This prospective observational multicenter cohort study was conducted from June 2021 to January 2024, with patient recruitment from June 2021 to December 2023 and follow-up completion in January 2024. In this study, 586 older adult patients (aged ≥60 years) scheduled for hemiarthroplasty due to hip fractures were initially enrolled. Inclusion criteria were age of 60 years or above and undergoing total hip replacement under uniform general anesthesia. Exclusion criteria were American Society of Anesthesiologists physical status classification of grade III or higher, Mini-Mental State Examination (MMSE) score below 23, history of central nervous system (CNS) disorders or psychiatric diseases, current use of sedatives or antidepressants, preoperative biochemical abnormalities indicating renal dysfunction or active hepatic disease, preexisting diabetes mellitus, history of neurosurgical or cardiac surgery, refusal to complete study protocols or inability to comprehend the study language, severe auditory or visual impairment, illiteracy, Parkinson disease, and alcohol or substance dependence. All baseline data (demographics and clinical characteristics), preoperative serum samples, and postoperative outcomes (POD incidence and cognitive function) were collected prospectively according to a predefined protocol without retrospective data retrieval.

All experimental procedures in this study were approved by the Ethics Committee of the Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine (approval number 2021-YS-237). The study was conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from each participant or their legally authorized representative prior to enrollment. To protect participant privacy and confidentiality, all personal identifiers were removed, and data were anonymized prior to analysis. Access to study data was restricted to authorized research personnel only. No financial compensation was provided to participants. All study procedures in this research have been registered at the Chinese Clinical Trial Registry [24] with the registration number ChiCTR-CPC-15006141.

All patients in the experimental group received a light diet preoperatively. Semiliquid feeding was initiated 6 hours after general anesthesia, followed by a gradual transition to a light regular diet.

Patients underwent standardized general anesthesia for hemiarthroplasty. Premedication included intramuscular midazolam (3 mg). Anesthesia induction was performed with intravenous fentanyl (0.1 mg), propofol (1.5 mg/kg), and vecuronium (0.1 mg/kg). Maintenance anesthesia was administered via a semiclosed circuit system using sevoflurane (1‐1.5 maximum allowable concentration) and oxygen (1 L/min). Continuous intraoperative monitoring included pulse oximetry, invasive arterial blood pressure waveform, end-tidal partial pressure of carbon dioxide, and 3-lead electrocardiography.

Hemiarthroplasty involved replacement of the fractured femoral head with a metallic prosthesis and insertion of a metal stem into the femoral shaft to enhance implant stability.

Preoperative blood samples were collected on the day of the surgery prior to entering the operating room. Postoperative samples were obtained immediately upon transfer from the postanesthesia care unit. Serum was cryopreserved for subsequent high-throughput targeted metabolomics profiling.

Among the 260 enrolled patients undergoing total hip replacement, POD was assessed twice daily (8 AM to 8 PM on postoperative days 1‐3) using the Confusion Assessment Method (CAM) [25,26] as recommended by the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition. Diagnostic criteria were (1) acute onset with fluctuating symptoms, (2) inattention, (3) disorganized thinking, and (4) altered consciousness. POD was confirmed if criteria 1 and 2 were present along with either criterion 3 or 4. Patients were categorized into POD and non-POD groups.

The CAM (POD diagnosis) was conducted by research nurses certified in the CAM scale. Before study initiation, all involved nurses completed certified training for the Chinese version of the CAM scale (certificate numbers archived). The assessment results were reviewed, and the POD diagnosis was confirmed by the designated psychiatrists (see the Perioperative Cognitive Assessment section), thereby avoiding potential bias from nurses’ subjective judgments.

Baseline demographic data, anesthesia- and surgery-related parameters, and other clinical variables were systematically documented. Clinical data collection was performed by research coordinators. Each center designated 1 coordinator (with a master’s degree or higher in nursing or a related field) responsible for recording baseline characteristics (age, sex, and BMI) and surgical and anesthesia parameters (operation duration and anesthesia duration). Data entry was monitored weekly by a clinical research associate who spot-checked 20% of cases to ensure accuracy.

Serum samples were analyzed using the Q300 metabolite assay kit (Metabo-Profile) [27]. Sample processing, including protein precipitation and derivatization, was automated on an Eppendorf epMotion workstation. Metabolite separation and detection were performed using ultrahigh-performance liquid chromatography (UPLC) coupled with tandem mass spectrometry (ACQUITY UPLC Xevo TQ-S system; Waters). The detailed sample preparation protocol and complete UPLC coupled with tandem mass spectrometry parameters are provided in Multimedia Appendix 1. Raw data processing, including peak integration and metabolite quantification, was conducted using the QuanMET software (version 2.0; Metabo-Profile). Quality control (QC) samples were analyzed intermittently to ensure analytical consistency.

To avoid multicenter data bias, the following standardization measures were implemented before and during the study.

To prevent knowledge of metabolomics results from influencing cognitive assessments or POD diagnosis, this study used an assessor-analyst double-blinding design.

A schematic overview summarizing the complete study workflow is provided in Figure S1 in Multimedia Appendix 1. This diagram visually integrates the key stages of the research, from participant recruitment and data collection to metabolomics profiling, machine learning model development, and the final predictive outcome for POD.

The study flowchart (Figure 1) illustrates the patient enrollment process in this multicenter observational cohort study. Ultimately, 260 patients met the inclusion criteria and completed the study protocol. Of these 260 patients, 56 (21.5%) were diagnosed with POD, constituting the POD group, whereas 204 (78.5%) without a POD diagnosis were categorized as the non-POD group. As shown in Table 1, no significant differences were observed between the 2 groups in demographic characteristics, surgical duration, or anesthesia time.

The patient flow throughout the study is detailed in Figure 1. All cases of attrition were meticulously documented in the study case report forms, including the specific reason and time point of withdrawal. This documentation was reviewed and signed by the principal investigator at each site to ensure adherence to ethical standards.

Data from patients who dropped out were handled as follows:

This rigorous approach to reporting patient flow and handling missing data ensures the transparency, reproducibility, and ethical integrity of our research findings. The final analytical cohort (N=260) represents a complete-case dataset for the planned metabolomics and machine learning analyses.

To compare metabolite profiles, we used UPLC coupled with triple quadrupole mass spectrometry for absolute quantification of 201 metabolites. As shown in Figure 2, absolute quantification of 201 metabolites revealed 41 (20.4%) significantly differentially expressed metabolites between the POD and non-POD groups (P<.05). These differential metabolites were primarily enriched within 4 core metabolic domains (amino acid metabolism, energy metabolism, lipid metabolism, and bile acid metabolism), suggesting that POD development is closely associated with disruptions in these key biological pathways.

Kyoto Encyclopedia of Genes and Genomes pathway enrichment analysis further confirmed that these differential metabolites were significantly enriched in key pathophysiological pathways, including alanine, aspartate, and glutamate metabolism (neurotransmitter synthesis); phenylalanine, tyrosine, and tryptophan biosynthesis (precursors for dopamine and serotonin); and the citrate cycle (TCA cycle and core energy metabolism; Figure 3).

By performing a union operation on the top 10 differentially abundant metabolites identified through RF model screening (Figure 4) and LASSO regression–based feature selection (Figure 5), we obtained 16 distinct metabolites (see Table S1 in Multimedia Appendix 1 for the complete list). These biomarkers span the aforementioned 4 core metabolic domains and collectively form the feature set for predictive modeling. Subsequently, predictive models were constructed using these 16 metabolites via gradient boosting, LR, and RF algorithms. Among these, the LR model demonstrated strong predictive performance, achieving an AUC of 0.855 (95% CI 0.800‐0.910; Figure 6).

Total samples were randomly partitioned into training and testing sets (7:3 ratio) via computerized randomization algorithms. Differential metabolites underwent rigorous LR modeling, demonstrating robust predictive performance with AUC values of 0.844 (95% CI 0.776–0.912) in the training cohort (Figure 7) and 0.856 (95% CI 0.741–0.971) in the independent test cohort (Figure 8), meeting TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) guidelines for clinical prediction models.

The final LR model was subjected to a comprehensive evaluation on the independent test set (n=78). The model demonstrated good overall accuracy, with a Brier score of 0.1156. The calibration plot (Figure 9) indicated excellent agreement between predicted probabilities and observed outcomes, with a calibration slope of 0.9817.

Decision curve analysis (Figure 10) revealed that using the metabolomics model for clinical decision-making provided a superior net benefit compared to strategies of intervening on all patients or intervening on no patients across a wide range of clinically reasonable probability thresholds (approximately 10% to 60%). This underscores the potential clinical utility of the model for targeting preventive interventions.

In light of the class imbalance, PR curves were analyzed (Figure 11). The model achieved an AP of 0.965 on the test set (training set AP=0.934). This exceptionally high AP, in the context of a 21.5% (56/260) event rate, confirms the model’s robust capability to identify patients with POD with high precision and recall. The PR curve is particularly informative for this imbalanced classification task as it focuses on the performance regarding the positive class (POD), unlike the ROC curve, which can be overly optimistic under class imbalance.

To facilitate clinical application, a nomogram was constructed based on the final LR model (Figure 12), allowing for the visual calculation of an individual patient’s POD risk using the 16 preoperative metabolite levels.

POD, a severe neurological complication in older adult patients, is associated with neuroinflammation, oxidative stress, energy metabolism dysfunction, and cholinergic system imbalance [31]. Using a multicenter high-throughput targeted metabolomics approach combined with machine learning, we identified 16 differentially expressed metabolites as key biomarkers for constructing a POD prediction model.

This multicenter metabolomics study identified 16 differentially expressed metabolites involved in core pathological pathways of POD in older adult patients, including energy metabolism, amino acid homeostasis, bile acid regulation, and oxidative stress. These metabolites provide multidimensional biomarkers for early POD prediction and novel targets for interventions (eg, improving mitochondrial function and regulating gut microbiota). The machine learning–based prediction model has potential for integration into perioperative management to achieve precise POD prevention through dynamic metabolic profiling. However, to facilitate its clinical translation, future studies should focus on (1) integrating these metabolites with readily available key clinical variables to develop a more universally applicable and practical multimodal prediction tool; (2) rigorously validating the refined model in independent, multiregional external cohorts to assess its real-world performance; and (3) further elucidating the potential causal associations between these metabolites and the development of POD, and exploring their specific roles across different surgical subtypes.