The results of our study indicate, that our population and patients' management constantly evolves. Our most recent patients are older, have more co-morbidities, and are more frequently undergoing off-pump surgery. In a most recent cohort, significant predictors for prolonged ventilation were aortic aneurysm surgery, combined procedures, valve procedures, preoperative renal dysfunction and preoperative stroke.
Cardiothoracic centres perform a considerable number of highly repetitive procedures. Together with a well organized local database, this may provide a solid basis for the application of various predictive models for all types of postoperative complications (neurological, respiratory, cardiovascular, infectious) or death. The literature abounds in various prediction models for postoperative complications, but their analysis may be quite confusing.
Prolonged ventilation is variably defined. It may be classified as postoperative ventilation continued for more than 12 hours [14], 24 hours [15, 16], 48 hours [17], 72 hours [18], or even 96 hours after the operation [19]. Not surprisingly then, the results obtained may be contradictory and of very limited practical applicability.
What is more, authors tend to narrow down the groups of subjects, creating predictive models for prolonged ventilation in patients after one type of procedure only - for example, after coronary revascularisation with cardiopulmonary bypass [14, 15], adult valve procedures [16, 20],or aortic aneurysm surgery [17]. The only benefit of this approach is the increased homogeneity of the studied group.
Mean extubation times were not the focus of this study. The primary objective was to obtain a practical possibility of predicting a successful execution of a theatre plan taking into account a limited number of intensive beds. It is well known that a patient whose ventilation takes longer than 48 hours, will constitute not only a medical, but also an organizational challenge for the department.
Originally, our intention was to develop a prediction model based on a large population of patients operated on between 2007 and 2010, on the basis of approximately 6,000 patients. Quite surprisingly, however, it turned out that the percentage of patients with prolonged ventilation (>48 hours), which for many years (2004-2008) remained at a fairly constant level (approximately 5%), suddenly in 2009 dropped sharply by half (to 2.5%), reaching values lower than those reported by authors of comparable reports [21]. More interestingly, the characteristics of patients also changed - in a more recent period, there were more elderly patients, with many co-morbitities. The surgical technique also changed - in a subgroup of patients undergoing coronary revascularization, an almost two-fold increase was observed in the percentage of procedures performed without cardiopulmonary bypass (from 37% to 68%). Finally, also in 2009, we changed the technique of postoperative ventilation, introducing a lung protective strategy with higher end-expiratory pressures (PEEP) and smaller tidal volumes for all operated patients, although this method is not aimed primarily at patients with healthy lungs [22]. Tidal volumes were reduced from 7-8 ml/kg to 5-6 ml/kg, whilst PEEP levels were increased from 0-3 cm H2O to 6-8 cm H2O. Additionally, patients started to be extubated directly from the ventilator, without a spontaneous breathing trial.
If we had analyzed in detail the differences between the periods of 2007-2008 and 2009-2010, we would probably have identified an even greater number of more or less important differences. This is not surprising, as this phenomenon occurs in every department - medical care continually evolves, changes are made, and the team climbs higher and higher in their skills as per the learning curve principle. The consequence of all these changes, however, is that the prediction models developed for a specific period of time may prove out of date, as was the case with prolonged ventilation. The model developed in 2007-2008 was not very useful for the population of patients operated on in 2009 and 2010.
Comparing the prediction models from 2007-2008 and 2009-2010, we found some very interesting relationships. It turned out that the most important predictors (in order - aortic aneurysm surgery, emergency surgery, complex procedures and valve procedures) remained in their places in the "ranking" of prediction, still significantly increasing the risk of prolonged ventilation. Some of the previous predictors, however, were missing - namely, urgent operation, advanced age (>65 years), presence of congestive heart failure (NYHA III or IV) and coronary revascularization with cardiopulmonary bypass.
The absence of these predictors in the new model may be logically explained. We have probably advanced more in the management of sick, elderly patients with overt heart failure. The percentage of urgent patients has increased from 15.6% to 28.5%, so urgent surgery has become the rule rather than the exception. Coronary revascularization with cardiopulmonary bypass surgery (CABG technique) was once reserved for more difficult coronary patients, whilst more straightforward cases tended to be scheduled for OPCAB. Today, the trend has reversed and the OPCAB technique is now considered suitable for elderly patients with various co-morbidities [23].
In our department we are interested in a realistic prediction of postoperative complications. The ultimate goal is to identify predictors and to apply them in practice - otherwise, it becomes art for art's sake. In accordance with these principles, our department has already introduced prediction models for renal replacement therapy, permanent neurological complications and prolonged ventilation (>48 hours). Each model was developed with a different method.
The primary factor in determining how to conduct a given analysis is always an audit revealing the incidence of complications over the previous few years. Our audit showed that the fractions of patients undergoing renal replacement therapy, and patients who experienced permanent neurological complications, were fairly constant and averaged between 2.5% and 2.7% respectively, whilst the proportion of patients with prolonged ventilation decreased suddenly in the last two years (from 5% to 2.5%).
For the prediction of renal replacement therapy, we re-introduced the 2007 model proposed by Wijeysundera et al., originally encompassing 20,131 Canadian patients [24]. When in 2008, it was concluded that the model worked very well also for the Polish population, it was introduced in our department. The model appears to be reliable, as the incidence of renal replacement therapy remains constant and, additionally, the demographic data of our patients seem to drift slowly towards the Canadian cohort [25].
For the prediction of permanent neurological complications, however, we decided to develop our own model, based on the analysis of our large cohort of 6,016 consecutive patients [26]. We were forced to ignore the existing models proposed in the literature, as they were inconsistent and resulted in conflicting results when tested on our population. We plan to use our model for at least another two years, as the incidence of permanent neurological injury in our department is also fairly constant. Both the above-presented models are likely to be subjected to routine verification in a couple of years, unless the data from internal audits indicate that the prevalence of any of these complications has suddenly changed.
Therefore, in the prediction of prolonged ventilation, we had to choose another solution. We decided to introduce a model based on an analysis of data derived only from the 2009-2010 period (due to the recent sudden decline in the proportion of patients with prolonged ventilation).
What is going to happen next? Does this mean that over the next few years, in predicting prolonged ventilation, we will be using the model based on the analysis of 2,192 patients only?
We do not know yet; the answer depends on the situation! In 2012 we plan to check whether the proportion of patients requiring prolonged ventilation has set on a new, lower level or still continues to change. If this proportion proves to be permanent - we will perform prospective validation of this model in a most recent population. If satisfactory, we will recalculate the data for a larger population (and thus for the entire period from 2008 to 2012) and to make the necessary corrections (which are likely to be minor). Otherwise, we will have to develop another model based on the previous two years ... unless, in the meantime, the medical literature offers us a better solution to this problem.
An important limitation of our study is in the retrospective data extraction, however a lot of work is carried out to confirm the high quality of the hospital database. Another limitation of the study was the exclusion of all patients in the categories "transplant" and "other", but it is always very difficult to create predictive models on the basis of such non-homogenous groups.
The results of our study indicate, that prediction game is particularly prone to bias and misinterpretation and prediction models should be always approached with caution. Moreover, even centre-specific, departmental prediction models should be updated, when major changes are noted in patients' demographics or management.