground truth.

This targets silent degradation
in specific patient subgroups.

Inputs

Required outputs

Discordance types

Examples:

Why it matters

AI can look stable
while failing on:

Validators need to know
where accuracy drops
before it shows up
as a headline metric failure.

Evaluation

The scorer checks that the response includes: