Myocardial damage prediction model after non-cardiac surgery using machine learning

In this study, machine learning techniques using the XGB algorithm were used to identify variables associated with MINS and create predictive models. The incidence of MINS, defined by elevated cTn above the upper limit of reference, was 22.0% in patients with normal cTn levels preoperatively. The top 12 variables retained in the prediction model were preoperative cTn level, intraoperative inotropic infusion, duration of surgery, emergency surgery, type of surgery, age, high-risk surgery, body mass index, chronic kidney disease, coronary artery disease, intraoperative was red blood cells. Cell transfusions, and current alcohol use. We created two models according to the number of variables, and the predictive model achieved an AUROC of 0.78 (95% CI 0.77–0.78) for the 12-variable model and 0.77 (95% CI 0.77–0.78) for the 6-variable model.

Current guidelines recommend selective monitoring of cTn postoperatively, but it remains difficult to predict the likelihood of MINS.^{2,3,Four,Five,6}This study included patients for whom preoperative and postoperative cTn levels were available to exclude patients with chronic cTn elevation. Two distinct mechanisms are involved in the development of MINS.Mismatches of oxygen supply and demand are more common than thrombosis, but risk factors for both mechanisms should be considered in the development of MINS¹²Moreover, non-ischemic causes contributing to cTn elevation are frequently found in the perioperative period, complicating the prediction of MINS.¹³Machine learning may be the right tool for interpreting interactive data from hospital electronic records and transforming them into knowledge.^TenIn this study, we curated real-world data directly from electronic hospital records of consecutive patients undergoing noncardiac surgery with normal cTn levels preoperatively to explore the effects of variables on postoperative cTn elevations. . We applied machine learning techniques to the XGB algorithm, which is known to be the best performing algorithm.¹⁴A previous study compared the performance of various machine learning algorithms for predicting mortality after MINS and showed that XGB was the best performing algorithm.¹⁵.

One of the problems in interpreting the results of machine learning techniques is that causal inference of observed data has not been resolved.¹⁶In other words, predictors from machine learning techniques are not necessarily the cause of the event¹⁶However, the variables selected for the predictive model showed clinical relevance. Our results showed that preoperative cTn had the greatest effect on her MINS, even though we only included patients with preoperative cTn levels within the normal range. In the perioperative period, even cTn levels within the normal range have been reported to be associated with outcome.¹⁷Current guidelines do not provide clear recommendations for preoperative cTn measurement.^{2,3,Four,Five,6}and only the Canadian Society guidelines mention the need for baseline cTn levels.³Our model supports that preoperative cTn levels may need to be measured in high-risk patients. Numerous variables in our model reflected myocardial strain due to surgical procedures, such as intraoperative inotropic drug use, emergency surgery, and operative time. The need for intraoperative inotropic infusions and red blood cell transfusions may also be associated with hypotension or anemia, increasing the risk of MINS.^18,19,20Additionally, the blood transfusion itself can be an additional burden^{twenty one,twenty two}On the other hand, this could also be due to pre-existing anemia, which needs further investigation. A higher incidence of MINS has been reported in thoracic surgery in which the pericardium was manipulated based on the extent of pneumonectomy.^{twenty three}and similar results were observed in our model.

Our model also retained known risk factors from patient characteristics such as age and history of cardiovascular disease.In expert opinion, postoperative monitoring of cTn was recommended for patients aged 45 years and older^{twenty four}and the cost of monitoring MINS was attractive per health promotion for patients aged 65 years and older^{twenty five}An association with body mass index was also reported. Although obese individuals are known to have an increased risk of cardiovascular disease and mortality, the ‘obesity paradox’ of lower mortality in mildly obese patients has been suggested for MINS and perioperative myocardial injury. increase.^26,27.

A strength of our model is that the variables are clinically plausible and readily available from routine medical records, thus potentially being adopted into routine clinical practice after further validation. For user convenience, we have provided multiple models based on a small number of retained variables and showed similar predictive values. We also provided estimated cutoff values ​​for each model according to the dataset. However, whether a model with more variables can provide good predictive value and whether it can provide optimal universally applicable cutoff values ​​needs further validation. Furthermore, the model’s low sensitivity limits its use as a screening test in clinical practice. Given its high specificity and low sensitivity, it seems reasonable to consider this model in excluding low-risk patients rather than selecting high-risk patients. This may help save limited medical resources from patients excluded from MINS. In this model, we included only preoperative variables, so it can be used preoperatively in clinical applications. Some variables were modifiable, but it is unknown if altering these variables would reduce her incidence of MINS. No effective preventive method for MINS has been established yet.^2,7, and saving limited resources from low-risk patients based on our model, may be a good starting point for early detection and treatment of patients with MINS. Preoperative medications were evaluated and none showed a significant effect on the development of MNS. This is consistent with previous findings that beta-blocker use reduced postoperative myocardial infarction but increased the incidence of stroke.²⁸Other cardiovascular agents such as preoperative aspirin, nitrous oxide, and clonidine showed nonsignificant results in preventing MINS.⁷.

Our study has several limitations that must be considered. First, the study used retrospective data from a single center, leaving the risk of confounding effects of unmeasured factors. Our analysis lacked detailed cardiac assessments such as echocardiography, as not all patients had such data. Preoperative results of other blood tests and intraoperative variables that could not be retained due to unavailability of data may need to be considered in future studies. To exclude patients with chronic cTn elevation, we enrolled patients with available preoperative cTn levels and excluded a large number of patients due to lack of preoperative cTn levels. Also, because the perioperative cTn was selectively measured, the incidence of MINS may have been overestimated, and patients who should have been evaluated with cTn may have been included. Additionally, postoperative cTn was not systemically monitored. There may have been some patients lost during cTn monitoring and we were unable to assess stepwise association. Furthermore, our study was performed between cTn I and results may vary between cTn assays. Further internal and external validation is required in clinically measured patients. Moreover, the definition of non-ischemic causes of cTn elevation was strictly applied due to the retrospective nature of the study, which may have caused selection bias. Further studies may need to develop different models, depending on the type of surgery or emergency procedure. In addition, our study population was high-risk patients with both pre- and postoperative cTn measurements, and therefore had a relatively high mortality rate. This may have also caused selection bias. Finally, perioperative management was poorly managed. Institutional protocol was followed based on current guidelines, which may have been updated during the study period. Despite these limitations, this is the first study to demonstrate a predictive model for MINS based on risk factors identified by machine learning techniques.