heartwise/deepcoro_clip_cardiosyntax

DeepCORO-CLIP SYNTAX (v5, 16-segment)

Fully automated estimation of a modified SYNTAX score from a complete multi-view coronary angiography study. Fine-tuned from the DeepCORO-CLIP video–language foundation model.

Under peer review — not for clinical use. The manuscript is under review at EuroIntervention (EIJ-D-26-00383); these results are not yet peer-reviewed. The model is a triage and consistency aid whose output a clinician confirms against the images. It is not an autonomous scorer, has not been prospectively validated, and carries no regulatory clearance.

What it predicts

Six heads on one study-level representation: continuous global, left and right SYNTAX scores, and the corresponding four-class severity (zero / 1–22 / 23–32 / ≥33).

The reference standard is a modified SYNTAX score

Read this before using the output. The score is 16-segment, segment-based and is not interchangeable with a core-laboratory SYNTAX score:

• 16 of the 25 SYNTAX segments are covered. The annotation instrument presents the Ramus/S12 and the first obtuse marginal (OM1/S12a) as separate entries for one anatomical segment; the scorer collapses them, keeping whichever complete profile carries the larger standalone contribution.
• Severe vessel tortuosity is never recorded and scores nothing.
• Of the five total-occlusion sub-modifiers, none is collected. The routine imputes the chronicity point for every occlusion and adds one point per additional occluded segment in a run; blunt stump, bridging collaterals and a side branch at the occlusion are never scored.
• The 1-vs-2-point bifurcation distinction rests on a single side-branch question rather than Medina notation.

All of these are one-directional: the score is biased downward, most in occlusive and multivessel disease. scoring/score_syntax_v5.py is the exact routine.

Performance

Held-out test partitions. The binary operating threshold (16.90) was selected once on the merged validation partition and held fixed — no threshold is tuned on test data.

	CardioSYNTAX (n=369)	MHI held-out (n=127)
AUROC (≥23)	0.97 (95% CI 0.94–0.99)	0.88 (95% CI 0.79–0.95)
Sensitivity	0.86 (0.72–0.97)	0.72 (0.50–0.91)
Specificity	0.93 (0.90–0.96)	0.80 (0.71–0.87)
NPV	0.98 (0.97–1.00)	0.95 (0.90–0.99)
Spearman ρ	0.77 (0.71–0.82)	0.80 (0.72–0.86)
ICC(2,1)	0.83 (0.78–0.87)	0.76 (0.67–0.82)
MAE (points)	3.36	4.98

High NPV at modest PPV: this is a rule-out step, not a confirmatory test.

Against physicians, on the 171 studies with three complete ratings: AI-SYNTAX accuracy 0.74 (95% CI 0.67–0.80), κ 0.71 (0.59–0.79); physicians' ordinal-median consensus 0.70 (0.63–0.76), κ 0.69 (0.57–0.78). Comparable to, not better than, the physician consensus — the intervals overlap substantially.

Limitations that bear on deployment

• No fully external validation. Both cohorts contributed training data; the MHI result is held-out internal performance at a second institution.
• Proportional bias persists. The model scores progressively lower than the reference as complexity rises (slope −0.168, p<0.001 Dataset A; −0.164, p=0.009 Dataset B).
• No segment-level attribution. The model does not localise the points it assigns; attention weights indicate which projections it used, not which structures.
• Untested in grafted anatomy — prior-CABG studies were excluded.
• Weakest in the moderate stratum, which is also where expert readers disagree most.

Training

Run 20260822-164504, best epoch 12 of 30 on validation loss. Merged CardioSYNTAX + MHI development set: 2,695 studies, 30,036 decodable videos, split 1,859 / 340 / 496 at patient level with a leakage audit. Video encoder 91.8% frozen. v5_20260822-164504/config.yaml has the rest.

v6 — CardioSYNTAX v3 (candidate, not yet deployed)

Run 20260929-203527, best epoch 19 of 30 on validation loss. Same data, split (1,859 / 340 / 496), reference standard, architecture, loss (Huber + CE) and hyperparameters as v5; trained with batch 1 × 2 gradient-accumulation steps (effective batch 2) and the epoch-9 DeepCORO-CLIP encoder checkpoint. The checkpoint has the same keys and shapes as v5 and loads in the existing service unchanged.

Held-out test, zero-clamped total score, v5 re-run on the identical split (95% bootstrap CIs):

	v5 CardioSYNTAX	v6 CardioSYNTAX	v5 MHI	v6 MHI
MAE (points)	3.23 (2.77–3.74)	3.21 (2.77–3.64)	4.90 (4.11–5.72)	4.47 (3.67–5.30)
Bias	−0.85	−0.35	−1.19	−0.57
ICC(2,1)	0.83	0.86	0.75	0.79
AUROC (≥23)	0.97	0.97	0.88	0.89

By reference severity (both cohorts, n=496), MAE (bias):

	0 (n=218)	1–22 (n=225)	23–32 (n=31)	≥33 (n=22)
v5	1.1 (+1.1)	4.5 (−0.7)	8.3 (−7.1)	14.6 (−14.6)
v6	1.0 (+1.0)	4.8 (−0.2)	7.4 (−4.6)	10.9 (−10.2)

Paired v6 − v5: overall MAE −0.13 (95% CI −0.34 to 0.08, not significant); MAE in SYNTAX ≥23 (n=53) −2.08 (−2.93 to −1.18). v6 reduces, but does not remove, the proportional underestimation in complex anatomy. Single training seed; confirm with additional seeds before replacing v5.

Loss variants tested alongside v6 (linear target scaling; SynSeq-style inverse-frequency sample weighting, arXiv 2609.27696) did not improve overall accuracy: weighting cut the ≥33 bias to about −2 but overestimated zero/low scores by 2–6 points (overall MAE 5.6–6.2).

Files

Path	What
v5_20260822-164504/models/best_model_epoch_12.pt	Current weights (v5, 16-segment)
v5_20260822-164504/config.yaml	Training configuration
v6_20260929-203527/models/best_model_epoch_19.pt	Candidate weights (v6, CardioSYNTAX v3) — not yet deployed
v6_20260929-203527/config.yaml	v6 training configuration
v6_20260929-203527/test_results.md	v6 vs v5 held-out test comparison (aggregate metrics only)
scoring/score_syntax_v5.py	Reference scorer (16-segment, S12 collapse)
scoring/score_syntax_v4.py	17-segment predecessor, imported by v5
rxt2fz27_20250711-171619/	Superseded. The original CardioSYNTAX-only model (epoch 17), kept because earlier work cites it. Trained on CardioSYNTAX alone against the 17-segment reference standard; do not use it for new work.

Version history

Version	Run	Reference standard	Development set
v6 (candidate)	20260929-203527, epoch 19	Modified SYNTAX, 16-segment	Merged CardioSYNTAX + MHI, 2,695 studies (same split as v5)
v5 (current)	20260822-164504, epoch 12	Modified SYNTAX, 16-segment	Merged CardioSYNTAX + MHI, 2,695 studies
v1 (superseded)	rxt2fz27, epoch 17	17-segment predecessor	CardioSYNTAX only

v5 supersedes v1 on every reported metric and is the only version the manuscript reports.

Citation

Manuscript under review. Cite the DeepCORO-CLIP foundation model (arXiv:2603.17675) until publication.