Hill-grade Knowledge Via Integrated Neural-network for Gastroscopy

Objective:

The primary goal of this study is to compare the accuracy in determining the Hill classification during gastroscopy between an artificial intelligence (AI) based system and physicians performing the examination. Secondary outcomes include evaluation of the per-class accuracy and other statistical measures such as precision, recall and f1 score.

Study Design:

Single center, endoscopist blinded study. The model considered in a previous study achieved a mean accuracy of 88%. All participants initially attended a lecture serving as a refresher regarding the Hill classification. Subsequently, physicians were asked to provide the Hill classification for test images expert annotated images depicting different Hill grades, achieving mean accuracy of 72%. Thus 127 paired measurements are required. Taking patient drop-out into consideration, at least 159 patients need to be recruited. Upon examination of the flap-valve during endoscopy, the physician is required to store an image of the flap-valve during retroflexion, which is part of the standard procedure, based on which they determine the Hill classification. The prediction of the AI model on this image is considered the model output and is considered the model's output. A group of three expert endoscopists determines the Hill classification for each image, based on majority vote, which is treated as the gold standard.

AI setup and limitations:

There are no limitations caused by the AI. The method performs a frame-by-frame analysis of the recording. These images are parsed from the AI based system in order to obtain predictions. The only interactions required with the method is a button press that initiates the examination recording process and a second button press to terminate the recording. This is performed at the beginning and end of the examination respectively. The model used in this study is an updated version of the model reported in a preliminary study, that has been trained with more data together with an auxiliary output for predicting if the Hill classification is relevant to the shown image.

Study population:

All adult patients appointed for gastroscopy that do not match the exclusion criteria will be asked for informed consent. Exclusion criteria include previous surgical interventions or altered anatomy that prevents the proper examination of the flap valve, examinations where the flap-valve is not inspected, and examinations where the expert committee does not produce a majority vote.

Intervention:

The physician performs the examination as usual. Upon inspection of the flap valve, the physician captures an image of the examination, as usual, and gives their assessment of the Hill grade. The output of the model for the same image is considered the model prediction. The physician is blinded to the model's prediction.

Inclusion criteria

• Adult patients (>18 years)
• Scheduled gastroscopy

Exclusion criteria

Examination level

• Previous surgical interventions or altered anatomy that prevents the proper examination of the flap valve
• Flap-valve not inspected

Data Level:

• Image during flap-valve inspection not stored
• Expert committee not resulting in a majority vote