Table 2:

Linear readout performance on COPDGene dataset. The values are the average of 5-fold cross validation with standard deviations. IFM yields improvements on all phenotype predictions.

Method	logFEV1pp	logFEV1FVC	CLE	CLE 1-off	Para-septal	Para-septal 1-off	mMRC	mMRC 1-off
Loss	R-Square		Accuracy (%)
$\mathcal{L}$ (baseline)	0.566±.005	0.661±.005	49.6±0.4	81.8±0.5	55.7±0.3	84.4±0.2	50.4±0.5	72.5±0.3
${\mathcal{L}}_{\epsilon }$, ε = 0.1	0.591±.008	0.681±.008	49.4±0.4	81.9±0.3	55.6±0.3	85.1±0.2	50.3±0.8	72.7±0.4
IFM, ε = 0.1	0.615 ±.005	0.691 ±.006	48.2±0.8	80.6±0.4	55.3±0.4	84.7±0.3	50.4±0.5	72.8±0.2
IFM, ε = 0.2	0.595±.006	0.683±.006	48.5±0.6	80.5±0.6	55.3±0.3	85.1±0.1	49.8±0.8	72.0±0.3
IFM, ε = 0.5	0.607±.006	0.683±.005	49.6±0.4	82.0±0.3	54.9±0.2	84.7±0.2	50.6 ±0.4	73.1 ±0.2
IFM, ε = 1.0	0.583±.005	0.675±.006	50.0 ±0.5	82.9 ±0.4	56.3 ±0.6	85.7 ±0.2	50.3±0.6	71.9±0.3