Single nodule inputs
Diameter and volume are linked as a sphere. Brock uses diameter. The growth component uses volume.
Single nodule output
CRS action tier
-
Position in cohort feature space
Run a cohort simulation to populate malignant and benign centroids.
Why this tier
Cohort parameters
Advanced cohort settings
These defaults are plausible synthetic Nordic-mixed-cohort priors for sandbox discussion, not empirical Nordic registry data. Type probabilities are normalised internally if they do not sum to 1.
Edit class-conditional distributions
| Feature | Malignant | Benign |
|---|---|---|
| log10 volume mu | ||
| log10 volume SD | ||
| Type probabilities, solid / part-solid / GGO | ||
| Spiculation probability | ||
| Upper lobe probability | ||
| Nodule count Poisson lambda plus 1 | ||
| Age mu / SD | ||
| Female probability | ||
| Emphysema probability | ||
| Family history probability | ||
| Pack-years Gamma shape / scale |
ROC and AUC for CRS
Euclidean grey-zone summary
Multicentre centroid decomposition
Each patient is described by its Gaussian-kernel proximity to four clinical centroids (malignant, ordinary benign, inflammatory benign, no prior scan), giving a soft local-centroid mixture score. The aim is to make ambiguous cases more interpretable, not to create a validated classifier.
Boundary robustness under measurement noise
Each patient is perturbed using the configured volume CV. Prior and current volumes are both perturbed when a prior scan exists. Pathway decisions are then reclassified to estimate decision fragility.
Feature-space PCA scatter
Feature vector: z(log10 current volume), z(G(VDT)), and unnormalised Brock_P. Triangles mark no-prior-scan patients.
Distance-ratio histogram
r = distance to malignant centroid / (distance to malignant + distance to benign). r near 0.5 is the grey zone.
Built-in pathway controls
Custom pathway editor
No free-text code is executed. The rule is an OR-combination of validated numeric fields.
Pathway performance table
Expected total harm
Harm weights are explicit and subjective. They are intended for sensitivity discussion, not clinical valuation.
Threshold sensitivity sweep
Brock and CRS sweep their score thresholds. GrowCAT sweeps the upper VDT bound with the lower bound fixed. Hybrid sweeps its low Brock threshold.
Gap population — who is reclassified between two pathways
The gap population is the set of patients whose referral eligibility changes when one pathway replaces another. Two pathways can have similar AUC or sensitivity while moving different patients in or out of eligibility; this panel describes who those patients are.
Harm-weight sensitivity — is the lowest-harm pathway stable?
Because the harm weights are subjective, each weight is independently halved and doubled while the others stay fixed. If the lowest-total-harm pathway changes, the harm-based ranking depends on that weight.
Information axes — what does an orthogonal test add?
Imaging + clinical risk (CRS) has a discrimination ceiling. This panel asks what a genuinely orthogonal information axis would buy: a blood protein panel (4MP-like) or a deep-learning CT model (Sybil-like). Set each axis's standalone discrimination and how redundant it is with CRS, then see the gain for everyone versus only the contested grey zone. Educational simulation on the synthetic cohort, calibrated to published numbers (4MP AUROC ~0.74; Sybil ~0.86) — not real performance.
Protein panel (4MP-like)
Deep-learning CT (Sybil-like)
Discrimination ladder
Complementarity at the chosen specificity
Among true malignancies: who does each axis catch, and how many does a new axis catch that the others miss?
Guideline takeaway
Model assumptions and limitations
- Brock-Nordic validation gap. The Brock 2013 PanCan score is used because it is a published continuous nodule model. It is not validated for Nordic populations and must not be treated as a Nordic clinical decision rule.
- Synthetic cohort caveat. The cohort generator uses editable, class-conditional assumptions for guideline-planning experiments. It is not a registry emulator.
- Inflammatory-spike model. Very rapid benign growth is explicitly simulated, and G(VDT) is non-monotonic so VDT below about 100 days is not automatically high risk.
- Measurement-noise assumption. Volume uncertainty is modelled as lognormal CV-based noise applied to both prior and current volumes before observed VDT is calculated.
- Harm-weight subjectivity. CT, PET, biopsy, false-positive biopsy, missed malignancy, and indeterminate penalties are explicit weights for comparison only.
- Four-centroid mixture caveat. The local-centroid mixture is an exploratory interpretability aid for ambiguous cases. It is not a validated clinical classifier.