An Interpretable and Clinically Deployable Machine Learning Model for Predicting Early Postoperative Pneumonia of Brain Tumor: a Multicenter Diagnostic Study
An Interpretable and Clinically Deployable Machine Learning Model for Predicting Early Postoperative Pneumonia of Brain Tumor: a Multicenter Diagnostic Study
Postoperative pneumonia (POP) is a common and serious complication after elective craniotomy for brain tumor resection. POP often develops within the first week after surgery and may lead to prolonged hospitalization, higher medical costs, and increased risk of severe illness. Because symptoms can be subtle in neurosurgical patients, POP may be detected late, limiting timely prevention and treatment.
This study will evaluate whether a machine-learning-based clinical decision support tool can help clinicians identify patients at high risk for POP early and improve perioperative preventive care. The tool uses routinely collected clinical information to estimate an individual patient's POP risk and provides an easy-to-understand explanation of key risk drivers. Based on the predicted risk level (low, moderate, high, or very high), the system suggests standardized preventive care pathways (e.g., perioperative airway management, targeted antibiotic strategies per local practice, and nutritional support), while allowing clinicians to override recommendations at any time.
Participants will be adults undergoing their first elective craniotomy for brain tumor resection at participating neurosurgical centers. The primary outcome is the occurrence of POP within 7 days after surgery, defined using CDC/NHSN criteria. Secondary outcomes include antibiotic use intensity, length of hospital stay, direct medical cost, and clinician decision confidence. Participants will be followed at postoperative days 1, 3, and 7 using electronic medical record review and phone confirmation when needed.
The goal of this study is to determine whether integrating an explainable AI risk prediction tool into routine care can reduce POP and improve the quality and efficiency of perioperative management after brain tumor surgery.
Rationale Postoperative pneumonia (POP) remains a frequent and clinically important complication after elective craniotomy for brain tumor resection, contributing to prolonged hospitalization, increased cost, and worse clinical outcomes. Conventional POP risk assessment is often experience-based or relies on simplified scoring approaches, which may not adequately capture nonlinear interactions among perioperative factors. This study implements an explainable machine-learning (ML) prediction model within routine perioperative workflows and evaluates whether model-assisted care can improve POP prevention and related resource utilization compared with usual care.
Decision support system
An explainable gradient boosting machine (GBM) model is used to estimate individual POP risk from routinely available perioperative variables. Interpretability is provided using SHAP-based explanations at two complementary levels:
Population level: summarizes the most influential predictors and selected interaction patterns to support clinical understanding and model governance.
Patient level: generates an individualized contribution visualization (e.g., waterfall-style), highlighting the main drivers of a specific patient's risk estimate.
The system automatically assigns a risk tier (low, moderate, high, or very high) and links each tier to standardized prevention pathway templates (e.g., airway management optimization, antimicrobial stewardship-consistent strategies per local policy, and nutritional support). The tool does not mandate treatment; clinicians may accept, modify, or override any suggestion.
Evaluation framework The overall project includes retrospective model development/optimization, prospective external validation/calibration, and a pragmatic implementation evaluation. The registered interventional evaluation uses a multicenter, cluster randomized crossover design with monthly alternating periods of model-assisted care versus usual care. Allocation procedures, eligibility criteria, planned enrollment, and endpoint definitions/time windows are specified in the corresponding record modules (Study Design, Arms/Interventions, Outcome Measures, and Eligibility) to avoid duplication in this section.
Implementation and integration The model is deployed as a lightweight web service with unified APIs and data-exchange formats to enable non-disruptive integration with hospital information systems (HIS) and electronic medical records (EMR). A web-based front end and a Python-based back end support RESTful calls and are designed for low-latency inference (target single-prediction latency <200 ms), suitable for perioperative and inpatient workflows.
Data governance and model updating To support long-term generalizability across hospitals and mitigate dataset shift, the project establishes a closed-loop maintenance process ("local de-identification → cloud retraining → model version management → edge deployment"). Model updates are version-controlled and deployed under governance procedures consistent with local regulations, institutional policies, and applicable ethics approvals.
Inclusion Criteria:
Exclusion Criteria: