MDT recommendation
|
Accuracy of prediction by
|
---|
Modality
|
Recommended
|
Best ML modela
|
ESMO Guidelines
|
NCCN Guidelines
|
---|
Strategy
|
N (%)
|
Sens/Spec
|
LR+
|
Sens/Spec
|
LR+
|
Pb
|
Sens/Spec
|
LR+
|
Pb
|
---|
Chemotherapy
|
Aggressive
|
582 (60)
|
0.93/0.89
|
8.8
|
0.55
c
/0.78
c
|
2.5
|
<0.01
|
0.97/0.12c
|
1.1
|
<0.01
|
Conservative
|
342 (35)
|
0.86/0.95
|
16.7
|
0.60
c
/0.82
c
|
3.3
|
<0.01
|
0.82/0.71c
|
2.9
|
<0.01
|
Endocrine
|
Aggressive
|
916 (91)
|
0.98/0.85
|
6.5
|
0.98/0.81
|
5.2
|
0.68
|
0.97/0.75
|
3.9
|
0.25
|
Conservative
|
794 (79)
|
0.97/0.65
|
2.8
|
0.99/0.36c
|
1.5
|
0.30
|
0.96/0.50
|
1.9
|
0.37
|
Trastuzumab
|
Aggressive
|
93 (20)
|
0.98/0.99
|
77.9
|
0.97/0.97
|
33.3
|
0.60
|
0.97/0.98
|
45.2
|
0.73
|
Conservative
|
86 (19)
|
0.95/0.99
|
122.9
|
0.97/0.96
|
24.0
|
0.24
|
0.92/0.98
|
37.7
|
0.28
|
- The sensitivity (sens), specificity (spec), and the positive likelihood ratio (LR+) when using the best machine learning models or guideline to predict MDT recommendations
- Note: aThe best models were ripple down rules for the chemotherapy decisions, polynomial SVM for the aggressive endocrine decisions, and ADTree for the remaining groups
- bpairwise comparisons of likelihood ratios using two-sided z-test (i.ebest model vs. guideline)
- cthe best model performed better than the guideline