keys/practice/README.md

Practice answer keys

One file per practice hospital: the flaws we put in (kind, severity, a plain description, and the evidence a correct find cites), plus reference measures comparing each hospital with hospital_a, the clean reference.

For each measure: value is what this hospital shows; chance_p95 and chance_p99 are what two clean samples of the same size show by chance 95% and 99% of the time; verdict is our call (within chance, borderline, weak, strong). A value above chance_p99 is unlikely to be chance.

Patients shared across practice hospitals under different ids: hospital_b & hospital_e: 2

Measures

• aaa_prev_diff: AAA prevalence difference (= EMD on the binary label)
• size_class_hellinger: Hellinger distance over aaa_size_class
• diameter_w1_mm: W1 over max_aortic_diameter_mm, mm
• age_band_tv: TV over age band
• sex_tv: TV over sex
• phase_tv: TV over contrast phase
• slice_class_tv: TV over slice_thickness_class
• vendor_tv: TV over scanner_vendor
• feature_c2st_auc: classifier two-sample AUC, aorta HU and log organ volumes
• embed_c2st_auc: classifier two-sample AUC, embedding PCA-32
• embed_mmd2: unbiased MMD^2, RBF, embedding PCA-32
• cmi_s_y_given_x: I(S;Y|X) in nats, Y = aaa, X = baseline predicted diameter
• label_pred_w1_mm: W1 over (label diameter - baseline predicted diameter), mm
• dice_gap: |mean baseline aorta Dice difference|
• aaa_balacc_gap: |baseline AAA balanced accuracy difference|
• shortcut_max_phi: max phi between one study_description or protocol_name value and aaa (values on 3 or more rows)
• shortcut_u: I(study_description; aaa) / H(aaa), plug-in (uncertainty coefficient, so prevalence alone does not raise it)
• neff_deficit_pid: 1 - n_eff / n, ICC of diameter by released patient_id
• neff_deficit_emb: 1 - n_eff / n, ICC of diameter by near-duplicate embedding cluster
• near_dup_rate: share of rows with a within-hospital embedding neighbour closer than 0.25