Carbon Dioxide, Blood Pressure, and Stroke: Comment

In their article, Vlisides et al.¹  present retrospective associations between hypotension, hypocarbia, and hypercarbia and clinically apparent postoperative strokes, with 122 used in the analysis. (Covert strokes are nearly 10 times as common,²  but were not considered.) They conclude that there are strong associations between mean arterial pressure (MAP) less than 55 mmHg and end-tidal pressure of carbon dioxide exceeding 45 mmHg. Both outcomes are biologically plausible and might guide clinical care, assuming that the relationships are proven causal in trials. However, some clarification about certain aspects of this report would be helpful, especially given previous analyses with 120 and 1,553 strokes that showed no association between hypotension and stroke.^3,4 

Case-control studies depend critically on correctly identifying case and control populations (avoiding selection bias). Curiously, patients with strokes in a limited population were matched to a much broader population from which there was a less rigorous attempt to identify strokes: “as stroke cases were initially screened and identified before control cases, when there were fewer cases overall in the Multicenter Perioperative Outcomes Group Database.” This is a highly unorthodox approach that leaves the broader population contaminated by an unknown number of unidentified strokes. More seriously, the broader population might differ in important but unidentified ways from the population from which strokes were extracted. A confirmatory sensitivity analysis restricted to the patients from whom strokes were drawn would be reassuring.

The size of the population from which strokes were identified remains unclear, but apparently was about 600,000 patients. The incidence of overt postoperative stroke is reported to be roughly 0.4 to 0.6% in various noncardiac populations.^4,5  Vlisides et al. identified 126 strokes, corresponding to an incidence of 0.02% (assuming 600,000 patients) which is about a factor of 25 less than expected. Consequently, there is substantial potential for selection bias because the identified stroke patients may poorly represent the full population who suffer postoperative strokes. To the extent that the analysis was based on a possibly nonrepresentative sample, the presented associations may be inaccurate.

Table 4 indicates that only 8% of the observed strokes were detected on the day of surgery, and less than 60% within 3 days. There is no plausible biologic mechanism to explain why intraoperative hypotension or hypercarbia would provoke strokes days after surgery. Furthermore, half the strokes were embolic. Again, there is no obvious reason why intraoperative hypotension or hypercarbia would provoke embolic strokes. Residual confounding is the most likely explanation for such biologically implausible associations.

Specifically, there are many factors such as cardiovascular disease that provoke both hypotension and stroke, thereby generating confounded noncausal relationships between one and the other. Vlisides et al. naturally adjusted for observed confounding, but there is always unobserved confounding. Furthermore, even known confounders are often poorly characterized such as only being known dichotomously (e.g., history of cardiovascular disease or not), which precludes sophisticated adjustment based on severity. As an example, the authors note that they had sparse data about β-blocker use—although acute β-blocker use is known to cause perioperative strokes.⁶  Adequacy of confounding adjustment is key to any observational analysis. Observational analyses differ with respect to confounding risk; the risk in this case seems especially high.

There were no significant bivariable associations for amounts of hypotension below a range of thresholds (table 2). For example, area less than 60 mmHg was 15 mmHg-min in stroke patients and 14 mmHg-min in reference patients. The analogous areas were 35 and 33 mmHg-min for area less than 65 mmHg. Consequently, all reported significance is consequent to statistical adjustment. Normally univariable associations are strong, and then adjusted downward in multivariable analyses. It is unusual for statistical significance to result entirely from multivariable adjustments. A corollary is that modest differences in statistical modeling, especially confounder adjustment, might substantially alter the results rather than simply fine-tuning obviously meaningful differences in the underlying data.

The risks of hypotension (and presumably pressure of carbon dioxide) accrue at the extremes. A consequence is that the most interesting conclusions are often based on limited data. For example, in this case, investigators are looking for the “cross” between serious hypotension and strokes, although few patients will have both—especially with only 122 analyzable strokes. Along those lines, it would be helpful to provide a histogram showing the distribution of blood pressures as a guide to the range of reliable risk estimates.

Vlisides et al. present their associations in terms of “an approximately 10 to 17% increased relative risk of stroke per 10 units (mmHg-min)” with MAP 55 mmHg or less. In our experience, across many hospitals, MAPs less than 55 mmHg are uncommon and short-lived because they inevitably treated. It would be helpful to know what fraction of the patients had 10 min with MAP less than 55 mmHg, and how many of them experienced a stroke. The authors also note the adjusted odds of stroke are 2.0 for patients who sustained “10 minutes with a MAP of 50 mmHg and ETco₂ of 28 mmHg.” Again, it would be helpful to know what fraction of patients met this stringent criterion. Assuming the reported control incidence, along with a factor-of-2 risk increase in patients who experienced substantial and sustained hypotension and hypocapnia, the number needed to treat would be 5,000—which does not seem to support the authors’ conclusion that these “risk factors can be modified to reduce the incidence of postoperative stroke.”

Dr. Sessler serves on advisory boards and has equity interests in Perceptive Medical (Newport Beach, California). He is a consultant for Edwards Lifesciences, a company that funds research conducted in the Department of Outcomes Research.