Patient Demographics and Radiograph Characteristics

For each dataset, 300 cases from 300 unique patients were included. All cases could be processed by SmartChest and were thus included in the per-protocol analysis.

Pneumothorax Dataset

The median age was 56 years old (interquartile range [Q1-Q3]: 34.8-71; range: 18-90). 44.3% were female, 37,7% were aged ≥65 years, and 58.2% were acquired as a PA view.

Among the 150 (50%) pneumothorax cases, 52% were classified as small and 1.5% were bilateral. The most frequent concomitant chest anomalies were pleural effusion (31.7%), catheters or tubes (28.0%), and atelectasis (17.7%). Most imaging sites were located in large urban (37.0%) or community (32.0%) areas. The most represented manufacturers were Samsung (36.5%), Canon (15.0%), Siemens (11.3%), Carestream (10.6%), Fujifilm (6.8%), and GE (5.1%), with other manufacturers each representing <5% of cases.

Pleural Effusion Dataset

The median age was 60 years old (interquartile range [Q1-Q3]: 42-73; range: 18-90). 47.3% were female, 41% were aged ≥65 years, and 63.2% were acquired as a PA view.

Of the 150 (50%) pleural effusion cases, 58% were characterized as small, and the effusions were mostly unilateral (42.6% right-sided and 29.5% left-sided). Most imaging sites were located in urban (40.3%) and large urban (34%) areas.

Standalone Performance of the AI Algorithm

To classify the presence or absence of a suspected pneumothorax, the deep learning (DL) software achieved a ROC-AUC of 0.989 (95% CI: 0.978–0.997), with the lower boundary of the 95% confidence interval exceeding the acceptance criterion of 0.95. It yielded 92.7% sensitivity (95% CI: 87.4–96.2) and 97.3% specificity (95% CI: 93.4–99.1).

For pleural effusion, it achieved a ROC-AUC of 0.975 (95% CI: 0.960–0.987) and yielded 93.3% sensitivity (95% CI: 88.1–96.4) and 90.0% specificity (95% CI: 84.1–94.1). A true positive and a false negative results are shown in Figure 1 and 2.

The forest plots in Figures 3 and 4 for pneumothorax and pleural effusion, respectively, provide an overview of the ROC-AUC values for each subgroup and each specific population. The software’s performance remained consistent across all age groups (18–21 years up to ≥65 years), across a variety of imaging systems, and for both frontal view types. For both pleural effusion and pneumothorax, SmartChest’s performance was high and consistent regardless of the lesion’s volume or location.