The scene will be familiar to many of us; our patient is leaving the ICU after their brutal clash with sepsis. They were really, REALLY sick when they came in but you and the team have done an amazing job. Aggressive cardiovascular support, meticulous attention to lung protective ventilation, intensive physiotherapy – you’ve done a cracking job in maximising their chances and they’ve pulled through. They are now stepping down to the ward and it’s all smiles, best wishes and high fives around the room. Job well done team!
But wait, one of your junior trainees has pointed out that their 2 year mortality is actually quite affected by this illness, and they’re not out of the woods yet…
The mood darkens slightly. The party hats are taken off. Killjoy! Can’t we celebrate the little victories?
And yet this is where we are in intensive care medicine much of the time. A spell on the unit is hardly a spa weekend that renews your vigour and health – the acute illness process is just the start. This topic is the focus of the paper from our recent journal club, available open-access through the link below. Specifically, it looks at the impact of sepsis on our patients in the long term.
Prescott H et al. Late mortality after sepsis: propensity matched cohort study. BMJ. 2016. 353. i2375
But wait, one of your junior trainees has pointed out that their 2 year mortality is actually quite affected by this illness, and they’re not out of the woods yet…
The mood darkens slightly. The party hats are taken off. Killjoy! Can’t we celebrate the little victories?
And yet this is where we are in intensive care medicine much of the time. A spell on the unit is hardly a spa weekend that renews your vigour and health – the acute illness process is just the start. This topic is the focus of the paper from our recent journal club, available open-access through the link below. Specifically, it looks at the impact of sepsis on our patients in the long term.
Prescott H et al. Late mortality after sepsis: propensity matched cohort study. BMJ. 2016. 353. i2375
What's it about?
Prescott and colleagues start with the uncomfortable truth that we’ve just highlighted; patients who survive their acute sepsis episode still have quite a high mortality in the next few months and years. This isn’t particularly news, but there is some debate as to what exactly this represents. One line of argument is that it is all down to the patient’s comorbidities, i.e. they weren’t very well to start with and this may have contributed to why they got so sick with this infection. The converse of this is that it is actually the repercussions of the sepsis that makes you more likely to die in the following months. The evidence for each of these viewpoints seems conflicting, hence the authors have set out to try and provide an answer.
What did they do?
The obvious challenge from a research point of view is how best to measure the impact of such an event (sepsis) with an appropriate control. There is clearly no way to randomise a member of the public to receive sepsis, so it becomes more about identifying an adequate control group for the patients who do get sepsis. The authors have tried their best to do this, and go one better.
They have used an existing database from which to draw their samples; the US Health and Retirement Study (HRS). This is a longitudinal survey of US citizens over the age of 50 with regular questionnaires on topics such as health and employment. This is also linked to the patient’s Medicare records, thus allowing a link between the patient’s hospital care. With a database size of 37,000 adults that are broadly representative of the US population, the authors felt that this was a suitable source to take their sample.
The design of their study was overall a propensity matched cohort study. They were able to identify their primary cohort fairly easily i.e. the patients who developed sepsis within their inclusion criteria. The challenge was to find an appropriate control group with which to compare it to. The propensity matching was the approach used to try and create this control group from the patients within their large database. Now this all seems a little bit of statistics voodoo to me, but essentially they started by determining which criteria it was important to be similar between the patients to minimise confounding. They determined this based on age, gender and their ‘centile risk of sepsis’. This last parameter appeared to be a composite of a number of impacting factors; ethnicity, BMI, marital status, self-reported health, wealth, ADL scores, recent sepsis and Charlston comorbidities. All fairly clever stuff, but I suppose it does very much depend how confident you are that it is these parameters which are the true representative features of similarity between the groups. Either way, they used this approach to 1:1 match the septic patients with 3 other cohorts; those patients not in hospital, those admitted with a non-septic infection, and those admitted with a sterile SIRS process e.g. pancreatitis.
Having identified their cohorts the outcome they were interested in was all cause mortality between 31 days and 2 years. They also made note of the actual mode of death for these patients who did go on and die within 2 years.
They have used an existing database from which to draw their samples; the US Health and Retirement Study (HRS). This is a longitudinal survey of US citizens over the age of 50 with regular questionnaires on topics such as health and employment. This is also linked to the patient’s Medicare records, thus allowing a link between the patient’s hospital care. With a database size of 37,000 adults that are broadly representative of the US population, the authors felt that this was a suitable source to take their sample.
The design of their study was overall a propensity matched cohort study. They were able to identify their primary cohort fairly easily i.e. the patients who developed sepsis within their inclusion criteria. The challenge was to find an appropriate control group with which to compare it to. The propensity matching was the approach used to try and create this control group from the patients within their large database. Now this all seems a little bit of statistics voodoo to me, but essentially they started by determining which criteria it was important to be similar between the patients to minimise confounding. They determined this based on age, gender and their ‘centile risk of sepsis’. This last parameter appeared to be a composite of a number of impacting factors; ethnicity, BMI, marital status, self-reported health, wealth, ADL scores, recent sepsis and Charlston comorbidities. All fairly clever stuff, but I suppose it does very much depend how confident you are that it is these parameters which are the true representative features of similarity between the groups. Either way, they used this approach to 1:1 match the septic patients with 3 other cohorts; those patients not in hospital, those admitted with a non-septic infection, and those admitted with a sterile SIRS process e.g. pancreatitis.
Having identified their cohorts the outcome they were interested in was all cause mortality between 31 days and 2 years. They also made note of the actual mode of death for these patients who did go on and die within 2 years.
What did they find?
After all the stats wizardry they ended up with 3 matched cohorts of different sizes: sepsis vs not in hospital (777 patients), sepsis vs infection (788 patients), and sepsis vs sterile inflammatory process (504 patients). This matching was done from a total of 960 patients who had sepsis from their database.
The mortality results they observe from these comparisons are interesting:
Interestingly the most common terminal illness in all groups was of further infection, either pneumonia or further sepsis. The similarity of this process across all the cohorts suggested that the increased mortality wasn’t particularly through a new modality, but rather an increased incidence of that modality overall.
The mortality results they observe from these comparisons are interesting:
- Sepsis vs not in hospital – adjusted odds ratio 3.5 (2.7 – 4.5)
- Sepsis vs infection – adjusted OR 1.6 (1.3 – 2.1)
- Sepsis vs sterile inflammatory process – adjusted OR 2.3 (1.7-2.1)
Interestingly the most common terminal illness in all groups was of further infection, either pneumonia or further sepsis. The similarity of this process across all the cohorts suggested that the increased mortality wasn’t particularly through a new modality, but rather an increased incidence of that modality overall.
Is it any good?
Now I found this quite a difficult question to confidently answer. There are many good points to this study. Firstly, they have managed to get a good large number of patients together to cover the specific areas that they wanted. In addition these patients had some pretty detailed demographic information available from the HRS survey (and high follow up rates), allowing a good degree of accuracy within their propensity matching magic. Perhaps more importantly they have gone about this study in a really good way to try and answer the clinical question they had at the start. I like the use of similar but different patient groups (sterile inflammation and standard infection) as contrasting cohorts as I think that this adds weight to their conclusions when considering the drawbacks of the study.
Now speaking of drawbacks, there is perhaps the one major issue with this study - propensity matching. Now I will get my confessions out that I am certainly no statistician, but I am willing to accept that the statistical basis of the matching is fairly sound. The main worry I have is about the step before this – how can we be confident about the parameters which we have chosen to match with? I know that authors love pointing out that groups are similar through their Chi squared’ and ‘t-tests’ etc. but all that shows is that they are similar for those particular parameters they are testing. On the surface it seems reasonable to accept that our cohorts are matched for age and comorbidities etc., but it remains a leap of faith that this makes them at a similar risk of dying from any particular illness, particularly here because I cannot see that these factors are well validated markers for this purpose. I know this will be the challenge whenever randomisation isn’t done, and it is clear that this is impossible in this case, so we may have to accept that this is the best we can do, but it leaves a sense of unease with the conclusions.
Other minor drawbacks relate to the methods of data collection. The patients are all rather elderly because of the nature of the HRS, limiting the generalisability of the results to younger patients with sepsis. The dependence on both appropriate coding and a claims based medical system is also an area with some potential for error. These limitations are not tiny, but in my mind are a rather more appreciable that that from the propensity matching.
Now speaking of drawbacks, there is perhaps the one major issue with this study - propensity matching. Now I will get my confessions out that I am certainly no statistician, but I am willing to accept that the statistical basis of the matching is fairly sound. The main worry I have is about the step before this – how can we be confident about the parameters which we have chosen to match with? I know that authors love pointing out that groups are similar through their Chi squared’ and ‘t-tests’ etc. but all that shows is that they are similar for those particular parameters they are testing. On the surface it seems reasonable to accept that our cohorts are matched for age and comorbidities etc., but it remains a leap of faith that this makes them at a similar risk of dying from any particular illness, particularly here because I cannot see that these factors are well validated markers for this purpose. I know this will be the challenge whenever randomisation isn’t done, and it is clear that this is impossible in this case, so we may have to accept that this is the best we can do, but it leaves a sense of unease with the conclusions.
Other minor drawbacks relate to the methods of data collection. The patients are all rather elderly because of the nature of the HRS, limiting the generalisability of the results to younger patients with sepsis. The dependence on both appropriate coding and a claims based medical system is also an area with some potential for error. These limitations are not tiny, but in my mind are a rather more appreciable that that from the propensity matching.
Final Thoughts
Overall, the journal club verdict was that there were a few potential limiting factors within the methodology of this study. However, the use of the multiple cohorts has provided some strength to the conclusion that this is a genuine trend and it also seems unlikely that there is another approach to better answer such a question. On balance, it seems that it is likely that even if you survive your sepsis episode, the toll of such an illness lasts for at least the next couple of years. This impact is separate from any underlying comorbidities, though I am unsure if it may be partly attributable to specific critical illness interventions and sepsis complications (mechanical ventilation, AKI) as much as the sepsis aetiology itself. I think there is increasing interest in the ongoing rehabilitation of patients after their critical illness has finished, and this paper really supports the idea critical care has to go beyond the critical care unit.
Thanks for reading and also a big thanks to all the members of the journal club who were involved in the discussions. The brave amongst you may like to have a look at the link below looking at propensity matching.
BW
Tom Heaton
Thanks for reading and also a big thanks to all the members of the journal club who were involved in the discussions. The brave amongst you may like to have a look at the link below looking at propensity matching.
BW
Tom Heaton
References & Links
- Prescott H et al. Late mortality after sepsis: propensity matched cohort study. BMJ. 2016. 353. i2375
- Stuart E. Matching methods for causal interference: a review and a look forward. Stat Sci. 2010. 25(1); 1-21
RSS Feed
