A Machine Learning Prediction Model for Postoperative Acute Kidney Injury in Non-Cardiac Surgery Patients

Lanyue Zhu

Status: Enrolling

Conditions: Kidney Injury, Acute

Treatments: Other: No intervention measures were used.

Study type: Observational

Funder types: Other

Identifiers

NCT07030166

2025ZDSYLL200-P01

Details and patient eligibility

About

Primary objectives of this study is to develop and validate a predictive model for acute kidney injury after non-cardiac surgery based on machine learning. Secondary objectives of this study is to incorporate frailty assessment as a new predictor into the model and measure its incremental value was measured.

Full description

The data of this study are divided into two parts: retrospective and prospective. The retrospective data were from the electronic medical records of adult patients who underwent non-cardiac surgery during hospitalization from July 2015 to June 2025. The ratio of the training set, the internal validation set and the test set was 7:1:2. The prospective data is an external (temporal) validation set. Data collection began in July 2025 and is expected to end in February 2026

Enrollment: 2,500 estimated patients

Sex: All

Ages: 18+ years old

Volunteers: No Healthy Volunteers

Inclusion criteria

• 18 years old or above
• Undergo non-cardiac surgery

Exclusion criteria

• At least one measurement of serum creatinine (SCr) was not conducted before and after the operation
• End-stage renal disease (ESRD) that has received dialysis within the past year
• Baseline SCr ≥ 4.5 mg/dl (because the clinical criteria for AKI based on elevated SCr may not be applicable to these patients)
• Acute kidney injury occurred within 7 days before the operation
• The operation time is less than 2 hours

Trial design

2,500 participants in 3 patient groups

Development group

Description:

The development group is used for fitting the model and optimizing the model. We used preoperative demographic characteristics (gender, age, BMI, marital status, and occupation, etc.), laboratory indicators (blood and urine routine, liver and kidney function, coagulation function and other blood test indicators), preoperative comorbidities and surgical information (surgical department, surgical grade, ASA grade, operation time, anesthesia method, intraoperative position, intake and output volume, vital signs, and intraoperative medication, etc.) Variables such as logistic regression, extreme gradient boosting, decision tree, random forest and Bayesian are used for screening, and multiple methods such as machine learning are employed for modeling.

Treatment:

Other: No intervention measures were used.

Testing group

Description:

The testing set is used for the initial performance evaluation of the model. We use indicators such as discrimination and calibration for model comparison and optimization to select the best model.

Treatment:

Other: No intervention measures were used.

External (time) validation group

Description:

The external (time) validation group is used for future generalization ability assessment. We prospectively collected patient-related data. In addition to the same variables as those in the development group and the testing group, we also evaluated and collected the frailty status of patients before the operation, and recorded prognostic indicators such as the incidence of in-hospital complications, in-hospital mortality, length of hospital stay and hospitalization cost of patients. We used the data from the external (time) validation group to validate the model performance, incorporated the frailty assessment as a new predictor into the model, calculated the incremental values and evaluated the performance of the updated model.

Treatment:

Other: No intervention measures were used.