Study design
This study was conducted in compliance with the Transparent Reporting of Multivariate Predictive Models for Individual Prognosis or Diagnosis (TRIPOD) statement (Additional file: Table S1)14Data from 2286 postoperative elderly ICU patients admitted to Dongyang People's Hospital between 2012 and 2021 were retrospectively analyzed. Inclusion criteria were ICU length of stay >24 hours, first admission to ICU, and age >65 years. Exclusion criteria were patients with hematological malignancy, liver cirrhosis, or previous splenectomy, patients who presented with thrombocytopenia before ICU admission, patients with heparin-induced thrombocytopenia (HIT), and patients with missing data >30%. The publicly available MIMIC database was used for external validation.
This study complied with all relevant local guidelines and regulations and was approved by the Ethics Committee of Dongyang People's Hospital (DRY-2022-YX-155). Because the study was retrospective and observational in nature, the Ethics Committee of Dongyang People's Hospital waived the requirement for informed consent. Data were anonymized before analysis was performed. Author XJ obtained permission to access the MIMIC-III database after completing the online training course “Protection of Human Research Participants” provided by the National Institutes of Health (certification number 7632299).
Data collection
Data collection was conducted using case information mining software from Shanghai LE9 Health Co., Ltd. Data collected included age, sex, Sequential Organ Failure Assessment (SOFA) score, Glasgow Coma Scale score, smoking history, history of alcohol use disorder, comorbidities such as hypertension and diabetes, types of vasoconstrictors, glucocorticoids, and antiplatelet drugs used in the first 24 hours after ICU admission, early biochemical indicators, and maximum, minimum, and mean vital signs at ICU admission.
The primary outcome was in-hospital mortality. Secondary outcomes included duration of mechanical ventilation, ICU stay, length of hospital stay, and hospital costs.
Diagnostic Criteria
Thrombocytopenia was defined as a platelet count <100 × 10 .9/L. Thrombocytopenia was defined as a platelet count <100 × 10 and was classified as mild, moderate, or severe.9/L, < 50 × 109/L, < 20 × 109/L, respectively15,16.
Data Processing
The primary objective of this study was to accurately predict the occurrence of thrombocytopenia, defined as a platelet count <100 × 109We aim to clarify the correlation between risk and outcomes in postoperative elderly ICU patients using a binary classification approach (yes/no) with machine learning models to facilitate direct application in clinical practice and ultimately support timely and effective interventions. A preliminary screening of 87 potentially relevant variables was performed. Data preprocessing was performed using the classification and regression training (CARET) R package. Seven variables with correlation coefficients >0.9 were excluded and the remaining 80 variables were screened using least absolute shrinkage and selection operator regression. The least absolute shrinkage and selection operator regularization parameter lambda was selected through 10-fold cross-validation using the cv.glmnet function, and variables were selected stepwise by controlling the value of lambda.17In addition, random forest method was utilized for variable selection.Finally, eight variables were included: initial platelet count, creatinine, hemoglobin, glutamic oxaloacetic transaminase (AST), minimum and maximum systolic blood pressure at ICU admission, and mean and maximum heart rate within 24 hours of ICU admission.
Interquartile ranges, i.e. the distance between the upper and lower quartiles of the boxplots, were used to detect outliers. Outliers were excluded and treated as missing values. Variables with more than 30% of values missing were removed. Multiple imputation was used to replace missing variable values.18,19Specifically, for each incomplete variable, an imputation model was specified using the other variables as predictors. At each iteration, the variables resulting from the imputation were used in the subsequent imputation model, and this process was repeated until convergence was achieved.
Model Building
First, we trained four machine learning (ML) models with Support Vector Machine (SVM), Bayesian Machine Learning, C5.0 Decision Tree, and eXtreme Gradient Boosting (XGBoost) algorithms using the CARET, XGBoost, C50, e1071, naivebayes, and gbm R packages. Hyperparameters were tuned by grid search. Then, we used the four models to create an integrated model using CaretEnsemble.20Samples were randomly split into training and validation sets in a 7:3 ratio. All ML models were evaluated using 10-fold cross-validation. The MIMIC dataset was used for external validation.
Model validation and evaluation
Model performance was evaluated by the area under the receiver operating characteristic curve (AUC). Model calibration performance was assessed using the calibration curve and confusion matrix was evaluated using accuracy, precision, specificity and recall as parameters with a cut-off point of 0.5.
Model description and application
Model explanation was done using ranked variable importance using the “varImp” function in the CARET R package. Additionally, individual explanations were provided by the Locally Interpretable Model-Agnostic Explanation (LIME) and iBreakdown algorithms.21,22Finally, a web-based calculator based on the optimal C5.0 algorithm model was developed to calculate the probability of thrombocytopenia during an individual patient's ICU stay by inputting eight parameters (initial platelet count at ICU admission, creatinine, hemoglobin, AST, minimum and maximum systolic blood pressure within 24 hours of ICU admission, and average and maximum heart rate) (Additional file: Figure S1). This tool was uploaded to GitHub and made publicly available. Furthermore, to differentiate the severity of thrombocytopenia, the same method was used to screen eight variables: SOFA score, platelet count, arterial pH, serum lactate international normalized ratio, serum sodium value, minimum systolic blood pressure, and norepinephrine usage. A prognostic model was constructed using the SVM algorithm incorporating these determinants. Similarly, a web-based calculator was created based on this model and can be accessed from the supplementary material as shown in Figure S2.
statistics
Descriptive statistics were analyzed using the CBCgrps R package.twenty threeMeasures that followed normal distribution were expressed as x ± s, and measures that did not follow normal distribution were expressed as M(P25, P75). Analysis of variance was used to compare continuous variables, and chi-square test was used to compare categorical variables. All statistical analyses were performed using R software (version 4.1.3 (https://mirrors.tuna.tsinghua.edu.cn/CRAN/)). Statistical significance was expressed as P< 0.05.
