The application of artificial intelligence methods to gene expression data for differentiation of uncomplicated and complicated appendicitis in children and adolescents - a proof of concept study –

According to the ethical vote, blood samples (> 5 ml) were only collected during routine blood test at the emergency department. Isolation of PBMCs was carried out within 1 h after blood collection. Therefore, blood from each patient was suspended in phosphate-buffered saline (PBS, 1:1 ratio) followed by density gradient centrifugation (Ficoll PM400, GE Healthcare, Pittsburgh, PA; room temperature, 30 min at 400 g). Thereafter, the monocyte layer was re-suspended with PBS and centrifugated (twice, 5 min each at 400 g), followed by a final re-suspension with 1 ml PBS, centrifugation and removal of the supernatant. Native cells were frozen at − 20 °C (Mr. Frosty, Waltham, MA) and finally stored at − 80 °C in liquid nitrogen. By using the NucleoSpin RNA Plus kit (Macherey-Nagel, Germany) total RNA was isolated from the PBMCs. RNA quality control (2100 Bioanalyzer, Agilent Technologies; RNA 6000 Pico Kit) and RNA quantity control (Nanodrop 2000 spectrophotometer; Thermo Scientific) were performed. Samples displaying a Bioanalyzer’s RNA integrity number (RIN) above seven fulfilled quality control standards and were labelled by generation of fluorescent cRNA (complementary RNA; Low Input QuickAmp Labeling Kit, Agilent Technologies). 1st strand synthesis using random primer/oligo-dT primer mixture, 2nd strand synthesis and synthesis of cRNA labelled with cyanine 3-CTP and were performed. Cy3 labelled cRNA (600 ng) was hybridized (65 °C for 17 h; ArrayXS Human Agilent microarray; design ID 79407; Agilent Gene Expression Hybridization Kit, Agilent Technologies; Oak Labs, GmbH, Henningsdorf, Germany) followed by microarray wash and scanning (SureScan Microarray Scanner; Agilent Technologies). According to manufacture’s protocols a resolution of 3-μm was used to generate 20 bit TIF files.

After appendectomy, the appendices were histopathologically examined by two pathologists of which the second examination was a blinded evaluation by a specialized pediatric pathologist. Thereafter, primary diagnosis had been corrected in two cases. Appendicitis was classified in accordance with Carr (9):

A supervised learning algorithm was used to analyze gene expression data and to build a prediction model for diagnosis of appendicitis based on relevant biomarkers. Usually, this is a two-step process summarized as discovery and validation [15]. Due to a limited number of included patients, we favored the widely accepted resampling technique to alternate between the discovery and the validation phase in up to thousand iterations. Thus, to develop biomarkers in this experimental setting, resampling was performed with the bootstrap method for dataset augmentation: Input data consisted of a large number of subsets of samples for varying threshold values. Model building with biomarker selection was carried out on a portion of these data (“discovery set”) with measurement of the performance in an independent data set (“validation set”). Input data consisted of n samples, each of these described by a set of p variables – represented by the biomarker values. Concretely, the data matrix consisted of n lines and p columns. Then, relevant biomarkers were identified: a sequence of distinct biomarker signatures {bm₁,. . ..,bm_j,.. ., bm_m} was built and subsequent implementation of a binary classification problem was performed, fitting the parameters of a logistic regression model on the discovery data X_discovery whose colums p_bmj were filtered according to the biomarker signatures. The quality of each biomarker signature was measured on the discovery data, while all performance values were obtained by measurement of the trained model on the validation data [15].

After definition of the best model, it was used to predict the diagnostic status of a patient with class probability. Output class probabilities can be interpreted as separation thresholds between class prediction. The thresholds represent trade-offs for the model to predict true/false positive and true/false negative rates. The diagnostic ability of the model with regard to sensitivities and specificities was tested on the validation dataset: true and false positive rates were counted at different thresholds [15]. For illustration of the results, a receiver operating characteristic (ROC) plot was created.

The primary endpoint of the study was the pretherapeutic distinctness of the histopathological entities demonstrated by an area under the curve (AUC) in the ROC analysis exceeding at least 50%. A relevant decisive degree was assumed at an AUC of at least 80%.

Regarding evaluation of the artificial intelligence approach concerning effective differentiation of the inflammatory entities, bootstrapping allowed for the calculation of standard errors and confidence intervals.

Statistical analysis regarding epidemiological and routine laboratory data was performed with Mann-Whitney-U-test for continuous and Chi-Square-test for categorical parameters. Welch’s t-test was used for statistical analysis of gene expression values after quantile normalization as previously described [10]. Values are shown as percentages or mean ± SD. Level of significance was p < 0.05. Statistical analysis was performed with GraphPad Prism software (version 9.1.0, La Jolla, CA).

Epidemiological data have already been published elsewhere [12]. Acute appendicitis was suspected sonographically in 33 otherwise healthy patients. After primary inclusion following informed consent, four cell samples (three samples of patients with GA, one of a patient with PA) were secondarily excluded due to signs of degradation within the RNA quality control. After two histopathological examinations the PA group consisted of 13 and the GA group of 16 samples. Mean age in the PA group was 11.5 ± 2.7 years and in the GA group 11.1 ± 2.6 years. Gender distribution was as follows: 4 patients in the PA group were female and 9 male, in the GA group 10 patients were female and 6 male. Mean CRP was significantly upregulated in patients with GA compared with those affected by PA (71.4 ± 58.81 mg/L vs. 23.05 ± 19.26 mg/L, p < 0.05). Mean eosinophilic granulocytes were significantly upregulated in patients with PA compared with patients with GA (0.8 ± 0.1 vs. 0.6 ± 0.18, p < 0.05) without any other significant differences in the differential blood counts.

Patient demographics, distribution of histopathological entities and mean symptom durations are illustrated in Table 1.

As published previously, out of a total of 56,666 analyzed genes 3594 (6.3%) were significantly differentially expressed [12]. Resampling allowed the definition of a best performing discriminatory biomarker signature which consisted of a set of four genes from of a total of 56,666 genes: ERGIC and golgi 3, regulator of G-protein signaling 2, Rho GTPase activating protein 33, and Golgi Reassembly Stacking Protein 2. Figure 1 displays the receiver operating characteristic curve with illustration of the particular sensitivities and specificities of the analyzed thresholds after resampling (mean values ± SE). The area under the curve (AUC) was 84% (SE 8, CL 68.2).