Table 2 Multivariable logistic regression of radiological findings and proxy follow-up within 90 days. Results of a logistic regression model including all 14 radiological findings simultaneously, adjusted for sex. Odds ratios (OR) with 95% confidence intervals (CI) and two-sided P-values are reported. Variance inflation factors (VIFs) ranged from 1.00 to 1.08, indicating limited collinearity. Statistically significant P-values (P < 0.05) are shown in bold. “No finding” served as the reference category and was not included as a predictor.

From: Exploratory associations between radiographic findings and metadata-derived proxies of 90-day follow-up in 112,120 ChestX-ray14 radiographs

Finding

Male N = 63,340 (n)

Female N = 48,780 (n)

Odds Ratio

95% CI (Lower)

95% CI (upper)

P-value

Edema

1204

1099

10.6

8.5

13.2

< 0.001

Pneumothorax

2717

2585

7.6

6.7

8.6

< 0.001

Effusion

7435

5882

4.0

3.8

4.3

< 0.001

Consolidation

2666

2001

3.9

3.5

4.3

< 0.001

Emphysema

1610

906

3.3

2.9

3.7

< 0.001

Pneumonia

838

593

2.7

2.3

3.2

< 0.001

Atelectasis

6906

4653

2.3

2.2

2.5

< 0.001

Infiltration

11,427

8467

1.9

1.8

2.0

< 0.001

Mass

3529

2253

1.3

1.3

1.4

< 0.001

Pleural thickening

2042

1343

1.3

1.2

1.4

< 0.001

Nodule

3685

2646

1.1

1.0

1.1

0.0357

Cardiomegaly

1307

1469

1.0

0.9

1.1

0.9304

Fibrosis

915

771

0.7

0.7

0.8

< 0.001

Hernia

96

131

0.7

0.5

0.9

0.0037

No finding

33,922

26,439

̶

̶

̶

̶

  1. Follow-up was defined using metadata as a proxy and should be interpreted cautiously. Radiological labels were NLP-derived and subject to variable accuracy (higher for pneumothorax, lower for infiltration [Wang et al.4]). “No Finding” served as the reference category and was not estimated. Results are exploratory and not definitive evidence of clinical practice.
Back to article page