PREDICTING ICU ADMISSION AND VENTILATOR DEPENDENCE IN ADVANCED LUNG CANCER PATIENTS USING CLINICAL AND IMAGING DATA

Abstract

Progressive tumour burden, respiratory compromise, treatment-related toxicity, infections, and systemic complications are risk factors for acute clinical deterioration, intensive care unit admission and ventilator dependence in advanced lung cancer patients. The ability to identify patients who are likely to need critical care support early would allow timely intervention in the patient, planning of resources and individualized clinical decision making. This study aims to develop a predictive model for ICU admission and ventilator dependence of advanced lung cancer patients based on a machine learning framework using a combination of clinical and imaging data. Clinical features were used at the patient level: age, cancer stage, performance status, oxygen saturation, respiratory rate, comorbidities, inflammatory markers, treatment history, and laboratory parameters, among others, as well as imaging-derived features such as tumor burden, presence of pleural effusion, lung infiltration, airway obstruction, metastatic involvement, and radiomic texture characteristics. Several supervised learning models such as Logistic Regression, Random Forest, Support Vector Machine, Gradient Boosting, XGBoost and multimodal models (deep learning) were trained and evaluated for the prediction of ICU admission and ventilator requirement. The accuracy, sensitivity, specificity, precision, F1-score and area under the receiver operating characteristic curve were used to assess model performance. The multimodal framework achieved good predictive results, and the proposed ensemble model and deep learning surpass the baseline classifiers. The feature importance analysis revealed that oxygen saturation, respiratory rate, ECOG performance status, tumor burden, presence of pleural effusion, inflammatory markers, previous chemotherapy, and imaging-based lung involvement were the most predictive features. These results indicate that the use of clinical parameters in combination with imaging parameters can be useful in early risk categorization during advanced lung cancer. The proposed framework could help clinicians recognize high-risk patients, optimize triage in the ICU, aid in monitoring the patient's breathing, and better prepare patients for critical care.