Table 2.

Procedures of predictive tool interventions in the 11 randomized controlled trials involving interventions evaluating deep learning tools.

Reference	Conditions	Sample size	Tools for intervention	Control	Algorithms	Tool function	Tool input	Tool output	How the output being used in clinical settings	Trial primary outcomes	Gold standard	Trial findings
	Upper gastrointestinal lesions	437	Routine EGD examination stratified by three types with the assistance of ENDOANGEL AI system	Routine EGD examination stratified by three types without AI	DCNN (VGG-16)	Assistive diagnosis	EGD images	A virtual stomach model monitoring blind spots; timing; scoring and grading	Experts referenced AI output to make EGD examination and monitor blind spots.	Mean blind spot rate	Experts	Positive
	Childhood cataracts	700	CC-cruiser web diagnosis platform	Regular ophthalmic diagnosis	DCNN (ImageNet)	Assistive diagnosis	Ocular images from slit-lamp photography	Diagnosis outcome; comprehensive evaluation; treatment recommendation	AI made diagnosis independently, and its results would be comparted with experts and not impact clinical decision making.	Accuracy of diagnosis	Experts	Negative
	Colorectal cancer	659	Routine colonoscopies with the assistance of an AI automatic quality control system	Routine colonoscopies	DCNN (AlexNet, ZFNet, YOLO V2)	Assistive diagnosis	Colonoscopy images	Location of colorectal polyps; timing; reminding retest and clean	Endoscopists referenced AI output to make endoscopic examination and report of polyps and adenomas.	Adenoma detection rate	Pathology	Positive
	Colorectal cancer	1058	Routine colonoscopies with the assistance of an automatic polyp detection system	Routine colonoscopies	Deep learning architecture	Assistive diagnosis	Colonoscopy images	Location of polyps; alarming	Endoscopists were required to check every polyp location detected by the system and report of polyps and adenomas.	Adenoma detection rate	Pathology	Positive
	Upper gastrointestinal lesions	303	Routine EGD examination with the assistance of WISENSE AI system	Routine EGD examination	DCNN (VGG-16 and DenseNet)	Assistive diagnosis	EGD images	A virtual stomach model monitoring blind spots; timing; scoring and grading; extracting frames with the highest confidence	Experts referenced AI output to make EGD examination and monitor blind spots.	Mean blind spot rate	Experts	Positive
	Colorectal cancer	704	ENDOANGEL-assisted routine colonoscopy	Routine colonoscopy	DCNN and perceptual hash algorithms (VGG-16)	Assistive diagnosis	Colonoscopy images	Timing; safe, alarm, and dangerous ranges of withdrawal speed for real-time monitoring; slipping warning	Operating endoscopists referenced AI output to make endoscopic examination and report of polyps and adenomas.	Adenoma detection rate	Pathology	Positive
Liu (2020)42	Colorectal cancer	1026	Routine colonoscopy with CADe assistance	Routine colonoscopy	DCNN-3D	Assistive diagnosis	Colonoscopy images	The probability of polyps in each frame; lesions alarming	Endoscopists focused mainly on the main monitor during the examination process, and a voice alarm prompted them to view the system monitor to check the location of each polyp detected by the system.	Detection rate of polyps and adenomas	Pathology	Positive
	Colorectal cancer	157	AI-assisted colonoscopy	Traditional colonoscopy	CNN (YOLO)	Assistive diagnosis	Colonoscopy images	Location of polyps	Endoscopists referenced AI output to make endoscopic examination and report of polyps.	Polyp detection rate	Not reported	Positive
Repici (2020)62	Colorectal cancer	685	High-definition colonoscopies with the AI-based CADe system	Routine colonoscopy	CNN	Assistive diagnosis	Colonoscopy images	Location of polys	Endoscopists referenced AI output to make endoscopic examination and report of polyps and adenomas.	Adenoma detection rate	Pathology	Positive
Wang (2020)61	Colorectal cancer	962	White light colonoscopy with assistance from the CADe system	White light colonoscopy with assistance from a sham system	Deep learning architecture	Assistive diagnosis	Colonoscopy images	Location of polyps; alarming	Endoscopists were required to check every polyp location detected by the system and report of polyps and adenomas.	Adenoma detection rate	Pathology	Positive
Blomberg (2021)58	Out-of-hospital cardiac arrest (OHCA)	5242	Normal protocols with alert	Normal protocols without alert	Speech recognition using deep neural networks	Assistive diagnosis	Emergency calls	OHCA Alert	Dispatchers in the intervention group were alerted when the machine learning model identified out-of-hospital cardiac arrest.	The rate of dispatcher recognition of subsequently confirmed OHCA	Danish Cardiac Arrest Registry	Negative

AI artificial intelligence, DL tools using deep learning algorithms, ML tools using machine learning algorithms, CNN convolutional neural networks, DCNN deep convolutional neural networks, CADe computer-aided detection, EGD esophagogastroduodenoscopy, OHCA out-of-hospital cardiac arrest.