Galectin-3 and S100A9: Novel Diabetogenic Factors Mediating Pancreatic Cancer–Associated Diabetes

Approximately 1% of new-onset diabetes is attributed to pancreatic cancer–associated diabetes (PCDM), which occurs in ∼50% of pancreatic adenocarcinoma patients within 24 months before the diagnosis of cancer (1–4). Pancreatic cancer (PC) is the fourth leading cause of cancer deaths in the U.S. and is projected to become the second leading cause of cancer deaths by 2030 (5). Surgical resection of tumor is the only potential cure, but ∼85% of patients present with unresectable tumors (6). Therefore, PC is the most lethal cancer, with a 5-year survival rate of only 7.7% (6), underscoring the urgent need for detecting PC at an early/resectable stage.

The intriguing phenomenon of PCDM may provide a window of opportunity for early detection of PC. At the onset of PCDM, the tumor is generally early and amenable to surgical resection (4). However, screening all subjects with new-onset diabetes with computed tomography or MRI for PC is not feasible, because only 1% of patients with new-onset diabetes older than 50 years are diagnosed with PC (7). Therefore, the majority of individuals with new-onset diabetes have type 2 diabetes, and thus a screening tool to identify those who are at high risk of PCDM and require further evaluations is needed. The Enriching New-Onset Diabetes for Pancreatic Cancer (ENDPAC) model has been developed to stratify subjects with new-onset diabetes older than 50 years into high, intermediate, or low risk of PC based on changes in weight, changes in blood glucose, and age at onset of diabetes (7). However, in that study, 25% of patients were categorized as intermediate risk by the ENDPAC model, with a 3-year PC incidence of only 0.5%; therefore, this subgroup needed to be further screened by biomarkers to identify high-risk patients for further evaluations (7).

Multiple lines of evidence indicate that PCDM is a paraneoplastic syndrome mediated by tumor-secreted factors (1,8,9), which may be used to distinguish PCDM from type 2 diabetes in subjects with new-onset diabetes. PCDM is characterized by marked peripheral insulin resistance, but the mediators of insulin resistance remain obscure (10–12). Patients with PCDM have decreased glucose metabolism despite a higher plasma insulin level compared with control subjects, and resection of the tumor reduces peripheral insulin resistance (11) and resolves diabetes in 57% of patients with PCDM despite a loss of islet mass after surgery (13). In vitro, tumor extracts from patients with PCDM reduce insulin-mediated glycogen synthesis in skeletal muscle (11), and conditioned media (CM) of PC cell lines impairs peripheral glucose metabolism (14) and glucose tolerance (15). Notably, PC harbors a chronic inflammatory microenvironment (16), and chronic inflammation is a crucial mediator of insulin resistance (17). Analysis of gene ontology also showed that secretory genes differentially expressed in tumors of patients with PCDM are related mainly to inflammation (18). Therefore, we hypothesized that insulin resistance and PCDM are induced by overexpression of specific inflammatory mediators in PC, and blood levels of those mediators may facilitate differentiation between PCDM and type 2 diabetes. In this study, we used proteomic and bioinformatics approaches to search for potential mediators of insulin resistance in PCDM and assessed the usefulness of their blood levels in distinguishing PCDM from type 2 diabetes.

The method of analysis is summarized in Supplementary Fig. 1A. We first performed proteomic analyses to analyze proteins in the CM of MIA PaCa-2 cells. CM was collected and fractionated by SDS-PAGE, and distinct proteins were identified using mass spectrometry (MS). The proteins that had ≥10-fold overexpression in the PC tissue transcriptome data set (GSE16515) from the Gene Expression Omnibus (GEO) database were further analyzed by ingenuity pathway analysis and glucose uptake assay to identify potential diabetogenic factors. The expression of those factors in PC tissues was assessed by immunohistochemistry, and their serum levels were measured by ELISA and compared between patients with PCDM, PC patients without diabetes, patients with type 2 diabetes, patients with pancreatic/peripancreatic tumors other than PC, and healthy control subjects.

CM from MIA PaCa-2 cells was collected and processed as previously described (19). In brief, supernatants were concentrated and desalted by centrifugation in Amicon 3-kDa cutoff spin columns (Millipore, Billerica, MA). The protein concentrations of supernatants were determined using BCA protein assay reagent from Pierce Chemical (Rockford, IL).

CM (40 μg) was manually excised from silver-stained gels and digested overnight with trypsin (Promega, Madison, WI) at 37°C. Tryptic peptides were resuspended in 0.1% trifluoroacetic acid and analyzed by liquid chromatography–tandem MS (LC-MS/MS) using an LTQ-Orbitrap Velos hybrid mass spectrometer (Thermo Fisher Scientific, Waltham, MA). Peptide separations were performed online with MS by nanoflow LC (Dionex Ultimate 3000; Dionex Corporation). Proteins were identified using Proteome Discoverer version 1.3 (Thermo Fisher Scientific) and Mascot Daemon search engine (version 2.3.02) to search against the Swiss-Prot 57.2 version of Homo sapiens (human) protein database containing 20,232 sequences.

Recombinant His-tagged galectin-1, galectin-3, and S100A9 were cloned into pET28a vector (Novagene, Bedford, MA) (Supplementary Table 1). Each recombinant protein was expressed by Escherichia coli strain BL21 after isopropyl β-D-1-thiogalactopyranoside induction and purified by an Ni²⁺-chelating Sepharose column (GE Healthcare, Chicago, IL). The endotoxin was removed by 1% Triton X-114 before use.

Glucose uptake of C2C12 myotubes was assessed by measuring the fluorescent glucose analog 2-[N-(7-nitrobenz-2-oxa-1,3-diazol-4-yl)amino]-2-deoxy-d-glucose (2-NBDG) (Molecular Probes, Eugene, OR), as described previously (20). In brief, C2C12 myotubes were treated with 25, 50, and 100 ng/mL of CM proteins or 50 and 100 ng/mL of purified recombinant galectin-1, galectin-3, S100A9, or commercial recombinant TIMP-1 (Thermo Fisher Scientific) depending on the experiments. The media was replaced with glucose-free DMEM for 1 h prior to addition of 100 nmol/L insulin for 30 min and then 50 μmol/L 2-NBDG for 20 min. 2-NBDG was measured by a microplate reader (Beckman Coulter, Brea, CA) with fluorometry excitation and emission wavelength at 485 and 535 nm, respectively.

Four commercial antibodies were used for immunoblotting (Supplementary Table 2). Protein extracts (20 μg) were separated by SDS-PAGE and transferred onto polyvinylidene fluoride membrane. For detection of target proteins, the membrane was incubated with specific primary antibodies and corresponding secondary antibodies conjugated with horseradish peroxidase. The bound antibodies were detected using enhanced chemiluminescence reagent (Millipore). Densitometry was performed using ImageQuant version 5.2 software from Molecular Dynamics (GE Healthcare).

Patients with PCDM (91 with histology/cytology-confirmed pancreatic adenocarcinoma with fasting blood glucose >126 mg/dL or HbA_1c >6.5% or 48 mmol/mol at diagnosis, without a history of diabetes or with a history of diabetes diagnosed within 24 months preceding the diagnosis of PC [3]) diagnosed at a tertiary referral center (National Taiwan University Hospital) were consecutively enrolled between January 2006 and September 2018 and randomly split into the training set (n = 50) and test set (n = 41). Tumor stage was defined according to the eighth edition tumor, node, metastasis (TNM) system of the combined American Joint Committee on Cancer (AJCC)/Union for International Cancer Control (UICC) (21). In addition, the training set also included five other subgroups of subjects (n = 50 in each subgroup), including PC without diabetes (fasting blood glucose <100 mg/dL and HbA_1c <5.6% or 38 mmol/mol, without the use of antidiabetic medication), acute or chronic pancreatitis, other pancreatic/peripancreatic tumors (e.g., intraductal papillary mucinous neoplasm or ampullary tumor), new-onset type 2 diabetes (fasting blood glucose >126 mg/dL and HbA_1c >6.5% or 48 mmol/mol without a history of diabetes), and healthy control subjects (normal fasting blood glucose and HbA_1c levels without the use of antidiabetic medication). The test set included 41 patients with PCDM and 41 patients with new-onset type 2 diabetes. The patients with new-onset type 2 diabetes and healthy control subjects were selected from subjects who underwent a comprehensive physical checkup at the same center, including blood tests, radiological and endoscopic examinations, and abdominal sonography, and were found to be free of malignancy and pancreatitis; none developed PC between enrollment and the time of this study. The diagnosis of diabetes in patients with PCDM and subjects with type 2 diabetes was based on fasting blood glucose and HbA_1c levels measured at the recruiting center. Blood samples were collected after an overnight fast, and those of PC patients were collected before treatment for PC. Serum was separated by centrifugation and stored at −80°C until use, and all samples were coded for blind analysis. Serum levels of galectin-3 and S100A9 were determined by sandwich ELISA (R&D Systems, Minneapolis, MN). The study was approved by the Institute Research Ethical Committee of National Taiwan University Hospital, and all participants provided informed consent.

Tumor tissue samples were processed for paraffin embedding and 5-μm sections were prepared. Standard avidin-biotin immunohistochemical analyses of the sections were performed according to the manufacturer’s recommendations (Leica Biosystems, Newcastle, U.K.). In brief, tissue sections were deparafinized, rehydrated, heated for antigen retrieval, blocked, and reacted with primary antibodies (Supplementary Table 2), secondary antibodies, and substrate 3,3′-diaminobenzidine. In addition, the sections were counterstained with hematoxylin. Expression of galectin-3 and S100A9 was scored using an H score ranging from 0 to 300, defined as [(% of 0) × 0] + [(% of 1+) × 1] + [(% of 2+) × 2] + [(% of 3+) × 3] (0, nil; 1+, low or weak; 2+, moderate; 3+, high or strong; %, percent of tumor cells of a certain intensity).

Values for all measurements were expressed as mean ± SD. The Kruskal-Wallis test was used for comparing continuous variables, and Dunn test was used for multiple comparisons. The ability of serum galectin-3 and S100A9 levels to differentiate between PCDM and type 2 diabetes was analyzed by logistic regression analysis. Optimal cutoff values of galectin-3 and S100A9 levels were derived from the training set using Youden index and validated in the test set. All statistical tests were two sided, and P values <0.05 were considered statistically significant. Statistical analyses were performed using Stata14 (StataCorp, College Station, TX).

We first assessed whether the CM of MIA PaCa-2 cells affected peripheral insulin sensitivity. We found that the CM of MIA PaCa-2 cells significantly reduced insulin-stimulated glucose uptake in C2C12 myotubes compared with control, supporting the concept that MIA PaCa-2 cells secrete proteins that induce insulin resistance (Supplementary Fig. 2).

Proteomic analyses of the CM of MIA PaCa-2 cells yielded 1,696 proteins, and 86 of those were upregulated in PC tissues according to microarray analysis (GSE16515) (Supplementary Fig. 1A). We further categorized these proteins into functional groups based on the Ingenuity Pathway Analysis (IPA; Ingenuity Systems; http://www.ingenuity.com). Ranking and significance of the biofunctions were tested by P values. IPA revealed that those proteins were involved in cell death and survival (n = 41), metabolic disease (n = 22), inflammatory response (n = 14), cancer/endocrine system disorders (n = 8), and cancer/gastrointestinal disease (n = 75) (Supplementary Table 3). Based on the notion that the diabetogenic factors might be linked to inflammation, metabolic disease, and cancer, we selected four candidate secreted proteins for further investigation, including galectin-1, galectin-3, S100A9, and TIMP-1, and verified their existence in the CM of MIA PaCa-2 cells (Supplementary Fig. 1B). Representative MS/MS spectra of identified peptides are shown in Supplementary Fig. 3.

To determine whether the selected candidate proteins can induce insulin resistance, we assessed their effects on glucose uptake of skeletal muscle cells in vitro. Pretreatment with galectin-3 and S100A9 suppressed insulin-stimulated glucose uptake in C2C12 myotubes in a dose-dependent manner, whereas galectin-1 and TIMP-1 did not inhibit glucose uptake (Supplementary Fig. 4). Therefore, galectin-3 and S100A9 were considered as potential diabetogenic factors.

We next evaluated the expression of galectin-3 and S100A9 in paraffin-embedded PC tumor tissues by immunohistochemistry. Tumors with PCDM (n = 22) had significantly higher expression of galectin-3 (P = 0.022) and S100A9 (P = 0.020) compared with tumors without DM (n = 11) (Supplementary Fig. 5).

To assess their potential effectiveness in distinguishing PCDM from type 2 diabetes, the serum levels of galectin-3 and S100A9 were measured in the training and test sets. In the training set, patients with PCDM had significantly higher levels of S100A9 and galectin-3 compared with patients with type 2 diabetes, other tumors, and pancreatitis (Table 1 and Fig. 1).

The area under the curve (AUC) for S100A9 and galectin-3 in differentiating between PCDM and type 2 diabetes was 0.79 (95% CI 0.70–0.87) and 0.73 (95% CI 0.64–0.83) (Fig. 2), respectively. The optimal cutoff for S100A9 was 59.0 ng/mL, with a sensitivity of 70.0% and a specificity of 74.0%. The optimal cutoff for galectin-3 was 6.5 ng/mL, with a sensitivity of 62.0% and a specificity of 80.0%.

In the test set, patients with PCDM also had significantly higher S100A9 and galectin-3 levels compared with patients with type 2 diabetes (both P < 0.001) (Table 1 and Fig. 1). The AUC in differentiating between PCDM and type 2 diabetes was 0.77 for S100A9 (95% CI 0.67–0.87) and 0.83 (95% CI 0.74–0.92) for galectin-3 (Fig. 2). The combination of S100A9 and galectin-3 did not perform significantly better than either S100A9 (Bonferroni-corrected P = 0.056) or galectin-3 (Bonferroni-corrected P = 0.312) alone. Using cutoff levels derived from the train set, the sensitivity and specificity of S100A9 were 69.8% and 58.1%, respectively. The sensitivity and specificity of galectin-3 were 72.1% and 86.1%, respectively.

The sensitivity of S100A9 and galectin-3 stratified by stages of cancer and duration of diabetes before diagnosis of PC is summarized in Supplementary Table 4. There were no significant differences between cancer stages (S100A9, P = 0.973; galectin-3, P = 0.534) and durations of diabetes before PC (S100A9, P = 0.934; galectin-3, P = 0.162).

To predict the risk of PC in patients with new-onset diabetes, Sharma et al. (7) proposed a simpler model including one predictor (weight loss ≥2.5 kg) and a full model that calculates the ENDPAC score based on three predictors (weight loss, change in blood glucose, and age), with scores of 0, 1–2, and 3 or greater categorized as low, intermediate, and high risk, respectively. When applied to subjects in this study, weight loss ≥2.5 kg achieved an AUC of 0.79 (95% CI 0.74–0.84) in differentiating between PCDM and type 2 diabetes, with a sensitivity of 60.4% and a specificity of 97.8% (Table 2 and Supplementary Fig. 6). Combining galectin-3 with weight loss achieved significantly better performance than weight loss alone (AUC 0.88 [95% CI 0.84–0.93], Bonferroni-corrected P < 0.001), with a sensitivity of 84.6% and a specificity of 82.4%. Combination with S100A9 also improved the performance of weight loss (AUC 0.89 [95% CI 0.84–0.93], Bonferroni-corrected P < 0.001), with a sensitivity of 92.3% and a specificity of 65.9%. Combining both galectin-3 and S100A9 with weight loss did not provide significant further benefit than combining weight loss with either galectin-3 or S100A9.

Calculating the ENDPAC score requires a blood glucose level that did not reach the level of diabetes (fasting blood glucose ≤125 mg/dL) and preceded the onset of diabetes by 3–18 months (7). Only 17 patients with PCDM and 20 patients with type 2 diabetes had such information to allow calculation of the ENDPAC score. In those patients, ENDPAC score achieved an AUC of 0.89 (95% CI 0.79–1.00) in differentiating between PCDM and type 2 diabetes. No significant improvement in performance was noted when ENDPAC score was combined with galectin-3 (AUC 0.94 [95% CI 0.86–1.00], Bonferroni-corrected P = 0.361), S100A9 (AUC 0.90 [95% CI 0.78–1.00], Bonferroni-corrected P = 1.000), or both galectin-3 and S100A9 (AUC 0.93 [95% CI 0.84–1.00], Bonferroni-corrected P = 0.637) (Table 2 and Supplementary Fig. 6).

Microarray data of PCDM and PC without diabetes from the GEO database (GSE15932) were compared to explore differences between PCDM and PC without diabetes. In PC without diabetes, 1,781 genes were upregulated compared with PCDM, and 4 of those were also among the 86 candidate biomarkers for PC identified by the above proteomic and microarray analyses, including galectin-1, SLC39A10, CTSC, and DNAJB1 (Supplementary Fig. 7). Taken together with the findings that galectin-1 was identified in the CM of PC cells and did not inhibit insulin-induced glucose uptake in myotubes (Supplementary Figs. 3 and 4), galectin-1 might be a potential biomarker for PC without diabetes.

This study used high-throughput proteomics coupled with transcriptomic and bioinformatic analyses to identify candidate diabetogenic factors and verified the ability of galectin-3 and S100A9 to induce insulin resistance and their differential expression in PCDM tissues. Furthermore, serum levels of galectin-3 and S100A9 were exclusively increased in patients with PCDM and distinguished PCDM from type 2 diabetes. These results support that galectin-3 and S100A9 are PC-produced diabetogenic factors that mediate peripheral insulin resistance/PCDM and may facilitate identification of patients at high risk of PCDM for further evaluations among subjects with new-onset diabetes.

Our results corroborated previous studies that observed overexpression of galectin-3 in human PC tissues (22,23). Galectin-3 is a β-galactoside–binding lectin with proinflammatory functions (24) and has been shown to promote proliferation and migration/invasion of PC cells (23,25). We further demonstrated that galectin-3 inhibited insulin-stimulated glucose uptake in muscle cells, and the finding that serum galectin-3 levels were exclusively increased in patients with PCDM might explain why previous studies evaluating galectin-3 as a biomarker for PC yielded mixed results. Xie et al. (22) found that serum galectin-3 levels were higher in PC patients compared with patients with benign pancreatic diseases and healthy control subjects, whereas Brandi et al. (26) found no significant differences in serum galectin-3 levels between PC patients and control subjects. These differences might be due to variations in the proportion of PC patients with PCDM between those studies.

S100A9 and S100A8, both members of the S100 family proteins, form the heterodimer calprotectin, which binds to Toll-like receptor 4 (TLR4) and amplifies the inflammatory responses of phagocytes (27). Previous studies have implicated S100A8/A9 in inflammation (27,28), gestational diabetes mellitus (29), and various cancers, including PC (28,30). Our results are in line with Wang et al. (31) who also noted overexpression of S100A9 in tumors of patients with PCDM, and we further demonstrated that S100A9 dose-dependently inhibited insulin-stimulated glucose uptake of muscle cells. Basso et al. (15) identified overexpression of a peptide corresponding to the N terminus of S100A8 in PCDM tissues and showed that the peptide could impair glucose catabolism in vitro and induce hyperglycemia in vivo (32). Taken together, these results support that S100A8/A9 mediates insulin resistance and PCDM.

PCDM is characterized by both β-cell dysfunction and marked peripheral insulin resistance. Whereas adrenomedullin and macrophage migration inhibitory factor have been shown to mediate β-cell dysfunction in PCDM (9,33), the mediators of insulin resistance remain unclear (10). A previous study suggested islet amyloid polypeptide as a possible mediator of insulin resistance in PCDM (34), but subsequent studies did not validate this observation (35). This study provides strong evidence for galectin-3 and S100A9 as PC-derived mediators of insulin resistance and lends support to the notion that overexpression of proinflammatory factors in PC induces chronic inflammation and subsequently insulin resistance and PCDM, and possibly other metabolic consequences such as cachexia. Further research is warranted to uncover the mechanisms by which galectin-3 and S100A9 induce insulin resistance.

Given the low incidence of PCDM among subjects with new-onset diabetes, a screening modality that can identify subjects at high risk of PCDM for further evaluations is needed to take advantage of new-onset diabetes as an opportunity for early detection of PC (3,7). Previous research showed that adrenomedullin inhibited insulin secretion of β-cells in vitro and in vivo, and plasma levels of adrenomedullin were higher in patients with PCDM compared with PC patients without diabetes and control subjects (9). However, the potential usefulness of adrenomedullin in distinguishing PCDM from type 2 diabetes has not been further validated. Other studies have adopted an alternative approach of developing predictive models based on clinical parameters. Boursi et al. (36) developed an 11-parameter predictive model that achieved 44.7% sensitivity, 94.0% specificity, and 2.6% positive predictive value when the threshold of predicted PC risk was set at 1% over 3 years. Given its high specificity, the model would subject only 6.19% of subjects with new-onset diabetes to further evaluations, but the model would miss 55.3% of patients with PCDM. The ENDPAC model required only three parameters, and Sharma et al. (7) showed that subjects with an ENDPAC score of 3 or higher were at high risk of PC requiring further evaluations, whereas those with a score of 0 or lower were at extremely low risk of PC. However, the incidence of PC in subjects categorized as intermediate risk by the ENDPAC model was low but not negligible; therefore, those subjects needed to be further screened with biomarkers to identify a subset for which further evaluation for possible PC would be cost-effective (7). Furthermore, calculation of the ENDPAC score requires nondiabetic blood glucose levels within 3–18 months preceding the onset of diabetes (7), which are often unavailable in real-life clinical situations as in our patients, limiting the clinical applicability of the ENDPAC model. Our results showed that combining galectin-3 or S100A9 with weight loss achieved comparable performance with the full ENDPAC model in distinguishing PCDM from type 2 diabetes, enabling screening of patients in whom ENDPAC scores cannot be calculated. Because of the small number of subjects in whom the ENDPAC score could be calculated, we could not assess whether galectin-3 or S100A9 might further enrich individuals categorized as intermediate risk by the ENDPAC score. Taken together, combined use of predictive models and biomarkers such as S100A9 and galectin-3 is needed to meet the contradictory needs of achieving a high sensitivity to avoid missing PC while maintaining a low false-positive rate to reduce unnecessary examinations. Further research is needed to validate our results and investigate how various predictive models and biomarkers should be combined.

In an exploratory analysis, we found four candidate biomarkers for PC without diabetes that warrant further study. Galectin-1 is of particular interest, as we found that it is a secreted protein of PC cells without diabetogenic function and differentially overexpressed in PC without diabetes, and Martinez-Bosch et al. (37) reported that plasma galectin-1 levels were significantly increased in PC patients compared with control subjects. Therefore, galectin-1 might be a potential biomarker for PC without diabetes and should be further evaluated in future study.

This study uncovered galectin-3 and S100A9 as novel mediators of insulin resistance in PCDM and validated their usefulness in distinguishing PCDM from type 2 diabetes in patients with new-onset diabetes. The notable finding that galectin-3 and S100A9 achieved ∼70% sensitivity in resectable stage I/II PCs, which are notoriously difficult to detect, lends further support to the notion that PCDM and its biomarkers provide a window of opportunity for detecting PC while it remains amenable to surgical resection, the only potential cure. A limitation was that only 40% of the included patients with PCDM had stage I/II PC, reflecting the fact that most PC patients have unresectable stage III/IV tumors at diagnosis. The effectiveness of using galectin-3 and S100A9 to screen subjects with new-onset diabetes and its impact on clinical outcomes need to be further validated prospectively. Second, the healthy control subjects were younger than other groups. To ensure that the control subjects were healthy without PC, other malignancies, pancreatitis, and diabetes, we recruited control subjects from individuals who were confirmed to be without those conditions after a comprehensive health checkup at the same referral center. Individuals who underwent the self-paid health checkup were mostly aged between 30s and 60s, and those with the above conditions were not eligible as control subjects; therefore, the selected control subjects were younger than other groups. However, the finding that control subjects had significantly lower serum galectin-3 and S100A9 levels than patients with PCDM could not be attributed to younger age, because other groups (type 2 diabetes, other tumors, and PC without diabetes) who were comparable with patients with PCDM with respect to age also had significantly lower galectin-3 and S100A9 levels compared with patients with PCDM. Last, because of the cross-sectional design, this study could not infer causality between galectin-3 or S100A9 and PCDM. A cohort study that prospectively follows subjects with new-onset diabetes is needed to verify the relation between galectin-3, S100A9, and PCDM.

In conclusion, galectin-3 and S100A9 are strongly associated with PCDM and may distinguish PCDM from type 2 diabetes in subjects with new-onset diabetes.

Acknowledgments. The authors thank the Eighth Core Lab (Department of Medical Research, National Taiwan University Hospital) and the Proteomics and Protein Function Core Lab (Center of Precision Medicine, National Taiwan University) for providing support for the study.

Funding. This work was supported by grants from the Ministry of Science and Technology, Republic of China (MOST 105-2314-B-002-114-MY2 and MOST 104-2320-B-002-045-MY3), National Taiwan University Hospital (104-P09 and 105-P01), and the Ministry of Education, Taiwan.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. W.-C.L. conceived and designed the study; acquired, analyzed, and interpreted data; and drafted the manuscript. B.-S.H. acquired, analyzed, and interpreted data and drafted the manuscript. Y.-H.Y., H.-H.Y., P.-R.C., C.-C.H., and H.-Y.H. acquired, analyzed, and interpreted data. M.-S.W. conceived and designed the study, analyzed and interpreted data, critically revised the manuscript for important intellectual content, and provided administrative and technical/material support. L.-P.C. conceived and designed the study; acquired, analyzed, and interpreted data; drafted the manuscript; and supervised the study. M.-S.W. and L.-P.C. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.