Study design
| Element | Detail | | --------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------- | | Design | Retrospective, 3,115 patients scanned by NCCT for acute ischemic symptoms | | Development set | 2,198 (1,108 AIS / 1,090 normal) — ~balanced | | Validation set | 917 (367 AIS / 550 normal) — "independent," ~40% AIS | | Image branch | DenseNet121 + BiLSTM (reads the axial slice stack as a sequence) | | Clinical branch | FT-Transformer embedding 11 clinical variables | | Fusion | Multimodal attention module that adaptively weights image vs. clinical features (Fig 1C compares 5 fusion strategies; 1D details the chosen one) | | Metrics | Accuracy, sensitivity, specificity from a confusion matrix |
Results — the three models compared
Reported fusion numbers are exact; the single-modality columns are my back-calculation from the reported deltas (confidence: high — simple arithmetic).
| Model | Accuracy | Sensitivity | Specificity | | ----------------------- | --------- | ----------- | ----------- | | Fusion (reported) | 89.7% | 91.6% | 88.5% | | Image-only (derived) | ~73.7% | ~75.7% | ~72.4% | | Metadata-only (derived) | ~88.2% | ~86.1% | ~89.6% |
SHAP top clinical predictors: NIHSS 0.280, hypertension 0.170, heart disease 0.150, A-fib 0.150, alcohol 0.094, smoking 0.082, diabetes 0.071.
Key insights
- The "synergy" is real against images, marginal against metadata. Fusion beats image-only by ~16 points across the board, but beats metadata-only by just +1.5% accuracy and +5.5% sensitivity — and is actually 1.1% worse on specificity. Most of the model's predictive power comes from the clinical variables, not the CT.
- What the images actually add is sensitivity. The image branch pushes the model to catch more true strokes (+5.5% sens over metadata alone) at a small specificity cost. For a triage tool meant to avoid missing strokes, that's a clinically sensible trade — but it's a modest gain, not a transformation.
- NIHSS dominates — and that's the main scientific caveat. NIHSS (SHAP 0.280) is nearly double the next feature. NIHSS is itself a clinician-administered stroke severity scale, so it's near-circular with the diagnosis: a high NIHSS already means stroke is clinically suspected. A reviewer will likely ask how much the model adds beyond what the ER physician already knows from the NIHSS. (Confidence: high that this will be raised; it's a standard critique.)
- NCCT genuinely is weak alone (~74% image-only accuracy), which is expected in the hyperacute window — this is the gap the paper is built to close, and it justifies the multimodal approach.
Limitations a poster reviewer may probe
- No AUC/ROC, no confidence intervals, no significance tests — only point estimates at one threshold. The 1.5% fusion-vs-metadata gap may not be statistically meaningful.
- Validation appears to be an internal split, not an external site — generalizability across scanners/hospitals is untested.
- AIS ground-truth definition not stated (presumably follow-up DWI-MRI) — worth having ready.
- "Spatial-temporal" is slightly loose wording: axial slices are spatial, not temporal; the BiLSTM models slice-order continuity.