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

In this study, we used machine learning techniques using the XGB algorithm to identify variables associated with MINS and generated prediction models. The incidence of MINS, defined by a cTn elevation above the upper reference limit, in patients with a normal preoperative cTn level was 22.0%. The 12 most important variables kept in our prediction models were preoperative cTn level, intraoperative inotropic drug injection, operation duration, emergency operation, type of operation, age, high-risk surgery, body mass index, chronic kidney disease, disease Coronary artery, red blood during the operation. Cell transfers and current alcohol use. We constructed two models according to the number of variables, and the prediction models 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 cTn monitoring after surgery, but difficulties remain in predicting the likelihood of developing MINS.2And3And4And56. In this study, we included patients with a normal cTn level before and after surgery to exclude patients with chronic cTn elevation. Two separate mechanisms are involved in the development of MINS. Although the mismatch of oxygen supply and demand outweighs thrombosis, risk factors for both mechanisms must be considered in the development of MINS.12. In addition, there are often non-ischaemic causes that contribute to elevated cTn in the perioperative period, which complicates the prediction of MINS.13. Machine learning may be a suitable tool for interpreting interactive data from electronic hospital records and transforming them into knowledge10. In this study, we curated real-world data directly from electronic hospital records of successive patients who underwent non-cardiac surgery with a normal preoperative cTn level and investigated the effects of variables on the postoperative cTn elevation. We applied machine learning techniques using the XGB algorithm, which is known as the best performing algorithm14. In our previous study, we compared the performance of different machine learning algorithms for predicting patients with post-MINS mortality, and it was found that XGB is the best performing algorithm.15.

One of the problems in interpreting the results of machine learning techniques is that the causal inference of the observational data is not resolved16. In other words, predictors from machine learning techniques are not necessarily causes of an event16. However, the variables selected for our predictive model showed clinical significance. According to our result, preoperative cTn showed the greatest effect on MINS, although we only included patients with preoperative cTn level within the normal range. In the perioperative period, a cTn level even within the normal range has been reported to be associated with outcome17. Current guidelines do not provide a clear recommendation for preoperative cTn measurement2And3And4And56and only guideline from the Canadian community indicates the need for a baseline cTn level3. Our model confirms that preoperative cTn level may need to be measured in high-risk patients. Several variables in our model reflected the myocardial burden from surgical procedures such as intraoperative inotropic drug use, emergency operation or procedure duration. The need for intraoperative inotropic infusions and red blood cell transfusions may also be associated with hypotension or anemia, which is associated with a higher risk of MINS18And1920. In addition, blood transfusions themselves can be an added burden21, 22. On the other hand, this may also be due to pre-existing anemia, and this needs further investigation. Regarding the types of surgery, there was no case that required intraoperative cardiopulmonary bypass. A higher incidence of MINS has been reported in thoracic surgery where the pericardium has been manipulated based on the extent of the lung resection23A similar result was observed in our model.

Our models also retained risk factors known from patient characteristics such as age and previous history of cardiovascular disease. Postoperative cTn monitoring is recommended for patients over 45 years of age as per expert opinion24and the cost of MINS monitoring was attractive per health gain for patients over 65 years of age25. An association with body mass index has also been reported. Although obese individuals are known to have a higher risk of cardiovascular disease and death, the “obesity paradox” of lower mortality in obese patients with cases of MINS and perioperative myocardial injury has been proposed.26, 27.

The strength of our models is that they can be adopted into daily clinical practice after further validation, because the variants are clinically compelling and readily available from routine medical records. For user convenience, we ran multiple models based on fewer retained variables and showed similar predictive values. We also provided estimated cut-off values ​​for each model according to our dataset. However, whether a model with more variables can offer superior predictive value and an optimal cut-off value that can be applied globally needs further validation. In addition, the low sensitivity of the model limits its use as a screening test in clinical practice. It seems reasonable to consider this model when excluding low-risk patients rather than selecting high-risk patients, given the high specificity and low sensitivity. This can help spare limited medical resources from patients who are excluded from MINS. In this model, we only included preoperative variables, so they can be used from the preoperative period when applied in clinical practice. Some of our variables were modifiable, but it is unclear whether modifying these variables can reduce the incidence of MINS. An effective method for preventing MINS has not yet been established2,7, and avoiding a limited resource of low-risk patients based on our model could be a good start for early identification and treatment of MINS patients. However, in this study, we evaluated several medications before surgery, but none of them showed a meaningful effect on the incidence of MINS. This is consistent with previous findings where the use of beta-blockers reduced myocardial infarction after surgery but increased the incidence of stroke.28. Other cardiovascular drugs including aspirin, nitrous oxide and clonidine in the preoperative period showed nonsignificant results for the prevention of MINS7.

Our study has several limitations that must be taken into account. First, this study used single center retrospective data, and there is a residual 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 laboratory tests of blood and intraoperative variables that cannot be retained due to lack of data availability may need to be taken into account in future studies. To exclude patients with chronic cTn elevation, we enrolled those with an available cTn level preoperatively, and several patients were excluded due to the lack of a preoperative cTn level. Furthermore, perioperative cTn was measured selectively, so the incidence of MINS may be overestimated, and there may be patients who should have been evaluated with cTn but were not. Furthermore, postoperative cTn was not systematically monitored. There may be patients who have lost cTn while monitoring, and the gradient correlation cannot be assessed. In addition, our study was performed on cTnI, and results may have differed according to the cTn assay. Therefore, for our model to be generalizable, it needs further internal and external validation, especially in patients whose cTn was measured routinely. In addition, the definition of a non-ischemic cause of cTn elevation was strictly applied due to the retrospective nature of the study, and this may have caused selection bias. In further study, different models may need to be developed according to the types of surgery and emergency procedures. In addition, our study population showed a relatively high mortality rate, because they were high-risk patients whose cTn was measured in both the preoperative and postoperative periods. This may also have caused selection bias. Finally, perioperative management was not well controlled. Although we followed the institutional protocol based on current guidelines, this may have been updated during the study period. Despite these limitations, this is the first study to demonstrate predictive models for MINS based on risk factors identified by machine learning techniques.

Leave a Comment