Deep learning in negative small-bowel capsule endoscopy improves small-bowel lesion detection and diagnostic yield.

OBJECTIVES: Although several studies have shown the usefulness of artificial intelligence to identify abnormalities in small-bowel capsule endoscopy (SBCE) images, few studies have proven its actual clinical usefulness. Thus, the aim of this study was to examine whether meaningful findings could be obtained when negative SBCE videos were reanalyzed with a deep convolutional neural network (CNN) model. METHODS: Clinical data of patients who received SBCE for suspected small-bowel bleeding at two academic hospitals between February 2018 and July 2020 were retrospectively collected. All SBCE videos read as negative were reanalyzed with the CNN algorithm developed in our previous study. Meaningful findings such as angioectasias and ulcers were finally decided after reviewing CNN-selected images by two gastroenterologists. RESULTS: Among 202 SBCE videos, 103 (51.0%) were read as negative by humans. Meaningful findings were detected in 63 (61.2%) of these 103 videos after reanalyzing them with the CNN model. There were 79 red spots or angioectasias in 40 videos and 66 erosions or ulcers in 35 videos. After reanalysis, the diagnosis was changed for 10 (10.3%) patients who had initially negative SBCE results. During a mean follow-up of 16.5 months, rebleeding occurred in 19 (18.4%) patients. The rebleeding rate was 23.6% (13/55) for patients with meaningful findings and 16.1% (5/31) for patients without meaningful findings (P = 0.411). CONCLUSION: Our CNN algorithm detected meaningful findings in negative SBCE videos that were missed by humans. The use of deep CNN for SBCE image reading is expected to compensate for human error.

Post-stroke respiratory complications using machine learning with voice features from mobile devices.

Abnormal voice may identify those at risk of post-stroke aspiration. This study was aimed to determine whether machine learning algorithms with voice recorded via a mobile device can accurately classify those with dysphagia at risk of tube feeding and post-stroke aspiration pneumonia and be used as digital biomarkers. Voice samples from patients referred for swallowing disturbance in a university-affiliated hospital were collected prospectively using a mobile device. Subjects that required tube feeding were further classified to high risk of respiratory complication, based on the voluntary cough strength and abnormal chest x-ray images. A total of 449 samples were obtained, with 234 requiring tube feeding and 113 showing high risk of respiratory complications. The eXtreme gradient boosting multimodal models that included abnormal acoustic features and clinical variables showed high sensitivity levels of 88.7% (95% CI 82.6-94.7) and 84.5% (95% CI 76.9-92.1) in the classification of those at risk of tube feeding and at high risk of respiratory complications; respectively. In both cases, voice features proved to be the strongest contributing factors in these models. Voice features may be considered as viable digital biomarkers in those at risk of respiratory complications related to post-stroke dysphagia.