Patient Demographics and Radiograph Characteristics

In this study, radiograph sets from 119 patients were analyzed, with a median age of 38 years (interquartile range [Q1-Q3]: 26-56.5 years; range: 16-89 years). Of these patients, 52 (43.7%) were female. The radiographs were evaluated for the presence of abnormalities, with 29 (24.4%) of the sets identified as abnormal. Among these abnormal sets, 10 (37.9%) were classified as critical in nature, and 17 (58.6%) contained a non-nodular pulmonary opacity.

Fig 1: Example of a chest radiograph analyzed by Milvue Suite, demonstrating a true positive finding for a pneumothorax by the AI.

Overall Performance of the AI Algorithm

The AI algorithm succeeded in interpreting all the 119 radiograph sets of the study dataset. It demonstrated a high level of accuracy in its analysis of the chest radiographs. 3 abnormal radiograph sets were missed, and 3 normal sets were incorrectly identified as abnormal, yielding an overall accuracy of 95%, a sensitivity of 88.5%, a specificity of 96.8%, a positive predictive value of 88.5%, and a negative predictive value of 96.8%. Examples of AI results are provided in Figure 1, 2, 3 and 4.

Fig 2: Example of a chest radiograph analyzed by Milvue Suite, demonstrating a true positive finding for a left rib fracture by the AI.

Fig 3: Example of a chest radiograph analyzed by Milvue Suite, demonstrating a false positive finding of a non-nodular pulmonary opacity by the AI.

When focusing specifically on critical findings, the pooled accuracy, Se, Sp, PPV and NPV were of 96.6%, 80%, 98.2%, 80%, and 98.2% respectively. For relevant findings, the pooled performances were slightly lower but still robust, with an accuracy of 93.3%, a sensitivity of 84.2%, a specificity of 95%, a PPV of 76.2%, and an NPV of 96.9%.

Fig 4: Example of a chest radiograph analyzed by Milvue Suite, demonstrating a false negative finding of a non-nodular pulmonary opacity by the AI (see the opacity in the right axillary region).

Performance Per Specific Findings

The AI algorithm's performance was particularly noteworthy in its evaluation of pneumothorax and fractures as it did not miss any cases of pneumothorax (n=2) or fractures (n=2), and there were no false positives reported for these findings, reaching a diagnosis performance of 100%. In addition, none of the 2 pulmonary nodule radiograph sets were missed.