A deep learning-based system capable of detecting pneumothorax via electrocardiogram.

PURPOSE

To determine if an electrocardiogram-based artificial intelligence system can identify pneumothorax prior to radiological examination.

METHODS

This is a single-center, retrospective, electrocardiogram-based artificial intelligence (AI) system study that included 107 ECGs from 98 pneumothorax patients. Seven patients received needle decompression due to tension pneumothorax, and the others received thoracostomy due to instability (respiratory rate ≥ 24 breaths/min; heart rate, < 60 beats/min or > 120 beats/min; hypotension; room air O saturation, < 90%; and patient could not speak in whole sentences between breaths). Traumatic pneumothorax and bilateral pneumothorax were excluded. The ECGs of 132,127 patients presenting to the emergency department without pneumothorax were used as the control group. The development cohort included approximately 80% of the ECGs for training the deep learning model (DLM), and the other 20% of ECGs were used to validate the performance. A human-machine competition involving three physicians was conducted to assess the model performance.

RESULTS

The areas under the receiver operating characteristic (ROC) curves (AUCs) of the DLM in the validation cohort and competition set were 0.947 and 0.957, respectively. The sensitivity and specificity of our DLM were 94.7% and 88.1% in the validation cohort, respectively, which were significantly higher than those of all physicians. Our DLM could also recognize the location of pneumothorax with 100% accuracy. Lead-specific analysis showed that lead I ECG made a major contribution, achieving an AUC of 0.930 (94.7% sensitivity, 86.0% specificity). The inclusion of the patient characteristics allowed our AI system to achieve an AUC of 0.994.

CONCLUSION

The present AI system may assist the medical system in the early identification of pneumothorax through 12-lead ECG, and it performs as well with lead I ECG alone as with 12-lead ECG.