Challenges involved in the application of artificial intelligence in gastroenterology: The race is on!

Table 1 Studies employing artificial intelligence/machine learning tools in the field of gastroenterology

Ref.	Parameters employed/study design	AI classifier	Sample size, control group, validation	Outcomes	Performance1
Prevention
Goshen et al[15]	EHR data/prospective validation	LR, DT, GB	688 flagged patients, NA, NA	High risk of CRC development	19 (7.5%) CRCs were found in 254 colonoscopies
Holt et al[16]	EHR data/case-control study	LR, DT, GB	1893641 patients, 15 controls matched to 1 case, NA	Early detection of CRC	0.536-0.624c-index
Huang et al[17]	EHR data/case-control study	LR, LASSO, SVM, KNN, RF	1035 patients, 3 controls matched to 1 NCGC case, 10-fold CV	Early detection of NCGC	LR: 72.4acc, 79.3sen, 70.4spe. LASSO: 75.1acc, 80sen, 73.6spe. SVM: 75.1acc, 76.3sen, 74.7spe. KNN: 78.1acc, 68.9sen, 80.8spe. RF: 72.2acc, 77.8sen, 70.6spe
Briggs et al[18]	EHR data/case-control study	RF, SVM, LR, NB, EGBDT	40348, 7471 cases/32877 controls), hold-out validation (75:25), 5-fold CV	Early detection of oesophago-gastric cancer	87-89acc, 0.81-0.87c-index
Diagnosis
Manandhar et al[20]	Gut microbiome data (fecal 16S metagenomic data)/case-control study	RF, SVM, DT, ANN	1429 patients, 729 IBD and 700 non-IBD patients, hold-out validation	Diagnosis of IBD	77-84acc, 0.41-0.82sen, 77-84sen, 46-64spe
Papa et al[21]	Gut microbiome data/case-control study	SVM, RF	105 children and young adults, 91 with IBD and 24 controls, 10-fold CV	Diagnosis of IBD in child population	0.83-0.91c-index
Pace et al[22]	Laboratory results, clinicopathological parameters/retrospective cohort study	ANN	159 patients, 103 with gastroesophageal reflux and 56 controls, 20-fold CV	Diagnosis of gastroesophageal reflux disease	78-100acc
Yepes et al[25]	Expression levels of 1046/315 microRNAs in gastric samples/retrospective cohort study	SVM, RF	648 gastric samples, 479 cancer and 169 controls, leave-one-out CV	Classification of non-tumor mucosa vs tumor sample	SVM: 94acc, RF: 89.3acc
Huang et al[26]	Serum expression levels of miR-21-5p, miR-22-3p, and miR-29c-3p/retrospective cohort study	Several	192, 122 cancer samples and 70 controls, leave-one-out CV	Identification of patients with gastric cancer	56-93acc
Alizadeh et al[27]	Serum expression levels of miR-663a, miR-1469, miR-92a-2-5p, miR-125b-3p, and miR-532-5p/retrospective cohort study	ANN	671, NR, 5-fold CV	Identification of patients with pancreatic ductal adenocarcinoma	93acc, 93sen, 92spe
Afshar et al[28]	Serum expression levels of miR-6726-5p, miR-7111-5p, miR-1247-3p, and miR-614/retrospective cohort study	ANN	200, 50 with colorectal cancer and 150 healthy controls, hold-out validation (70:15:15)	Identification of patients with colorectal cancer	100acc, 1c-index, 100sen,spe
Duttagupta et al[29]	Expression levels of 847 microRNAs in the peripheral blood/case-control study	SVM	40, 20 patients with ulcerative colitis and 20 controls, 10-fold CV	Identification of patients with ulcerative colitis	92.3-92.8acc, 87.8-89.5sen, 96.2-96.8spe
Treatment
Morilla et al[30]	Colonic microRNA profiles (9 microRNAs) and five clinical factors/retrospective cohort study	DNN	76 patients, 22 responders and 54 non-responders, hold-out validation (47:29)	Prediction of response to treatment of patients with active severe ulcerative colitis	80-931, 0.80-0.91c-index
Takiyama et al[31]	Esophago-gastro-duodenoscopy imaging/retrospective cohort study	CNN	17080 images, 363 larynx/2142 esophagus/13048 stomach/1528 duodenum, hold-out validation	Anatomical classification among the larynx, esophagus, stomach, and duodenum	0.99-1.00c-index
Rogers et al[35]	Data from baseline impedance, nocturnal baseline impedance, and acid exposure time/retrospective cohort study	DT	335 patients, 210 with gastroesophageal reflux and 115 controls, NR	Prediction of response to treatment with proton pump inhibitors for patients with gastroesophageal reflux disease	0.31-0.938c-index
Takayama et al[36]	Clinicopathological parameters, treatment data/retrospective cohort study	ANN	90 patients, 32 non-responders and 58 remission-effect, hold-out validation (54:36)	Prediction of the need for operation for UC patients treated with cytoapheresis	96sen, 97spe
Das et al[37]	Laboratory results, clinicopathological parameters/retrospective cohort study	ANN	587 patients, 246 patients with major stigmata-emergent endoscopy to 162 patients, hold-out validation (194:193:200)	Prediction of major stigmata of recent hemorrhage	89acc, 89-96sen, 63-89spe
				Prediction of the need for emergent endoscopy	61-81acc, 61-94sen, 48-82spe
Chu et al[38]	Laboratory results, clinicopathological parameters/retrospective cohort study	Several	189 patients, NR, hold-out validation (122:67)	Prediction of the source of GIB	69.7-94.3acc, 0.658-0.999c-index, 90.1-98.0sen, 89-100spe
				Prediction of the need for blood resuscitation	64.7-94.1acc, 0.381-0.993c-index, 90.3-93.9sen, 18.4-95.5spe
				Prediction of the need for emergent endoscopy	62.7-83.3acc, 0.404-0.913c-index, 80.1-89.1sen, 13.8-85.7spe
				Prediction of disposition	58.4-89.7acc, 0.324-0.972c-index, 81.9-92.9sen, 18.4-90.9spe
Loftus et al[39]	Laboratory results, clinicopathological parameters/retrospective cohort study	ANN	147 patients, 41% of patients with severe lower GIB and 13 patients needed surgical intervention, hold-out validation (103:44)	Prediction of severe lower GIB	0.979c-index
				Prediction of the need for surgical intervention	0.954c-index
Ayaru et al[40]	Laboratory results, clinicopathological parameters/retrospective cohort study	GB	300 patients, 88 with severe bleeding, 53 with recurrent bleeding, and 35 requiring intervention, hold-out validation (170:130)	Prediction of severe lower GIB	78-83acc
				Prediction of recurrent bleeding	88-88acc
				Prediction of the need for intervention	88-91acc
Prognosis
Sato et al[41]	Laboratory results, clinicopathological parameters, tumor characteristics/retrospective cohort study	ANN	395 patients, 281 alive at 1-year-89 alive at 5 years, hold-out validation (53:27:20)	1-yr and 5-yr survival of patients with esophageal cancer following surgery	0.883-0.884c-index, 78.1-80.7sen, 84.7-86.5spe
Shung et al[42]	Laboratory results, clinicopathological parameters/retrospective cohort study	GB	2357 patients, 1109 requiring either transfusion or hemostatic intervention, hold-out validation (1958:399)	Stratification of risk of patients with gastrointestinal bleeding	0.91c-index
Zhou et al[43]	Laboratory results, clinicopathological parameters, tumor characteristics/retrospective cohort study	Several	2012 patients, 405 patients with recurrence, hold-out validation (80:20)	Recurrence of gastric cancer following surgery	0.77-0.80c-index

¹Test set results are reported.

Acc: Accuracy (%); c-index: under the receiver operating curve or C-index; Sen: Sensitivity (%); Spe: Specificity (%); ANN: Artificial neural network; CD: Chron’s disease; CNN: Convolutional neural network; CRC: Colorectal cancer; CV: Cross-validation; DNN: Deep neural network; DT: Decision tree; EGBDT: Extreme gradient boosting decision tree; EHR: Electronic healthcare record; GB: Gradient boosting; IBD: Inflammatory bowel disease; KNN: K-nearest neighbor; LASSO: Least absolute shrinkage and selection operator; LR: Logistic regression; NA: Not applicable; NB: Naïve Bayes; NBI: Narrow-band imaging; NCGC: Noncardia gastric cancer; NPV: Negative predictive value; NR: Not recorded; RF: Random forest; SVM: Support vector machine; UC: Ulcerative colitis.