The Problem Intraoperative hypothermia (IOH) - when body temperature drops below 36 degrees Celsius during surgery - affects up to 61% of patients undergoing thoracoscopic lung cancer surgery. This is notably higher than general surgical procedures.
Serious Consequences IOH is not just a minor discomfort. It leads to coagulation dysfunction, increased blood loss, postoperative delirium, surgical site infections, and cardiovascular complications. It also extends hospital stays and raises ICU admission rates.
Why Lung Cancer Patients Are Especially Vulnerable Patients undergoing thoracoscopic lobectomy face unique challenges: they often have low body mass index, comorbid chronic diseases like diabetes, prolonged exposure to laminar-flow operating room environments, and large volumes of intraoperative irrigation fluids.
The Study Goal Researchers at a tertiary hospital in Wuhan analyzed 717 patients who underwent thoracoscopic lung cancer surgery between 2022 and 2023. They used a machine learning algorithm called random forest (RF) to build a risk prediction model, then used SHAP (SHapley Additive exPlanations) to make the model's decisions understandable to clinicians.
Study Design This was a retrospective analysis of 813 patients initially screened, with 717 meeting inclusion criteria (adults diagnosed with lung cancer, undergoing elective thoracoscopic lobectomy, with complete medical records). The dataset was split 70/30 into training (502 patients) and testing (215 patients) sets.
27 Predictive Factors The researchers identified 27 clinical variables spanning three domains: demographic characteristics (age, BMI, diabetes status, preoperative temperature), biochemical indicators (hemoglobin, platelet count, albumin), and intraoperative monitoring data (surgery duration, anesthesia duration, infusion volume, blood loss).
Temperature Monitoring Core body temperature was measured using a nasopharyngeal temperature probe, recorded every 5 minutes from induction of anesthesia until end of surgery. Operating room temperature was kept constant between 22-24 degrees Celsius to reduce environmental interference.
Random Forest Algorithm The RF algorithm was chosen for its ability to capture complex nonlinear relationships between variables and automatically handle feature interactions - something traditional logistic regression cannot do. Bootstrap validation with 1000 replicates was used for internal validation.
How Well the Model Performed The random forest model achieved an AUC (area under the ROC curve) of 0.753, meaning it correctly distinguished hypothermia from non-hypothermia cases about 75% of the time. The recall rate of 0.87 means it correctly identified 87% of patients who would develop hypothermia - crucial for a tool meant to catch high-risk patients.
Clinical Utility Confirmed Decision curve analysis showed positive net benefit across a threshold probability range of 0.05 to 0.95, confirming the model provides real clinical value compared to treating everyone or treating no one.
The Six Most Important Variables Using SHAP analysis, the six most influential predictors of IOH were: (1) intraoperative infusion volume, (2) surgery duration, (3) patient age, (4) anesthesia duration, (5) body mass index (BMI), and (6) hemoglobin level.
Interaction Effects SHAP revealed important interactions: age and BMI interact such that elderly patients with low BMI face sharply elevated risk. When age exceeds 60 years, the SHAP value increases exponentially. Hemoglobin interacts with anesthesia duration - longer procedures amplify the temperature risk associated with low hemoglobin.
Intraoperative Fluid Volume Patients who received more than 1500 mL of intravenous fluid had significantly higher IOH risk. This is mechanistically explained by the fact that every 1000 mL of unwarmed fluid can drop core body temperature by 0.25 degrees Celsius. Warming IV fluids to 36-37 degrees is now recommended.
Surgery and Anesthesia Duration Procedures lasting beyond 180 minutes showed sharply increased hypothermia risk. After 120 minutes, temperature drops become significant, and by 180 minutes the risk rises dramatically. General anesthetics impair the brain's temperature regulation center, and longer anesthesia deepens this effect.
Age and BMI Patients over 60 face heightened risk. Low BMI (under 18.5 kg/m2) means less insulating fat and less heat-generating muscle. Interestingly, higher BMI patients have some protection because fat tissue helps maintain core temperature by triggering vasoconstriction. Elderly patients with low BMI represent the highest combined risk group.
Hemoglobin's Underappreciated Role Low hemoglobin reduces oxygen delivery to tissues, which decreases metabolic heat production. This is a new finding - most prior research focused on hemoglobin in postoperative recovery rather than intraoperative temperature management. The model revealed this independent contribution for the first time in thoracoscopic lung surgery.
From Black Box to Transparent Tool Traditional machine learning models are often dismissed in clinical settings because doctors cannot understand why they make a prediction. The SHAP method converts the RF model's output into intuitive visual explanations showing which factors raised or lowered a specific patient's risk.
Personalized Prevention Planning Knowing a patient's specific risk profile enables targeted interventions before surgery begins. For high-risk patients, clinicians can proactively warm IV fluids, use forced-air warming blankets, schedule shorter procedures when possible, and increase monitoring frequency.
Embeddable in Clinical Systems The researchers suggest this prediction algorithm could be embedded directly into anesthesia monitoring equipment, providing real-time alerts to nursing staff when a patient's risk factors indicate high IOH probability.
A New Standard for Risk Assessment Compared to previous nomogram-based models for IOH prediction in thoracoscopic surgery, the RF-SHAP approach captures nonlinear relationships and feature interactions that simpler models miss. The combination of predictive power plus interpretability represents an advance over both pure statistical models and uninterpretable black-box AI.
Single-Center Data All 717 patients came from one hospital in Wuhan, China. Patient populations, surgical techniques, and equipment may differ at other institutions. External validation using large, multicenter datasets is needed before broad clinical deployment.
SHAP Explains Correlation, Not Causation While SHAP identifies which variables matter most, it is fundamentally a correlation tool grounded in game theory. Clinical decisions must still be made in conjunction with biological reasoning and clinical judgment - the model is a decision support tool, not a replacement for expertise.
Deep Learning on the Horizon The current study used a traditional RF algorithm. Future iterations could leverage deep learning, which can extract patterns from unstructured data like electronic medical records, patient lifestyle history, and medical images. Deep learning models integrated with clinical decision support systems could predict hypothermia risk even more precisely.
Prospective Validation Needed The retrospective design means outcomes were analyzed after the fact. A prospective study where the model is actually used to guide clinical decisions - and outcomes are measured - would provide the strongest evidence for real-world benefit.