背景：随着我国老龄化程度的加深，老年外科手术患者数量不断攀升，其手术安全也得到广泛关注。手术并发症是影响手术安全的重要方面。但在所有可能的术后并发症的影响因素中，输血的效应仍存在争议。具体来说，在非老年人群中，流行病学的“黄金标准”随机对照试验并没有发现输血在特定血红蛋白阈值范围内（如8~10g/dL，即是否应输血的灰色区间）是否影响患者死亡率和并发症发生率，即从节约用血角度支持限制性输血策略。这些证据大多数集中在骨科和心脏手术患者。然而，在更大范围手术患者中开展的许多观察性研究发现，输血会增加患者围术期不良结局的风险。面对这种输血悖论，迫切需要找出两类研究设计下研究结果产生分歧的原因，发展弥合分歧的方法，推动老年人群相关证据的产生以指导输血实践。 此外，准确预测手术并发症风险增加的老年患者，对指导临床医生对高危老年患者提前预防及减少患者不良结局发病率和死亡率具有重要意义。目前已有根据人口学特征及部分术前因素建立的手术风险预测模型，但这些模型纳入因素、结局定义不一致，且大多数只囊括术前危险因素，然而术后并发症风险是由术前、术中、术后三个方面共同影响的，牵一发而动全身，并且传统模型未能考虑预测变量之间的复杂相互作用，不受变量分布形态及变量间关系影响的机器学习方法则能很好克服这个缺点，既往研究广泛表明机器学习的预测准确度较高，结合SHAP值对机器学习进行解释更将充分发挥机器学习方法在临床预测方面的优势与潜力。

方法：本研究首先基于一种新的观察性研究设计来解决上述分歧，命名为基于血红蛋白的输血研究设计。我们将研究人群限制在血红蛋白稳定在7.5g/dL~9.5g/dL范围内的老年（≥60岁）手术患者，以此排除血红蛋白高值（>9.5g/dL）下的不合理输血、大出血（出血量≥500mL）及严重贫血（<7.5g/dL）等容易造成严重混杂的因素，进而利用倾向性评分匹配美国麻醉医师协会（ASA）评分（综合反映患者术前状态）、手术时长等关键协变量，之后利用Logistic回归控制倾向性评分未能控制的剩余变量，从而探究输血对患者术后不良结局的影响。不良结局定义为死亡（住院或出院后30天内）和住院期间可能的并发症,包括缺血性事件（心肌梗死、中风和急性肾功能衰竭）；感染（手术部位感染、肺炎、败血症、感染性休克和尿路感染）；以及其他（需要心肺复苏的心脏骤停、心力衰竭、再次插管、术后≥48小时的机械通气、肺不张、呼吸衰竭、伤口裂开、切口愈合延迟、肺栓塞、静脉血栓形成和多器官功能障碍综合征）。进一步利用“生物医学研究中机器学习预测模型的开发和报告指南：多学科观点”推荐的随机森林及XGBoost模型，并补充传统的Logistic模型作对照，比较模型之间AUC值并利用SHAP值的方法对机器学习模型进行解释。

结果：本研究共纳入6141名普外科手术患者，其中662名（10.78%）的患者接受了红细胞输注。输血组和非输血组异质性较大，特别是在术中大出血（输血 vs. 未输血：37.9% vs. 2.1%）和低血红蛋白（输血 vs. 未输血：29.7% vs. 22.6%）两个指标上。通过排除基础人群中大出血患者，将血红蛋白水平限制在7.5~9.5g/dL（即研究人群，n=715），并运用倾向性评分匹配的方法后，输血组与未输血组的异质性大大降低（关键变量的标准化均值差均小于10%）。同时，输血和不良结局之间的关联在患者异质性降低的过程中也发生了质的转变，基础人群中，输血与术后不良结局相关（OR：2.68，95%CI：[1.86, 3.88]）；而在研究人群中，输血与不良结局没有关联（0.77，[0.32, 1.86]），进一步用倾向性评分匹配后输血与不良结局仍无关联（0.66，[0.23, 1.89]）。本研究发现在7.5~9.5g/dL的研究范围内，输血对老年患者术后不良结局可能具有保护作用，但结果尚不具备统计学显著性。此外，结果显示随机森林模型（AUC: 0.756, 95%CI：[0.701, 0.810]）的预测性能最好，尽管和Logistic回归模型（0.750, [0.697, 0.803]）相比差异没有统计学意义(Delong检验P=0.466)。模型校准曲线也显示模型的校准度较优。ASA≥Ⅲ、入住ICU、血清白蛋白较低、72小时内输血量超过4单位和长手术时间是对术后不良结局影响最大的前5位风险因素。

结论：本研究结果与为数不多的涉及老年手术患者的随机对照试验结果一致，为普通外科老年人输血安全提供了新的参考依据。与此同时，研究结果也提示在基于复杂的观察性临床数据开展关联性研究时，应充分重视研究设计而非单纯的统计校正方法。最后，本研究建立了一个基于围术期风险因素的术后不良结局的预测模型，为真实世界条件下的临床研究提供了新的利于解释的预测方法。

Background: With the deepening of China's aging, the number of elderly surgical patients continues to rise, and their surgical safety has also received widespread attention. Surgical complications are an important aspect that affects the safety of surgery. However, the effect of blood transfusion remains controversial among all possible factors contributing to postoperative complications. Specifically, in the non-elderly population, the "gold standard" randomized controlled trial of epidemiology did not find whether blood transfusion within a specific hemoglobin threshold range (such as 8~10g/dL, that is, the gray interval of whether blood should be transfused) affected the mortality and complication rate of patients, that is, to support the restrictive transfusion strategy from the perspective of blood conservation. Most of this evidence focuses on orthopedic and cardiac surgery patients. However, many observational studies conducted in patients undergoing more extensive surgery have found that blood transfusions increase the risk of adverse perioperative outcomes in patients. In the face of this paradox of blood transfusion, there is an urgent need to identify the reasons for the divergence between the results of the two types of study designs, develop methods to bridge the gap, and promote the generation of relevant evidence in the elderly population to guide the practice of blood transfusion. In addition, accurately predicting elderly patients at increased risk of surgical complications is of great significance to guide clinicians to prevent and reduce adverse morbidity and mortality in high-risk elderly patients. At present, there are surgical risk prediction models based on demographic characteristics and some preoperative factors, but the factors and outcome definitions of these models are inconsistent, and most of them only include preoperative risk factors, but the risk of postoperative complications is affected by the three aspects of preoperative, intraoperative and postoperative, involving the whole body, and the traditional model fails to consider the complex interaction between predictors, and machine learning methods that are not affected by the variable distribution pattern and the relationship between variables can overcome this shortcoming. Previous studies have widely shown that machine learning has high prediction accuracy, and interpreting machine learning combined with SHAP value will give full play to the advantages and potential of machine learning methods in clinical prediction.

Methods: This study is first based on a new observational study design to resolve the above differences, named Hemoglobin-based Transfusion Study Design. We limited the study population to elderly (≥60 years old) patients with stable hemoglobin in the range of 7.5g/dL~9.5g/dL, so as to exclude unreasonable blood transfusion, major bleeding (bleeding volume ≥500 mL) and severe anemia ( <7.5g/dL) and other factors that are prone to serious confounding, and then the propensity score was used to match the American Society of Anesthesiologists (ASA) score (comprehensively reflecting the patient's preoperative state), surgery duration and other key covariates, and then logistic regression was used to control the remaining variables that could not be controlled by the tendency score, so as to explore the impact of blood transfusion on the adverse postoperative outcomes of patients. Adverse outcomes were defined as death (within 30 days of hospitalisation or discharge) and possible complications during hospitalisation, including ischaemic events (myocardial infarction, stroke, and acute renal failure); infections (surgical site infections, pneumonia, sepsis, septic shock and urinary tract infections); and others (cardiac arrest requiring cardiopulmonary resuscitation, heart failure, re-intubation, mechanical ventilation ≥ 48 hours postoperatively, atelectasis, respiratory failure, wound dehiscence, delayed incision healing, pulmonary embolism, venous thrombosis, and multiple organ dysfunction syndrome). Further use the random forest and XGBoost models recommended by "Guidelines for the Development and Reporting of Machine Learning Predictive Models in Biomedical Research: A Multidisciplinary Perspective", and supplement the traditional Logistic model as a comparison, compare AUC values between models and use SHAP values to explain machine learning models.

Results: A total of 6141 patients undergoing general surgery were included in this study, of which 662 (10.78%) received red blood cell transfusion. Heterogeneity was large in the transfusion and non-transfusion groups, particularly in the measures of intraoperative haemorrhage (transfusion vs. no transfusion: 37.9% vs. 2.1%) and hypohaemoglobin (transfusion vs. non-transfusion: 29.7% vs. 22.6%). By excluding patients with major bleeding in the basal population, limiting the hemoglobin level to 7.5~9.5g/dL (i.e., the study population, n=715), and using the method of propensity score matching, the heterogeneity of the transfusion group and the non-transfused group was greatly reduced (the standardized mean difference of the key variables was less than 10%). At the same time, the association between blood transfusion and adverse outcomes also changed qualitatively in the process of reducing patient heterogeneity, with blood transfusion associated with adverse postoperative outcomes in the basal population (OR: 2.68, 95% CI: [1.86, 3.88]); In the study population, blood transfusion was not associated with adverse outcomes (0.77, [0.32, 1.86]), and there was no association between blood transfusion and adverse outcomes after further matching with propensity scores (0.66, [0.23, 1.89]). This study found that blood transfusion may have a protective effect on adverse postoperative outcomes in elderly patients within the study range of 7.5~9.5g/dL, but the results are not statistically significant. In addition, the results showed that the random forest model (AUC: 0.756, 95% CI: [0.701, 0.810]) had the best predictive performance, although there was no statistically significant difference compared with the logistic regression model (0.750, [0.697, 0.803]) (Delong test P=0.466). The model calibration curve also shows that the model is well calibrated. ASA≥III., ICU admission, low serum albumin, transfusion volume of more than 4 units within 72 hours, and long operative time were the top five risk factors with the greatest impact on adverse postoperative outcomes.

Conclusion: The results of this study are consistent with the results of the few randomized controlled trials involving elderly surgical patients, and provide a new reference for the safety of blood transfusion in the elderly in general surgery. At the same time, the results suggest that when conducting correlation studies based on complex observational clinical data, full attention should be paid to study design rather than purely statistical correction methods. Finally, this study establishes a prediction model for postoperative adverse outcomes based on perioperative risk factors, which provides a new explainable prediction method for clinical studies under real-world conditions.