Early Warning and Stratified Diagnosis of Postoperative Respiratory Failure Based on Ventilator Waveform Image Features

II. Study Objectives

1. Primary Objective To identify and validate the associations between ventilator waveform image features in postoperative patients and the occurrence, progression, and severity stratification of postoperative respiratory failure (PRF).
2. Specific Objectives

• Early warning: To identify combinations of ventilator waveform image features that emerge before the clinical diagnosis of postoperative respiratory failure and have predictive value for its occurrence.
• Risk stratification: To develop a ventilator waveform feature-based model to stratify the risk of postoperative respiratory failure into low-, intermediate-, and high-risk groups.
• Severity stratification: To develop a ventilator waveform feature-based model to stratify patients who have developed postoperative respiratory failure into mild, moderate, and severe categories.
• Algorithm development: To develop and validate an algorithmic framework for automated extraction of ventilator waveform features, early warning, and risk/severity stratification.

III. Study Content

1. Study Design This is a prospective observational cohort study.
2. Study Procedures and Content 2.1 Inclusion Criteria

• Age ≥18 years;
• Patients receiving mechanical ventilation after surgery and transferred directly to the ICU;
• Patients expected to require or actually requiring mechanical ventilation for more than 24 hours after surgery;
• Patients or their legally authorized representatives providing written informed consent. 2.2 Exclusion Criteria
• Respiratory failure occurring before surgery;
• Severe pulmonary disease diagnosed before surgery, including but not limited to severe chronic obstructive pulmonary disease (COPD; GOLD stage 3-4), severe interstitial lung disease (PaO₂/FiO₂ <300 mmHg), active pneumonia, and severe pulmonary hypertension;
• Preoperative mechanical ventilation for >48 hours;
• A decision made immediately after surgery to withhold or withdraw active treatment, including invasive mechanical ventilation;
• Expected survival of less than 48 hours;
• Inability to connect to or record the required ventilator waveform data and/or electronic medical record data;
• Other severe underlying diseases, including but not limited to end-stage liver disease, dialysis-dependent renal failure, and advanced malignancy with a short expected survival. 2.3 Sample Size The sample size will be calculated based on an expected primary endpoint event rate of approximately 20%, with α=0.05 and a statistical power of 1-β=0.80. Considering a 15% attrition rate, the required total sample size is estimated to be approximately 250 patients, with approximately 50 expected endpoint events. This sample size is expected to satisfy the event-per-variable requirements and facilitate robust variable selection and model development. 2.4 Study Population and Stratified Sampling The study population will consist of patients receiving mechanical ventilation who are transferred to the ICU after surgery at our hospital. To ensure adequate representation of different surgical types and disease severity levels and to facilitate subsequent stratified analyses, stratified consecutive enrollment will be adopted.
• Stratification factors: Two levels of stratification factors will be used: surgical grade (Grade I, II, III, and IV) and postoperative disease severity. Postoperative severity will be stratified according to the APACHE II score at ICU admission: ≤15, 16-24, and ≥25, corresponding to mild, moderate, and severe groups, respectively.
• Target stratum proportions and dynamic monitoring: The number of enrolled patients in each stratum will be reviewed monthly. If enrollment in any stratum deviates from the target by more than ±10%, corrective measures will be implemented by adjusting recruitment channels or proactively contacting high-risk clinical departments. At least 25 patients will be enrolled in each stratum to facilitate basic subgroup analyses.
• Stratified sampling procedures: The "surgical type" and "APACHE II score" fields will be mandatory in the EDC system. Stratification labels will be independently confirmed by the investigator and the attending physician. If a patient undergoes multiple surgical procedures, the patient will be classified according to the primary surgical procedure.
• Stratification in statistical analyses: Surgical type and disease severity will be incorporated as stratification variables during model development and evaluation. Predictive performance and calibration will be reported separately for each stratum. 2.5 Data to Be Collected Ventilator waveform data will be prospectively collected, including pressure, volume, flow, pressure-volume loops, and other relevant waveform information. Ventilator waveforms will be continuously exported through the ventilator's digital interface with high temporal resolution and timestamps. During mechanical ventilation, data will be collected in real time on a daily basis and archived in hourly or event-based segments for subsequent feature extraction and analysis. Electronic medical record data will also be prospectively collected, including demographic characteristics (age, sex, and BMI); preoperative comorbidities (chronic pulmonary disease, cardiovascular disease, diabetes, and hepatic and renal dysfunction); smoking history; preoperative medications (e.g., immunosuppressive agents); surgical type and duration; anesthesia modality; blood loss; blood transfusion volume; APACHE II score; vital signs (HR, BP, SpO₂, and body temperature); respiratory support parameters; blood gas analysis results; chest imaging; laboratory tests (complete blood count, liver and renal function, CRP, PCT, IL-6, and coagulation parameters); fluid input and output; therapeutic interventions (vasoactive agents, antibiotics, sedative and analgesic agents, neuromuscular blocking agents, prone positioning, and ECMO); and postoperative complications (site of infection, microbiological findings, shock, AKI, and other organ dysfunctions). Patients will be followed until 90 days after surgery or hospital discharge/death, with outcomes and complications recorded. Follow-up will be performed daily during the first postoperative week and once weekly thereafter. The primary outcome will be the occurrence of postoperative respiratory failure (PRF). Secondary outcomes will include time to PRF onset, 28-day mortality, 90-day mortality, evolution of disease severity, duration of mechanical ventilation, ICU length of stay, total hospital length of stay, and total hospitalization costs. 2.6 Identification and Selection of Early Warning Features Data sources in this study will include raw ventilator waveforms (pressure, flow, volume, airway pressure-volume loops, etc.), vital signs, baseline clinical information, laboratory tests, imaging data, and treatment records. Early warning features will be selected using the following hierarchical workflow:
• Candidate feature screening: Time-domain and frequency-domain waveform features will be extracted using signal processing techniques. End-to-end waveform features will also be extracted using 1D-CNN or Transformer-based models. Structured clinical variables will include demographic characteristics, comorbidities, surgical duration, blood loss, medication use, postoperative blood gas parameters, inflammatory markers, and other relevant variables.
• Preprocessing and missing-data control: Time-series signals will undergo denoising, normalization, and fixed-window segmentation. Missingness rates of structured variables will be assessed, and variables with >30% missing data will be considered for exclusion. Multiple imputation will be used to address missing data.
• Univariable screening: Each candidate early warning feature will undergo univariable analysis. The preliminary inclusion threshold will be set at p<0.10 or AUC>0.55.
• Regularization and model-driven feature selection: LASSO regression will be used for penalized selection of a large number of candidate features. In parallel, random forest or XGBoost models will be used to calculate feature importance. After cross-validation, a stable set of features will be retained.
• Robustness testing and interpretability filtering: Bootstrap stability analysis will be performed for the selected features. Waveform and clinical features that are clinically interpretable or can be mapped to underlying physiological mechanisms will be prioritized to improve clinical acceptability.
• Determination of the final feature set: The final feature set will be determined by jointly considering statistical significance, contribution to model performance, missingness, and clinical interpretability. The final feature set will be confirmed by a biostatistician using cross-validation in the training set. 2.7 Follow-up Design From postoperative day 0 to day 7, patients will be followed daily, with mechanical ventilation parameters and waveforms, blood gas analyses, vital signs, adverse events (e.g., reintubation, pneumonia, pneumothorax, and ARDS), and therapeutic interventions recorded. From postoperative day 8 until hospital discharge, patients will be followed at least once weekly. After transfer to another department or hospital discharge, telephone or outpatient follow-up will be conducted to record survival status, readmission, recovery of respiratory function, and quality of life on postoperative days 14, 30, and 90. All follow-up forms will be predefined as mandatory fields in the EDC system, and the follow-up method and completion status will be documented. For incomplete follow-up, the reasons for loss to follow-up will be recorded. Pre-specified missing-data handling methods will be applied during statistical analysis, together with sensitivity analyses. Dedicated research nurses or follow-up coordinators will be responsible for follow-up. Follow-up completion rates and key clinical events will be reported at weekly meetings. If the completion rate falls below the predefined target, corrective measures will be initiated. 2.8 Early Warning Model Development This study will use a strategy of parallel validation of deep learning and conventional models. The primary model will be a hybrid neural network integrating 1D-CNN/Transformer and MLP architectures. LASSO, Cox regression, XGBoost, and LightGBM models will be used as baseline comparators to comprehensively evaluate the predictive performance and interpretability of different algorithms for postoperative respiratory failure. The sample will be divided into training, validation, and test sets at a ratio of 6:2:2. Stratified sampling based on endpoint events and key stratification factors will be performed during dataset splitting to maintain comparable event proportions across datasets. Model performance and interpretability will be evaluated using the AUC, sensitivity, specificity, calibration curves, and decision curves. Model interpretability will be assessed using visualization methods such as SHAP and attention mechanisms. Model robustness will be evaluated using bootstrap confidence intervals, subgroup performance analyses, and error analyses. Finally, the warning threshold will be selected in the validation set based on the Youden index or clinically acceptable false-positive and false-negative rates. The corresponding sensitivity, specificity, and false-positive rate will then be reported in the test set. 2.9 Statistical Methods Statistical analysis will primarily involve survival analysis, including Kaplan-Meier analysis and Cox proportional hazards regression. Model development will use training/validation datasets and cross-validation. Model performance will be evaluated based on the AUC, sensitivity, specificity, and calibration. Multiple imputation will be used to handle missing data, and relevant confounding factors will be adjusted for in multivariable analyses.

1. Names, Sources, Collection Period, Acquisition, Processing, and Destruction of Medical Records/Specimens

The electronic medical records and ventilator waveform data used in this study will be obtained from the electronic medical record system and monitoring equipment of postoperative patients at West China Hospital, Sichuan University. Data will be collected during the study implementation period, from December 2025 to December 2026. Data will be obtained by the research team after approval by the ethics committee and acquisition of informed consent from the patients. Personal identifiers will be removed, and the data will be coded and de-identified for analysis. All original data and specimens will be retained for three years after completion of the study in accordance with the hospital's research management regulations. Upon expiration of the retention period, the data and specimens will be destroyed after approval by the ethics committee, thereby ensuring patient privacy and data security.

IV. Quality Control and Quality Assurance

To ensure the scientific rigor, standardization, and reliability of the study, systematic quality control and quality assurance measures will be implemented throughout the study:

• Laboratory testing: All laboratory tests will be performed in qualified medical laboratories in strict accordance with national clinical laboratory standards. Calibrated equipment and standardized reagents will be used, with regular quality control and inter-method comparison testing to ensure the accuracy and reproducibility of laboratory results.
• SOP implementation: All study procedures will be conducted in accordance with established standard operating procedures (SOPs), covering data collection, specimen processing, ventilator waveform image analysis, and data entry, thereby ensuring procedural consistency and traceability.
• Researcher training: Before study initiation, all participating personnel will receive standardized training covering the study protocol, ethical requirements, data security, and SOP implementation to ensure a thorough understanding of and compliance with the study procedures.
• Participant compliance: Preoperative education and postoperative follow-up will be used to enhance participants' understanding of and cooperation with the study, minimize loss to follow-up and missing data, and improve study completeness.
• Data collection, management, and analysis: An electronic data capture (EDC) system will be used, with built-in logic checks and missing-data alerts. Data will be independently entered and verified by two personnel to ensure accuracy. Statistical analyses will be performed by qualified biostatisticians and independently reviewed.
• Study monitoring and supervision: A project monitoring team will be established to regularly monitor study progress, data quality, and compliance with ethical requirements through on-site or remote monitoring. Identified problems will be promptly corrected and communicated to ensure that the study continuously complies with scientific and ethical standards.

Through these multilevel and full-cycle quality control and quality assurance measures, the authenticity, completeness, and reliability of the study data will be ensured, providing a solid foundation for the clinical translation of the study findings.

1. Ethical Principles and Requirements for the Clinical Study

The clinical study will comply with the Declaration of Helsinki of the World Medical Association, the Ethical Review Measures for Biomedical Research Involving Human Subjects issued by the former National Health and Family Planning Commission of the People's Republic of China, and other applicable regulations and requirements.

The study will specifically implement the principles and requirements of informed consent, privacy protection, free participation and compensation, risk control, protection of vulnerable participants, and compensation for research-related injuries.

The clinical study will only be initiated after approval of the study protocol by the Ethics Committee. Before enrollment, the investigator is responsible for providing the participant and/or their legally authorized representative with a complete and comprehensive explanation of the study objectives, procedures, and potential risks. Written informed consent must be obtained before participation.

Participants will be informed that participation in the clinical study is entirely voluntary. They may refuse to participate or withdraw from the study at any stage without discrimination, retaliation, or any adverse impact on their medical care or rights. The informed consent forms will be retained as part of the clinical study documentation for inspection and verification. The privacy of participants and the confidentiality of their personal data will be strictly protected.

VI. Study Timeline

December 2025: Obtain ethics approval and finalize the informed consent form; conduct investigator training; establish the data collection system and standard operating procedures; complete equipment commissioning and laboratory quality-control preparation.

January 2026-September 2026: Systematically collect postoperative patients' ventilator waveform data, clinical information, and laboratory parameters; perform data organization, coding, and specimen processing in accordance with the SOPs; conduct regular data quality checks and monitoring.

October 2026-November 2026: Perform feature extraction and deep clinical phenotype mining based on multidimensional data; develop interpretable AI predictive models and conduct internal validation; optimize model parameters to ensure predictive performance and interpretability.

December 2026: Conduct external data validation and clinical feasibility assessment; prepare the study report and academic manuscripts; archive data and specimens and process or destroy them in accordance with applicable regulations.