Project · Medical · 2026 · solo
Predicting aortic-wall inflammation in abdominal aneurysms
Four feature pipelines for the HISA inflammation score on 369 patients. Charts plateau near chance; slide embeddings reach a test AUC of 0.717.
Problem
HISA grades how inflamed an aneurysmal aortic wall is, and it is only available after surgery, from tissue. Kolbinger et al. predict it from whole-slide images with attention-based MIL (AUC 0.77–0.79). My question: how much of that signal survives in simpler, tabular representations?
Method
Four pipelines, compared by 5-fold cross-validated AUC with a single use of the held-out test set: pre-operative clinical variables; clinical plus structural histology (with SelectKBest and RFECV); all variables including inflammation grades, kept on purpose as a leakage ceiling; and CTransPath H&E embeddings mean-pooled per patient, then PCA(150).
Results
Pre-operative and structural variables carry almost no signal; mean-pooled slide embeddings recover much of it, still short of the paper’s attention-based model. Validation is internal only, so these AUCs are likely optimistic compared with the paper’s external cohorts.
Measurements
| Setting | Value | Note |
|---|---|---|
| Clinical only (Naive Bayes) | 0.614 | recall (high) 0.11 |
| Clinical + structural (logistic reg.) | 0.591 | |
| H&E mean-pooled CTransPath (logistic reg.) | 0.717 | CV AUC 0.820 ± 0.054 |
| Combined incl. inflammation grades (random forest) | 0.961 | leakage: upper bound only |
| Reference: attention-MIL on whole slides | 0.77–0.79 | Kolbinger et al., external validation |
Held-out test AUC for HISA (binary), best model per pipeline. Model selection on 5-fold CV AUC; test n = 71.
Stack
Python · scikit-learn · pandas · NumPy · h5py · Matplotlib