Objective To develop a machine learning (ML) model to predict the risk of death in intensive care unit (ICU) patients with chronic obstructive pulmonary disease (COPD), explain the factors related to the risk of death in COPD patients, and solve the "black box" problem of ML model. Methods A total of 8088 patients with severe COPD were selected from the eICU Collaborative Research Database (eICU-CRD). Data within the initial 24 hours of each ICU stay were extracted and randomly divided, with 70% for model training and 30% for model validation. The LASSO regression was deployed for predictor variable selection to avoid overfitting. Five ML models were employed to predict in-hospital mortality. The prediction performance of the ML models was compared with alternative models using the area under curve (AUC), while SHAP (SHapley Additive exPlanations) method was used to explain this random forest (RF) model. Results The RF model performed best among the APACHE IVa scoring system and five ML models with the AUC of 0.830 (95%CI 0.806 - 0.855). The SHAP method detects the top 10 predictors according to the importance ranking and the minimum of non-invasive systolic blood pressure was recognized as the most significant predictor variable. Conclusion Leveraging ML model to capture risk factors and using the SHAP method to interpret the prediction outcome can predict the risk of death of patients early, which helps clinicians make accurate treatment plans and allocate medical resources rationally.
Objective To evaluate the performance of lightweight Chinese large language models (LLMs) in answering specialized lung cancer questions, and to explore the impact of retrieval-augmented generation (RAG) on model performance. Methods Eleven lightweight Chinese LLMs with parameter sizes ranging from 7B to 32B were included. A lung cancer-specific evaluation dataset consisting of 200 questions [100 A1-type (basic knowledge) and 100 A2-type (clinical case) questions], constructed based on clinical guidelines and thoracic surgery textbooks, was used for assessment. Model performance was evaluated under two conditions (with and without RAG). Accuracy was used to assess model performance, and response latency was recorded to reflect inference efficiency. An accuracy–latency scatter plot was constructed for descriptive analysis of overall model performance. Results All models successfully completed the evaluation. With the introduction of RAG, the overall average accuracy improved from 61.68% to 76.36%. Smaller models demonstrated the most significant improvement (e.g., the accuracy of DeepSeek-7B increased from 32.50% to 60.00%, P<0.001). The average response latency increased from 12.58 s to 13.80 s. The Qwen3 series showed the best overall performance, and Qwen3-32B achieved the highest accuracy under both conditions (76.50% and 84.00%, respectively). After RAG integration, performance differences among model families were markedly reduced. Based on the accuracy-latency trade-off, Qwen3-32B achieved the best balance between accuracy and response latency under the baseline condition, whereas Qwen3-14B demonstrated superior overall performance in terms of accuracy, latency, and computational cost after RAG integration. Conclusion The integration of RAG technology improves the ability of lightweight Chinese LLMs to answer specialized lung cancer questions. Under the dual practical constraints of limited computational resources and medical data security requirements, the "lightweight model+RAG" technical framework may represent a promising deployment solution.