In this study on hospitalised patients with COVID-19, we found a statistically significant association between a lower PaCO₂ at admission and increased risk of delirium during the hospital stay after adjusting for hypoxia. The RR remained in the same range and statistically significant when adjusting for the potential confounder inflammation, and when adjusting for predisposing factors for delirium such as age, dementia, frailty, multimorbidity and polypharmacy. The results of subgroup and secondary analyses suggest that the results were not biased by the administration of supplementary oxygen during blood gas sampling, ICU admission and mechanical ventilation, missing PaCO₂ values, or the fact that some patients died in hospital without developing delirium. In conclusion, our results suggest that a lower PaCO₂ contributes to the risk of delirium independently of other well-established risk factors in hospitalised patients with acute hypoxia due to COVID-19. Our findings support our hypothesis that a high HVR increases the risk of delirium in patients with acute hypoxia.

To the best of our knowledge, an association between hypocapnia and delirium in patients with acute hypoxia caused by an infectious disease has not previously been published. Our finding is in line with previous research on acute hypoxia from real and simulated high altitude, where it has been observed that an impairment of cognitive function is associated with both a low PaCO₂ and a high HVR^7,8. Additionally, in patients undergoing surgery, intraoperative hypocapnia has been found to increase the risk of postoperative delirium in a dose-dependent fashion^18,19. In sepsis, which is another condition characterised by hyperventilation, associations between delirium and reduced cerebral blood flow and cerebral oxygenation have been found²⁰.

The RR of delirium associated with a lower PaCO₂ was amplified and reached statistical significance at the 0.05-level only after adjusting for SaO₂. Hypoxia is an independent risk factor for delirium⁴, and obviously, an individual’s PaCO₂ is also affected by their SaO₂ mediated through hyperventilation. Additionally, hypoxia has a vasodilatory effect on cerebral arteries^16,21 and thereby counteracts the vasoconstrictory effect of hypocapnia. However, cerebral ischaemia is possibly augmented if hypocapnic vasoconstriction is combined with hypoxemia. Therefore, SaO₂ is an obvious co-pathophysiological factor and statistical confounder between PaCO₂ and delirium, and it makes sense that the effect of PaCO₂ on delirium risk becomes more evident after SaO₂ is adjusted for in the model.

When investigating the ability of PaCO₂ to correctly classify patients with delirium, we found a relatively low AUC of the ROC curve of 0.62 with a lower bound of the 95% CI just above 0.5. Although our regression analyses provided evidence that PaCO₂ is associated with delirium independently of other well-established risk factors, the AUC indicates that PaCO₂ alone is only marginally better than a random classifier of delirium. This is not surprising, considering the numerous factors that can precipitate delirium and the fact that diagnosing delirium is sometimes very difficult.

One imminent consequence of hyperventilation and hypocapnia is an associated respiratory alkalosis, and we initially hypothesised that the effect of hypocapnia on the risk of delirium could be mediated through a rise in arterial pH in two ways: First, a rise in arterial pH increases the affinity of oxygen to haemoglobin²², possibly causing less oxygen to be released from haemoglobin in the brain. Secondly, hypocapnic cerebral vasoconstriction is probably mediated through a fall in the concentration of H⁺ ions in the extracellular and cerebrospinal fluid²³, and may therefore also be affected by changes in arterial pH. Contrary to our hypothesis, the RR of delirium associated with a lower PaCO₂ increased after adjusting for pH, which suggests that the rise in arterial pH following hypocapnia actually protects against delirium, possibly because the left shift of the oxygen-haemoglobin dissociation curve increases the blood’s ability to transport oxygen to hypoxic brain tissue²². In patients both with and without delirium, median time from symptom debut to hospital admission was more than a week and arterial pH was just at the upper normal range of 7.45, with no differences between the groups. Additionally, patients with delirium had a significantly lower arterial bicarbonate than patients without delirium. It is therefore likely that the hypocapnic patients in our study had hyperventilated for many hours or days before being admitted to hospital, allowing time for the kidneys to compensate for the respiratory alkalosis and attenuate a rise in arterial pH. Our results suggest that a low PaCO₂ increases the risk of delirium independently of arterial pH, and that the risk of delirium may actually increase after pH is normalised through metabolic compensation.

An alternative explanation for the association between PaCO₂ and delirium is that delirium may increase anxiety and hence hyperventilation, which causes hypocapnia. However, this is unlikely for several reasons: first, patients in our study with delirium were significantly more hypoxic at admission, and therefore experienced a stronger stimulus to hyperventilate, than patients without delirium. Secondly, less than half of the patients who experienced delirium during hospital admission had delirium in the emergency department, although all arterial blood gases were sampled at this point. Thirdly, patients with delirium had a higher respiratory rate and lower PaO₂/FiO₂-ratio at admission than patients without delirium. It has previously been found that COVID-19 patients with more vigorous hyperventilation at admission are at increased risk of barotrauma to the lungs²⁴. Barotrauma can exacerbate respiratory failure and trigger excess inflammation, both of which are independent risk factors for delirium. These observations make it very likely that hyperventilation anticipated delirium among the hypocapnic patients in our study.

A possible confounder between PaCO₂ and delirium is sedative medication because the use of sedatives can affect both respiratory drive and delirium risk⁴. Although no patients in the study had been given sedatives in the hospital when arterial blood was sampled at admission, it cannot be ruled out that some patients had been taking sedative medication such as opioids or benzodiazepines before being admitted to hospital. Because sedatives increase delirium risk and reduce respiratory drive, delirious patients affected by sedatives would probably have a higher PaCO₂. Therefore, it is very unlikely that the observed association between a low PaCO₂ and delirium could be explained by the use of sedatives in some patients.

Two observations concerning the regulation of cerebral blood flow may challenge our hypothesis: first, climbs to high altitudes induce a gradual onset of hypoxia and hypocapnia that may resemble the clinical course of COVID-19. While high altitude dwelling can unquestionably affect brain function negatively^8,25, this setting is characterised by a pronounced increase rather than decrease in the cerebral blood flow because hypoxic cerebral vasodilatation more than offsets the effect of hypocapnia²⁶. Secondly, in mechanically hyperventilated patients with elevated intracranial pressure, the effects of hypocapnia on global cerebral blood flow vanes almost completely after a few hours²⁷. In these contexts, it is uncertain if prolonged hypoxia-driven hyperventilation during COVID-19 could cause a sustained cerebral vasoconstriction and cerebral ischemia predisposing to delirium. However, non-sedated hypoxic COVID-19 patients are different from mechanically ventilated patients with elevated intracranial pressure because the latter are usually heavily sedated and have normoxia or even hyperoxia. Additionally, delirium, unlike high altitude dwelling, appears to be characterised by a decrease rather than increase in the cerebral blood flow^20,28, suggesting that different pathophysiological mechanisms affect the cerebral blood flow. It has also been observed during hypoxia that there are marked regional differences in the blood flow to the brain, with the brainstem being more responsive to hypoxic vasodilatation than the cortex²⁹. Finally, the hypoxia-mediated increase in cerebral blood flow is reduced in persons with a high HVR because of more pronounced hypocapnic cerebral vasoconstriction³⁰. In conclusion, it appears as if the combination of hypocapnia and hypoxemia in persons with a high HVR can cause a sustained tissue hypoxia in cortical areas of the brain, and that the blood flow and oxygen supply to these areas is not necessarily adequate although the global cerebral blood flow is normal or elevated.

This study has some limitations. First, delirium was diagnosed retrospectively based on available patient records, and not bedside. Although the majority (69%) of patients retrospectively found to have delirium also had delirium explicitly diagnosed in their patient records, it is well-known that delirium in hospitalised patients is underdiagnosed³¹. It is likely that some cases of delirium were missed because of insufficient descriptions of mental status and behaviour in the patient record. Additionally, descriptions of the patient may have been misinterpreted, so that patients with dementia or intellectual disabilities, or patients receiving sedatives were wrongly diagnosed with delirium.

Secondly, the generalisability of our findings may be limited because of the study’s single-centre design. Patients were included during the first pandemic wave in 2020. At that time, the treatment of COVID-19 was much debated, particularly the role of antiviral and anti-inflammatory drugs and when to initiate mechanical ventilation. Additionally, isolation of hospitalised patients with COVID-19 probably restricted access to delirium-preventive measures such as physiotherapy and contact with family members. In particular, respiratory physiotherapy appears to be important to improve gas exchange and prevent delirium in ICU patients with COVID-19³². It is possible that patients admitted to other hospitals at the time received different treatment and delirium-preventive measures, which could have affected their delirium risk.

Thirdly, this study may be limited by a high number of missing observations in two important variables: PaCO₂ and ferritin. Although the significant association between PaCO₂ and delirium persisted when the missing PaCO₂ values were accounted for with multiple imputation, that analysis generally requires that data are missing at random, which may not be the case. Patients where PaCO₂ was not measured did not have a significantly different risk of delirium compared to patients where PaCO₂ was measured, but they did have less severe respiratory failure. Therefore, we suspect that inclusion of patients to our primary analyses was biased towards more severe disease and hence a higher delirium risk. It is possible that this bias persisted even after imputation of missing PaCO₂ values.

Regarding the inflammatory marker ferritin, there appears to be a clear association between increasing inflammation and delirium risk³³. It is possible that this inflammation also increases the tendency to hyperventilate, which makes inflammation a statistical confounder between hypocapnia and delirium (Fig. 1). When adjusting for inflammation, we chose to include both CRP and ferritin, which were the only two inflammatory markers measured in our study. It has previously been found that ferritin but not CRP is associated with the presence of delirium during COVID-19³⁴. Additionally, in patients admitted to ICU, CRP is not associated with delirium when adjusting for other variables³⁵. Therefore, ferritin appears to be a better biomarker of inflammation anticipating delirium than CRP. Nevertheless, we chose to include both ferritin and CRP in our model because ferritin was missing in almost half of patients. We suspect that patients with severe disease were more likely to have an extensive analysis of inflammatory markers including ferritin in their blood samples, possibly introducing a bias. Although our results showed that the RR of delirium associated with a lower PaCO₂ remained statistically significant after adjusting for CRP and ferritin, we acknowledge that our multivariate model where we attempted to adjust for inflammation may be biased. However, we note that our sensitivity analysis on the possible impact of unmeasured confounding yielded an E-value of 2.73. This indicates that unmeasured inflammation would have to increase the risk of both delirium and hypocapnia considerably to completely abolish the observed association between delirium and PaCO₂.

Lastly, a measure of PaCO₂ is not a direct measure of the HVR. Although we observed that patients with delirium had a significantly higher respiratory rate at admission than patients without delirium, an accurate measurement of the HVR would have required careful monitoring of the ventilatory minute volume³⁶. PaCO₂ measured at admission is merely a snapshot of the patient’s ventilation at present and may rapidly change because of changes in supplementary oxygen or anxiety during examinations in the emergency department. However, the differences in arterial bicarbonate between patients with and without delirium and the close to normal pH in both groups indicating metabolic compensation of the respiratory alkalosis, strongly suggests that hypocapnia was not a transient phenomenon in the emergency department. Because a low PaCO₂ is the direct consequence of a high HVR and there is a biological rationale to suspect that hypocapnia increases the risk of delirium, it makes sense to investigate associations between PaCO₂ and delirium. Nevertheless, although we present evidence that a lower PaCO₂ increases the risk of delirium independently of other well-established risk factors, it is not possible to conclude from our findings that hypoxic patients with a low PaCO₂ always have a high HVR. Future studies should therefore investigate associations between hypoxia, hypocapnia and delirium by measuring the HVR directly.