The other day I was par-oozing through my twitter feed when I saw this paper pop up from BMJ. My 8-month-old son was born 4 hours shy of 37 weeks and is considered pre-term. Since my wife would fit into this study, and I love her very much, this sparked my interest.
Unbeknownst to me there were quite a few studies across Europe in the late 70’s and 80’s that discovered an association between pre-term delivery and early death in women. (1,2) The studies were wrought with flaws so not much attention was paid to them. (3) This is difficult data to track because you really have a long horizon before you get the outcome. A prospective study of this magnitude would be very challenging to pull off. Fortunately, there are a few countries around the world that collect as much health data on their citizens as they please. Two in particular are Sweden and Denmark. They have nationalized health systems and a national database of health information. This has been in place for roughly 40 years in both countries. They are probably the largest and most extensive health databases in the world. Many observational studies have come from them.
The mechanism behind early death in women after pre-term delivery is unknown. Most of the correlation has been linked to underlying health issues. There was no assessment for confounding in previous studies so it was pretty much assumed this was the cause. However, new data has linked certain inflammatory markers released during pre-term delivery and their contribution to heart and lung disease. These markers are common players on the inflammatory pathway, so their causal relationship is quite loose. These early deaths often occur 10-30 years after the pre-term delivery further convoluting the correlation.
This study was put together using Sweden’s national health database. It is a retrospective cohort. The control group were parous women with no pre-term deliveries. The experimental group were parous women with at least 1 pre-term delivery. The study period began in 1973 with the formation of the data base. It ended December 31st, 2016. The primary endpoint was mortality. The experiment group was sub-divided into extremely pre-term (22-27wk), very preterm (28-33wk), late preterm (34-36wk), and early term (37-38wk). Specific causes of death were eligible for analysis if there were > 500 deaths due to that disease.
Validity of Methods
In this study there are several factors one would want to consider to determine validity. First, was the prognosis similar for individuals in the control and experimental group, and if it was not, was an effort made to correct for any differences. Second, were the circumstances and methods for detecting the outcome similar in both groups. Finally, was follow-up sufficient enough to detect the outcome in both groups without a significant loss to follow-up.
The life expectancy of an individual has a lot do to with their underlying economic status, access to life sustaining resources, education level, and any underlying disease. (4-7) The data in this study was collected from a national birth registry. The researchers had access to common diseases that complicate pregnancy, such as diabetes, hypertension, and pre-eclampsia. They also had access to demographic data such as age, education level, and income. They were able to collect data on smoking status and BMI after 1982.
They found that mothers that delivered pre-term were more likely to be younger, poorer, with lower education level, overweight, smokers, and a higher prevalence of chronic illnesses. They performed a sensitivity analysis to correct for these potential confounders in their survival model. So, the prognosis in pre-term mothers was worse than full-term mothers and it was factored into the analysis.
The circumstances and methods for detecting the outcome were the same in both groups. Sweden has a national death registry. The outcomes were collected from this database. Classification codes were used to determine the cause of death. Although, cause of death is often inaccurate on death certificates, it is reasonable to assume that inaccuracies were equal in both groups. (8) So, in this case the detection of the outcome is valid.
Follow-up in this study was sufficiently long. As we will see in the results section, they had a very large cohort to analyze with a sufficient number of events to have adequate power. Since it is a national database it is unlikely there was a significant loss to follow-up. This would be considered significant if up to 10% of the data was lost. This would be a substantial exodus from the country of Sweden. Follow-up was sufficiently long. The mean and median follow-up period for each individual in the study was about 33 years.
In summary, even though the prognosis for the two groups differed at the beginning of the analysis there was a substantial effort to correct for these differences. The outcome detection is without bias, and follow-up was sufficient. I would consider the methods to be valid. On to the results.
This was a massive study with a ton of analysis. I will do my best to summarize. There were 2.1 million individual women included. The number of deaths during the study period was 76,000. They had 50.7 million total years of follow-up. Mean age at time of delivery was 27, and mean age at the end of observation period was 50.4 for the 2.1 million women involved. Obviously, some were observed longer than others, but these were the averages. As mentioned above pre-term mothers were more likely to be younger, poorer, less educated, overweight, smokers, and have more underlying diseases.
I’m not going to through a ton a numbers into this section because it would just be overbearing. I will do my best to summarize their findings. They used Cox Hazard models for the analysis. Of note, none of the hazard ratios were greater than 2.0, which would be a doubling of the death rate.
Each week of pregnancy was associated with a 4% lower risk of death up to term. The longer the baby was in there, the weaker the correlation. The inverse was true as well, the earlier the delivery the higher correlation with early death. The risk was highest in the first 10 years, and then reduced slightly with each decade after this. Based on the death rate between the groups it was estimated that there was 1 excess death in women for every 73 pre-term births throughout the study period. None of their confidence intervals crossed 1 and they were all fairly narrow.
Their sensitivity analysis did not demonstrate substantial change in the end results when measured variables were assessed for confounding.
In a sub-set analysis, they grouped siblings of pre-term mothers with the pre-term mothers and compared them to see if the siblings died early as well. There was no significant change in the survival model.
In the sub-group analysis for cause of death, cardiovascular disease, respiratory disease, and diabetes related death had the largest differences. The greatest number of deaths were cancer related, but this effect was less compared to the others.
In this study the authors found an association between early death in mothers with pre-term deliveries. It is very difficult to find fault with their conclusion. The study methods and results were robust. They went to great lengths to analyze the data for confounding. They assessed for possible genetic/familial patterns.
There were four specific patterns that strengthened the results. First, a linear relationship between the number of weeks premature and the risk of early death was found. Second, the effect lasted for up to four decades. Even though the hazard ratio waned as time passed, the number of deaths continued to compound. This would be expected if there continued to be an effect from the underlying exposure. Third, there was a linear relationship between number of pre-term deliveries and early death. Finally, their results were similar to previous published reports.
The study had several strengths. The sheer size of it is impressive. They had a large number of events. The data was reliable. The statistical analysis was thought out and executed well. The results have narrow confidence intervals with a moderate effect size. The data was evenly distributed. There were only a few weaknesses. The study population is Swedes or immigrants to Sweden limiting its applicability to the rest of the world. It is an observational study, so proving causation is challenging due to the risk of residual confounding. Data on tobacco use was missing for the first 10 years of the study. It is likely that the prevalence of tobacco use was highest in this decade. (9) There was no assessment of correlation in optimized mothers. What I mean by optimized are those mothers without underlying diseases with strong home support systems.
So, what could be driving this correlation? There are two scenarios to consider. First, would be confounding. The underlying health of the mother is what leads to an early death, and the pre-term delivery is either a random variable that plays no role, or is a result of the underlying health (Chicken = unhealthy mother; Egg = pre-term birth). Second, the pre-term delivery sets of a cascade of events that changes the mother’s health forever which eventually leads to higher disease burden throughout her life (Chicken = pre-term birth; Egg = unhealthy mother).
I’ll start by addressing the first scenario. In the study they did a robust assessment for confounding. Cox proportional hazard models were used. I won’t go much into the math, more on the model building (try not to fall asleep). To build a cox hazard model you start by assessing if each variable is statistically associated with the outcome. Once you have assessed this you place all statistically significant variables into a model. The model is tested for statistical significance. It must be statistically significant and change the hazard ratio to show true confounding. The model will likely show some variables to be significant, and others not. All non-statistically variables are excluded (almost done). The calculation is run again. If the model is not significant after elimination of non-significant variables, then there is no confounding. If it is significant without changing the original hazard ratio, then the confounder is insignificant. (take a deep breath)
One of the limitations with hazard models is there is no way to determine if non-significant variables have synergistic effect. For example, high BMI and diabetes alone may not be significant, but when combined they may. Or smoking and poor economic status combined may be significant, but alone they aren’t. Like most things in medicine, individual risk factors have small effects on endpoints, but when combined with others they cause big trouble.
It would be a bad idea to group risk factors together and run hundreds of models. Given the number of variables in this study, the number of models they could create would be in the billions. By random chance a few hundred million of those groups would be statistically significant. The significance is this scenario would be indistinguishable from random luck. Not to mention you would need some serious computing power to run that analysis if you wanted to be done in this lifetime.
Fortunately, there are instances that real world experience helps us understand a correlation even though it is difficult to prove with a statistical model. If you don’t take care of yourself, you begin to accumulate diseases, these diseases increase your risk of early death. In this case, health status and economic status are the chicken, and pre-term birth is the egg.
The second category involves pro-inflammatory pathways that lead to pre-term labor. Several of the inflammatory proteins that are involved in pre-term labor are also involved in the development of cardiovascular and chronic lung diseases. (10) I tend to ignore loose associations like this because inflammatory pathways tend to be similar throughout the body. The big difference is the trigger. However, this study corresponded with several other studies performed in different populations of women. The signal continues to increase in strength. It would be difficult to discover if some women have a predisposition to over express this inflammatory pathway, or if pre-term labor awakens certain dormant genes in the inflammatory matrix.
The authors recommend more aggressive screening for vascular disease and age appropriate screening for common cancers to mitigate the risk in these women. This is pretty much done here the US and most other developed countries for women that see their doctor regularly.
If you let your imagination run a bit, it’s possible there are several inflammatory markers that induce the pre-term birth, and continue to wreak havoc for years to come. If there is a marker, you better believe your neighborhood biotech company can develop an antibody to suppress it. Fetuses would remain in the womb longer, the inflammatory pathway would be struck down, and women would live longer healthier lives. There could be more to be done with this research than meets the eye.
I reviewed this article because it hit close to home. My son is 8 months old and he is considered pre-term. By a whopping 4 hours. My wife would have been in the experimental group. When I showed her this study her eyes got wide and she said, “I knew something bad was going to happen to me.”
I would advise my female patients that find themselves in this scenario to not fret over this study. Preserve your health to the best of your ability. If you are living an unhealthy lifestyle and have had a pre-term birth, this should be a warning to shape up. Since my wife lives a healthy lifestyle and is fortunately not afflicted with a terrible disease, I’m assured she will continue to be healthy. It is more logical that the underlying health of the mother is the chicken, in this chicken or the egg scenario, and not the pre-term birth. As more research on the mechanism of this correlation is conducted, we will learn more. That is all for now.
1. Davey Smith G, Hyppönen E, Power C, Lawlor DA. Offspring birth weight and parental mortality: prospective observational study and meta-analysis. Am J Epidemiol. 2007;166(2):160-169. doi:10.1093/aje/kwm054
2. Friedlander Y, Paltiel O, Manor O, et al. Birthweight of offspring and mortality of parents: the Jerusalem perinatal study cohort. Ann Epidemiol. 2007;17(11):914-922. doi:10.1016/j.annepidem.2007.07.099
3. Wu P, Gulati M, Kwok CS, et al. Preterm Delivery and Future Risk of Maternal Cardiovascular Disease: A Systematic Review and Meta-Analysis. J Am Heart Assoc. 2018;7(2). doi:10.1161/JAHA.117.007809
4. Chetty R, Stepner M, Abraham S, et al. The Association Between Income and Life Expectancy in the United States, 2001-2014. JAMA. 2016;315(16):1750-1766. doi:10.1001/jama.2016.4226
5. Uchendu FN. Hunger influenced life expectancy in war-torn Sub-Saharan African countries. J Health Popul Nutr. 2018;37(1):11. doi:10.1186/s41043-018-0143-3
6. Steingrímsdóttir OA, Næss Ø, Moe JO, et al. Trends in life expectancy by education in Norway 1961-2009. Eur J Epidemiol. 2012;27(3):163-171. doi:10.1007/s10654-012-9663-0
7. Britton A, Shipley M, Singh-Manoux A, Marmot MG. Successful aging: the contribution of early-life and midlife risk factors. J Am Geriatr Soc. 2008;56(6):1098-1105. doi:10.1111/j.1532-5415.2008.01740.x
8. Eriksson A, Stenlund H, Ahlm K, et al. Accuracy of death certificates of cardiovascular disease in a community intervention in Sweden. Scand J Public Health. 2013;41(8):883-889. doi:10.1177/1403494813499653
9. Tobacco Trends Brief | American Lung Association. Accessed August 27, 2020. https://www.lung.org/research/trends-in-lung-disease/tobacco-trends-brief/overall-tobacco-trends
10. Cappelletti M, Della Bella S, Ferrazzi E, Mavilio D, Divanovic S. Inflammation and preterm birth. J Leukoc Biol. 2016;99(1):67-78. doi:10.1189/jlb.3MR0615-272RR