Skip to main content

Table 3 Feature importance, macro area under receiver operating characteristic curves (macro-AUC), and optimal thresholds of the synovial features in distinguishing OA and RA patients

From: Machine learning identification of thresholds to discriminate osteoarthritis and rheumatoid arthritis synovial inflammation

Feature

Feature importancea

macro-AUC

Optimal threshold

OA vs RAb

Mast cells

0.34

0.80

Present vs none

Automated cell density

0.25

0.88

<3400c cells/mm2

Fibrosis

0.11

0.84

Focal and widespread vs none

Lining hyperplasia

0.10

0.78

Normal lining or 2–3 cells thick vs >3–4 cells thick or > 4 cells thick

Fibrin

0.05

0.68

None vs present

Sub-lining giant cells

0.05

0.57

None vs present

Lymphocytic inflammation

0.04

0.69

None and mild (0–1 perivascular aggregates per low power field) vs marked (both perivascular and widespread interstitial aggregates) and band-like

Neutrophils

0.02

0.60

None vs present

Detritus

0.01

0.64

Absent vs present (small or large particles)

Plasma cells

0.01

0.66

<50% plasma cells

Binucleate plasma cells

0.01

0.60

None vs present

Synovial giant cells

0.01

0.58

None vs present

Germinal centers

0.01

0.51

None vs present

Mucoid change

0.00

0.50

No optimal threshold

Russell bodies

0.00

0.56

None vs present

  1. macro-AUC macro area under the receiver operating curve
  2. aFeature importance scores represent scores for the supervised machine learning model including all fourteen pathology scores and the computer vision-generated cell density
  3. bSee the Appendix for a full list of categorical variables
  4. cComputer vision-quantified cell density measured in mean cells per mm2 of tissue