HPLC-grade acetonitrile and formic acid were purchased from Tedia (Fairfield, OH, United States). 3-Hydroxykynurenine and 3-hydroxyanthranilic acid were obtained from Sigma-Aldrich (St. Louis, MO). Tyrosine and xanthurenic acid were purchased from Shanghai Aladdin Bio-Chem Technology Co., LTD.

This prospective observational cohort study protocol (SH9H-2020-T149) was approved by the Institutional Review Board of the Ninth People’s Hospital of Shanghai. The study was part II of the prospective study entitled “Risk factors for paediatric emergence agitation and postoperative pain following maxillofacial surgery and analysis of serum or urine metabolomics in children with agitation—an exploratory study” and was registered at ClinicalTrials.gov (NCT04807998). The study was conducted at the Departments of Anaesthesiology, Otolaryngology, and Oral and Maxillofacial Surgery of the Ninth People’s Hospital of Shanghai Jiaotong University’s School of Medicine from April 2021 to September 2021. It was in compliance with the Declaration of Helsinki for Medical Research. Written informed consent was obtained from parents or guardians of all patients before recruitment.

Paediatric patients aged 3–7 years with an American Society of Anaesthesiologists (ASA) physical status of I or II who were scheduled for tonsillectomy and/or adenoidectomy under sevoflurane-based anaesthesia were recruited. The exclusion criteria included the presence of 1) developmental delays or intellectual disability; 2) a history of neurological or psychiatric diseases; 3) severe liver and kidney dysfunction or other heart and lung diseases; 4) metabolic diseases or inherited diseases; 5) surgery with total intravenous anaesthesia or combined intravenous-inhalation anaesthesia and; 6) under medications or healthy supplements, e.g. vitamins.

All patients had fasted from 22:00 onward before surgery. They all received routine premedications of dolasetron 0.3–0.6 mg/kg and dexamethasone 0.25–0.3 mg/kg (i.v. injection). Non-invasive blood pressure, ECG and pulse oximetry were continuously monitored throughout anaesthesia and surgery. Anaesthetic induction included the sequential administration of intravenous midazolam, propofol, fentanyl and rocuronium to facilitate endotracheal intubation. Anaesthesia maintenance was 1.5%–2.5% sevoflurane supplemented with an intravenous administration of remifentanil (0.05–0.3 μg/kgmin) when necessary to maintain a bispectral index range from 40 to 60. Patients were transferred to the post-anaesthetic care unit (PACU) after surgery completion. The tracheal tube was removed in PACU when spontaneous breathing and swallowing cough reflexes were fully recovered.

During recovery, emergence agitation and postoperative pain were assessed and recorded according to the Paediatric Assessment of Emergence Delirium (PAED) score and the Face, Legs, Activity, Cry and Consolability scale (FLACC) by a researcher who was blinded with the study protocol. PAED and FLACC scores were assessed every 10 min until the patient returned to the ward. Emergence agitation was considered once the patient had a PAED score equal to or greater than 12. Once EA occurred, they were treated with esketamine 0.1–0.25 mg/kg, sufentanil 0.05–0.1 μg/kg or fentanyl 0.5–1 μg/kg administered intravenously as necessary.

Once EA patients were identified, gender- and age-matching non-EA patients, who had the same type of surgery in a similar time period, were randomly selected to match the same number of EA patients for metabolomics analysing comparisons (Figure 1). Patients’ demographics in both groups including age, gender, height, weight, BMI, ASA physical status classification, length of surgery, anaesthesia or anaesthetic dose were collected.

The primary outcome of our study was the usefulness of preoperative urine in metabolomics in predicting the occurrence of postoperative EA. The secondary outcomes were FLACC scores and post-operative complications or adverse events including postoperative nausea and vomiting (PONV), respiratory (e.g., cough, laryngospasm and bronchospasm) and cardiovascular (bradycardia and hypotension) complications.

A first morning urine sample from each patient was collected on the operation day. Urine samples were centrifuged at 13,000 g at 4 C for 10 min, and then, the supernatants were kept at −80°C for further analyses. Prior to liquid chromatography–mass spectrometry (LC–MS) analysis, 400 μL acetonitrile was added to 100 μL aliquots of freshly thawed urine samples. The mixture was, then, vortexed for 2 min and centrifuged at 13,000 g for 10 min at 4°C. The supernatant was transferred to vials for the LC–MS analysis. To monitor the stability of instrument analysis, a quality control (QC) sample was made by combining 10 μL aliquots of each urine sample. Before analysing the entire list of samples, the QC sample was injected six times consecutively to balance the system. Then, the QC sample was injected into every ten samples throughout the entire sample analysis. Variables with a coefficient of variation (CV) ≥ 20% in QC samples were excluded from the data matrix.

The metabolite analysis was performed using a Vanquish UHPLC system (Thermo Fisher) coupled with a Q Exactive Plus mass spectrometer (Thermo Fisher, Milan, Italy) (Lavarello et al., 2021). An InfinityLab Poroshell 120 EC-C18 (2.1 × 100 mm, 2.7 μm, Agilent) column was used for chromatographic separation. The column temperature was set at 45°C. Ultrapure water with 0.1% formic acid (v/v) and acetonitrile (Tedia, Fairfield, OH, United States) with 0.1% formic acid (v/v) constituted phase A and phase B of the mobile phase, respectively. The gradient elution programme for urine analysis was as follows: 1% B (0–1 min), 1%–20% B (1–5.5 min), 20%–30% B (5.5–6 min), 30%–35% B (6–8.5 min), 35%–70% B (8.5–10.5 min), 70%–100% B (10.5–11 min), 100% B (11–13.5 min) and 1% B (13.5–15.5 min). The flow rate was 0.4 ml/min, and the sample injection volume was 1 μL.

Mass detection was performed using a heated electrospray source probe both in positive and negative modes with the following settings: scan mode: DDA mode, one full scan followed by five MS/MS scans. The scan range was set from 67 to 1,000 m/z. The scanning resolutions of MS1 and MS2 were 70,000 (at m/z 200) and 17,500, automatic gain control was 1 × 10⁶ and 5 × 10⁵, and the maximum injection times were 100 ms and 50 ms, respectively. The spray voltages in positive and negative modes were 3,200 and 2800 V, respectively. The auxiliary gas temperature was 300°C, and the capillary temperature was 320°C. Sheath gas and auxiliary gas were set at 30 and 10 a.u., respectively. The s-lens RF level was 50 V. The collision energy was set to 15 and 30 to fragment the ions.

Samples were analysed in a random manner. All the data were acquired by Xcalibur 3.0 (Thermo Fisher). Then, the LC–MS raw data were converted via Progenesis QI v2.3 software (Waters Corp., Milford, MA, United States) for peak finding and alignment. Compound identification was achieved by comparing the precision m/z, MS2, RT and other information collected by mass spectrometry with authentic standards or searching the Human Metabolome Database (http://hmdb.ca/) or METLIN (https://metlin.scripps.edu/).

Due to the complexity of untargeted metabolomics, there are currently no standard methods for sample size estimation. Given that this was a prospective preliminary study, we randomly recruited and collected samples with 33 patients per group, which met the requirement for adequate statistical modelling of metabolomics study (Nyamundanda et al., 2013).

A total of 224 paediatric patients were enrolled in the study; however, 147 patients were excluded for the following reasons: 4 declined, 1 had a history of neurological disease, 1 was given drug treatment after admission, 2 were not obtained urine samples, 26 with total intravenous anaesthesia, 112 with combined intravenous-inhalation anaesthesia and 1 failed to score PAED due to post-operative complication. Ultimately, 77 patients completed the study, and their morning urine samples were obtained (Figure 1). Among them, 33 (42.9%) patients developed EA in PACU. Another 33 age- and gender-matched patients without EA were selected to comprise the non-EA group. There were no significant differences in age, gender, height, weight, BMI, length of surgery, anaesthesia or anaesthetic dose between patients in EA and non-EA groups (Table 1). PAED scores and FLACC scores in the EA group were significantly higher than those of the non-EA group (p < 0.01). The incidence of post-operative adverse reactions that occurred in PACU such as PONV, cough, laryngospasm and bronchospasm was not statistically different between the two groups (p > 0.05).

A typical urine LC–MS total ion chromatogram is shown in Supplementary Figure S1. As shown in PCA score plots (Supplementary Figure S2), QC samples were clustered closely together, indicating the high stability of the LC–MS system during the entire sequence. As shown in Figure 2, the urine samples of the EA group and the non-EA group were clearly divided into two clusters in the OPLS-DA score plots. In the positive ion mode, OPLS-DA model parameters R2Y and Q2 were 0.81 and 0.40, respectively. In the negative ion mode, OPLS-DA model parameters R2Y and Q2 were 0.71 and 0.22, respectively. The corresponding 200-fold randomization permutation test results showed that OPLS-DA models were valid without overfitting.

In our study, metabolites with VIP >1, p < 0.05 (FDR<5%) and FC > 1.5 were selected as potential biomarkers. As listed in Table 2, 16 metabolites associated with an increased risk of EA were identified, of which nine metabolites were aromatic amino acids and their metabolites. Others included several acylcarnitine, pyridoxamine, porphobilinogen, 7-methylxanthine and 5′-methylthioadenosine. The heatmap of differential metabolites between EA and non-EA groups is shown in Figure 3. All the 16 metabolites exhibited higher levels in the EA group than in the non-EA group. Using the metabolic pathway analysis function of MetaboAnalyst 5.0, several related metabolic pathways with p < 0.05 were found, including phenylalanine metabolism, phenylalanine, tyrosine and tryptophan biosynthesis, and tryptophan metabolism.

The LASSO regression selected the two most predictive metabolites, L-tyrosine and pyridoxamine (Supplementary Figure S3). Then, the backward stepwise logistic regression was used to examine the association between the two metabolites and occurrence of EA. The result showed that only L-tyrosine (p = 2.0 × 10^–4) was retained in the model. The odds ratio (OR) (95% CI) was 5.27 (2.20–12.63). The AUC of ROC (95% CI) was 0.81 (0.70–0.91). The sensitivity and specificity were 0.67 (0.51–0.83) and 0.88 (0.77–0.99), respectively. The positive predictive value (PPV) and negative predictive value (NPV) were 0.846 and 0.725, respectively. At internal 10-fold cross-validation, the model was shown to be robust with an AUC of 0.78 (0.67–0.90) (Figure 4).