Esophagitis

Twelve datasets (3496 patients) with treatment outcomes described in previous studies were collected from public repositories (www.cancerdata.org) or provided by collaborators. Table 1 characterizes these datasets. Given availability, some datasets consist of subsamples of or contain fewer/more patients and/or features than the cohorts described in the original studies. Two datasets were excluded after a preliminary analysis (these datasets are also not mentioned in Table 1) where none of the studied classifiers resulted in an average AUC above 0.51, which is evidence that they contain no discriminative power. Datasets without discriminative power are not suitable for this analysis as we would be unable to determine differences in discriminative performance across classifiers. The patient cohorts of 2 datasets, Wijsman et al.20 (link), 21 , partially overlap but each dataset lists a different outcome (esophagitis and pneumonitis). Datasets were anonymized in the analysis because their identity is not relevant for interpreting the results and to encourage investigators to share their datasets.
Nonbinary outcomes were dichotomized, for example, overall survival was translated into 2‐yr overall survival in the dataset of Carvalho et al.10 (link). Missing data were imputed for training and test sets (the splitting of datasets into training and test sets is described in Section 2.C) by medians for continuous features and modes for categorical features based on the training set. Basing the imputation on the training set avoids information leakage from test to training sets. Categorical features in training and test sets were dummy coded, that is, representing categorical features as a combination of binary features, based on the combined set for classifiers that cannot handle categorical features (Table 2). Dummy coding on the combined set ensures that the coding represents all values observed in a dataset. Features with zero variance in training sets were deleted in the training set and in the corresponding test set. In addition, we removed near‐zero variance features for glmnet to avoid the classifier implementation from crashing during the fitting process. Features in training sets were rescaled to the interval [0,1] and the same transformation was applied to the corresponding test sets. Rescaling is needed for certain classifiers, e.g., svmRadial. All these operations (imputation, dummy coding, deleting (near‐)zero variance features, rescaling) were performed independently for each pair of training and test sets (step 2 in Fig. 1).

Previous works have described the methods for MR analysis of summary data based on two studies^{53 (link),54 (link)}. Here, we used all variants in the tissue-grouped variant sets as proposed instruments to measure their associations to ten outcomes with summary data available from the FinnGen study⁴⁵ : T2D (n total = 215,654; n cases = 35,607), diabetic nephropathy (n total = 213,746; n cases = 3,283), CKD (n total = 216,743; n cases = 3902), peripheral artery disease (PAD) (n total = 213,639; n cases = 7098), heart failure (HF) (n total = 218,208; n cases = 13,087), stroke (n total = 180,862; n cases = 18,661), myocardial infarction (MI) (n total = 200,641; n cases = 11,622), diabetic retinopathy (n total = 216,666; n cases = 14,584), diabetic neuropathy (n total = 163,616; n cases = 1415) and esophagitis (n total = 190,442; n cases = 747) as a control (case definition for all outcomes can be found on Supplementary Table 4). SNP-exposure associations were retrieved from the summary statistics from Vujkovic et al.^{8 (link)}, and SNP-outcome associations come from summary statistics from Finngen⁴⁵ . All tissue-grouped variant sets were considered composed of sufficiently strong instruments based on their F-statistics, considering an F-statistic > 10 as strong enough instrument to avoid weak instrument bias^{55 (link)} (Table 1). Using fixed effects IVW analyses, we combined the effects of the individual genetic instruments to obtain a genetically determined association between exposure and outcome under the assumption of the absence of horizontal pleiotropy. Some variants from the tissue-grouped variant sets were removed from MR analyses due to being palindromic genetic instruments with intermediate allele frequencies. Estimates from the IVW analyses can be interpreted as the odds ratio for the outcome trait(s) per 2.72-fold increase in the odds of T2D (i.e., a one unit change in genetic liability to T2D on the log odds scale). We also run sensitivity analyses using methods MR Egger^{22 (link)}, Weighted Median^{23 (link)} and Weighted Mode^{24 (link)}. Analyses were run using the R-based package ‘TwoSampleMR’^{56 (link)}.

Subjects who underwent the upper endoscopy were recruited from two hospitals as follows: (1) Suzhou: The First Affiliated Hospital of Soochow University and (2) Jintan: Affiliated Hospital of Jiangsu University, between 2015 and 2021. In the two centers, subjects were excluded if they have (1) esophagitis of other etiologies, e.g., pills-induced esophagitis, eosinophilic, radiation, and infectious esophagitis; (2) esophageal varices; (3) esophageal squamous cell cancer. This study was approved by the Ethics Committee of The First Affiliated Hospital of Soochow University and conducted in accordance with the Helsinki Declaration of 1975 as revised in 2000 (the IRB approval number 2022098). All participants signed statements of informed consent before inclusion. Besides, the Z-line endoscopic images were also obtained from an open dataset, HyperKvasir, which now is the largest dataset of the gastrointestinal endoscopy (https://datasets.simula.no/hyper-kvasir/) [12 (link)]. The dataset offers labeled/unlabeled/segmented image data and annotated video data from Bærum Hospital in Norway. The characteristic of the datasets was shown in Figure 1. Each endoscopic image of Z-line was determined and labeled as normal, LA classification A + B (LA A + B), or LA classification C + D (LA C + D) by three rich-experienced endoscopists, based on the LA classification. The endoscopic devices in our hospital include Olympus GIF-Q260, GIF-H290, and Fuji EG-601WR, while in the HyperKvasir dataset, they include Olympus and Pentax at the Department of Gastroenterology, Bærum Hospital.