The challenge of postoperative pulmonary complications: Postoperative pulmonary complications (PPCs) - including pneumonia, respiratory failure, prolonged air leak, and pleural effusion - occur in 15-40% of NSCLC patients after thoracoscopic surgery (VATS). They extend hospital stays, increase costs, and worsen survival outcomes.
Current limitations in risk assessment: Surgeons currently assess PPC risk qualitatively based on clinical experience and basic spirometry. No validated, personalized, automated tool existed to provide a quantitative risk estimate that integrates multiple patient-specific factors.
Study approach: This study used 286 NSCLC patients who underwent video-assisted thoracoscopic surgery (VATS) to build and validate an AutoML model incorporating the novel IBSLO (Improved Blood-Sucking Leech Optimizer) metaheuristic for feature selection and hyperparameter tuning.
Outcome: The resulting model achieved AUC-ROC of 0.9038 in training and 0.8775 in testing, with a user-friendly clinical decision support system (CDSS) deployed in MATLAB for bedside use.
What is AutoML: Automated Machine Learning (AutoML) is a framework that automates the process of selecting the best machine learning algorithm and tuning its parameters. Rather than manually testing dozens of models, AutoML searches the parameter space efficiently to find optimal configurations.
The IBSLO algorithm: IBSLO stands for Improved Blood-Sucking Leech Optimizer - a metaheuristic optimization algorithm inspired by the feeding behavior of leeches. It balances exploration (searching new areas of the parameter space) and exploitation (refining promising solutions) to find the globally optimal set of features and model parameters.
Why a novel optimizer matters: Standard AutoML tools like grid search are computationally expensive. IBSLO's bio-inspired search strategy efficiently navigates high-dimensional feature spaces, making it practical for clinical datasets that may have hundreds of potential predictors but limited sample sizes.
Dataset composition: 286 NSCLC patients from a tertiary hospital were included. Features included preoperative laboratory values, pulmonary function tests, surgical approach, demographics, comorbidities, and perioperative parameters. The dataset was split 70:30 for training and testing.
Leukocyte count: Elevated white blood cell count before surgery indicates active inflammation or infection, significantly increasing PPC risk. High leukocyte counts predict impaired healing and infectious complications post-VATS.
Body mass index (BMI): Both underweight (BMI less than 18.5) and obese (BMI greater than 30) patients are at higher risk of PPCs. Obese patients have reduced functional residual capacity and are prone to atelectasis, while underweight patients have depleted respiratory muscle mass.
Surgical approach: The specific VATS approach - single port vs. multi-port, lobectomy vs. segmentectomy - significantly influences PPC risk. More extensive resections and technically challenging procedures carry higher complication rates.
Age, blood loss, and CRP: Older age (particularly above 65) correlates with diminished pulmonary reserve. Intraoperative blood loss greater than 200 mL signals surgical complexity and hemodynamic stress. C-reactive protein (CRP) - an inflammatory marker - predicts healing capacity and infection susceptibility.
Training performance: On the training dataset, the IBSLO-optimized AutoML model achieved AUC-ROC of 0.9038, with accuracy of 87.3%, sensitivity of 85.1%, and specificity of 88.9%. These results indicate strong discriminatory ability.
Test set generalization: On the held-out test set, AUC-ROC was 0.8775 - only modestly lower than training performance, suggesting the model generalizes well without significant overfitting. This is particularly notable given the relatively small dataset of 286 patients.
Comparison to conventional risk scores: The IBSLO-AutoML model outperformed both clinical scoring tools (such as the Thoracoscore) and individual machine learning algorithms (logistic regression AUC 0.78, random forest AUC 0.83) when tested on the same dataset.
Feature importance analysis: SHAP (SHapley Additive exPlanations) analysis confirmed that CRP and leukocyte count dominated model predictions - patients with CRP above 5 mg/L and leukocyte count above 10x10^9/L had more than three-fold higher PPC risk.
CDSS deployment: The model was packaged into a clinical decision support system (CDSS) using MATLAB. Clinicians can input six patient parameters and receive an individualized PPC risk probability - enabling real-time risk assessment at the bedside or in preoperative clinics.
Interpretable outputs: Beyond the risk probability, the CDSS presents a SHAP-based waterfall chart showing which factors are increasing vs. decreasing the patient's risk, and by how much. This transparency helps surgeons understand and trust the model's outputs.
Practical workflow integration: The tool is designed to fit into existing preoperative assessment workflows. Anesthesiologists or surgical nurses can input readily available clinical values, making the tool accessible without specialized data science training.
Patient counseling applications: The individualized risk estimates can be used to inform patient consent discussions - explaining concrete percentage risks rather than qualitative 'high' or 'low' risk categories, improving shared decision-making.
Small sample size: With only 286 patients, the dataset limits the number of features that can be reliably included and the precision of performance estimates. Larger multicenter studies are needed to confirm these findings.
Single-center bias: Surgical practices, patient populations, and complication definitions vary between hospitals. External validation at multiple centers is critical before deploying the CDSS in different healthcare settings.
Static model without temporal updates: As surgical techniques evolve (robotic VATS, new stapling devices), the model needs periodic retraining on updated data. Procedures change over time, and a model trained in 2022 may not reflect 2028 practice.
Expanding outcomes: Future versions should predict not just any pulmonary complication but specific complication types (pneumonia vs. prolonged air leak vs. respiratory failure), each of which has different management implications and risk factors.