AI for Gastric POCUS ( Point-of-care Ultrasound)

Gastric (stomach) Point-of-care ultrasound (POCUS) is an ultrasound examination done at bedside to assess the stomach. It is a validated non-invasive way to find out what is the content in the stomach and its volume. Gastric POCUS is increasingly used before surgery to determine the risk of gastric contents going into the lungs (possibly causing a lung infection and breathing problems) and guide anesthetic management whenever the doctors are not certain about the stomach content based on clinical information.

Gastric POCUS is a relatively new skill for anesthesiologists. While, obtaining the required images is relatively straightforward, the interpretation of such images, however, requires advanced training. Preliminary data have suggested that Artificial Intelligence (AI)-based programs and devices can help in image capturing and its interpretation for other ultrasound applications. This study will be the first to the researcher's knowledge to develop an AI algorithm to enhance anesthesiologists' ability to recognize a full stomach using gastric POCUS. The goal of this observational study is to train and test an AI (Artificial Intelligence)-based program to assist anesthesiologists in the interpretation of stomach ultrasound images and differentiate a "full" from an "empty" stomach.

This is an observational prospective cohort study that follows the CONSORT (Consolidated Standards of Reporting Trials)-AI extension reporting guidelines.

The researchers expect to enroll 30 healthy volunteers for the study.

Following a period of fasting for solids for at least 8 hours and clear fluids for at least 2 hours from the time of study visit. An anesthesiologist or sonographer with a minimum previous experience of 50 gastric ultrasound examinations will perform a standardized gastric ultrasound exam.

A baseline ultrasound examination will be conducted first with the participant lying on their back with the head elevated at 30 degrees (supine position) and then again with the participant lying on their right side (right lateral decubitus position(RLD)).

The same procedure will be repeated twice after ingestion of

1. various volumes of water (100-500) determined at random 2. 1 of 5 different solid or thick fluid foods also determined at random (a banana, an apple, a cup of yogurt, a croissant or a muffin).

Each one of the 30 participants will be randomized to 1 of 5 different volumes of water (100ml, 200ml, 300ml, 400ml, 500ml). Then ultrasound images will be obtained. Subsequently, each participant will also be randomized to 1 of 5 solids (a banana, an apple, a cup of yogurt, a croissant or a muffin) in a 1:1:1:1:1 ratio. A computer-generated list of random numbers for each participant will be created.

The investigators plan to collect 90 10-second clips in total, and each clip can be deconstructed into 10 frozen frames per second, for a total of 100 frozen frames per clip. The investigators expect to generate 9,000 individual images, 80% of which will be used to train the model, 10% to fine-tune and 10% to test the model accuracy. The three de-identified clips from each participant will be normalized and annotated by consultant anesthesiologists to indicate orientation (medial or lateral, cephalad or caudad) and identify relevant structures, as well as the type of content and antral CSA in the right lateral decubitus in case of fluid.

All the collected images will then be fed to an AI to generate computational data.

• To measure the accuracy of the AI model in differentiating a empty from a full stomach [Time frame: Through study completion, an average of 2 years]
• To measure the balanced accuracy of the AI model [Time frame: Through study completion, an average of 2 years]
• To measure the precision of the AI model [Time frame: Through study completion, an average of 2 years]
• To measure the recall of the AI model [Time frame: Through study completion, an average of 2 years]
• To evaluate the model's classification performance across different confidence thresholds. [Time frame: Through study completion, an average of 2 years]
• To evaluate the model's performance in detecting different anatomical structures (e.g., organs, vessels). [Time frame: Through study completion, an average of 2 years]
• To measure the latency and average inference time per image [Time frame: Through study completion, an average of 2 years]