To determine which factors are associated with sudden unexpected infant death (SUID) by time of day.
Data were analysed from the National Fatality Review Case Reporting System (2006‐2015). Out of 20 005 SUID deaths in 37 states, 12 191 (60.9%) deaths had a recorded nearest hour of discovery of the infant. We compared distribution patterns between time of death and 118 variables to determine which were significantly correlated with SUID time of death using advanced statistical modelling techniques.
The 12‐hour time periods that were most different were 10:00 to 21:00 (daytime) and 22:00 to 09:00 (nighttime). The main features that were associated with nighttime SUID were bed sharing, younger infants, non‐white infants, placed supine to sleep and found supine, and caregiver was the parent. Daytime SUID was associated with older infants, day care, white infants, sleeping in an adult bed and prone sleep position. Factors not associated with time of death were sex of the infant, smoking and breastfeeding.
Sudden unexpected infant death deaths that occur at night are associated with a separate set of risk factors compared to deaths that occur during the day. However, to minimise risk, it is important to practice safe sleep guidelines during both nighttime and daytime sleep.
Tatiana M. Anderson and Kelty Allen are joint first authors.
National Fatality Review Case Reporting System
sudden infant death syndrome
sudden unexpected infant death
Gradient boosted decision trees models
Sudden unexpected infant deaths (SUID) peak at night. The 12‐hour time period that was most different was 10:00 to 21:00 (daytime) and 22:00 to 09:00 (nighttime). SUID deaths that occur at night are associated with a separate set of risk factors compared to deaths that occur during the day. To minimise risk, it is important to practice safe sleep guidelines during both nighttime and daytime sleep.
Sudden unexpected infant death (SUID) is a term used to describe the sudden and unexpected death of a baby less than 1‐year old in which the cause was not obvious before investigation.
Most analyses of circadian variation in cases of sudden infant death have solely focused on SIDS, not SUID, cases. The majority of SIDS deaths occur at night
There are differences between day and night sleep that could result in disparate diurnal pathogeneses in cases of sudden infant deaths. Several theories have been posed as to why risk factors differ between day and nighttime SIDS including altered patterns of circadian rhythm,
Many of the previous studies comparing risk factors to time of death were completed before the Back to Sleep campaigns in the early 1990s that led to a significant decrease in SIDS/SUID rates. In addition, sample sizes for the studies above were limited to only a few hundred cases and nearly every study was conducted in countries outside of the United States. As the deaths are rarely observed, the exact time of death is often difficult to estimate.
This study utilises advanced modelling techniques to analyse over 20 000 SUID cases between 2006 and 2015 in the United States to identity which variables significantly correlate with time of death.
The data set was provided by the National Fatality Review Case Reporting System (NFR‐CRS). It includes 20 005 infant deaths across 37 states between 2006 and 2015 where the Child Death Review team indicated primary cause of death to be asphyxia, injury‐undetermined, injury‐unknown, SIDS, medical‐undetermined or medical‐unknown. All these deaths were considered to be SUID and are the subject of this study.
We excluded infants that were born at <28 weeks’ gestation. This exclusion avoids misclassification due to preterm birth.
Reported time of discovery to the nearest hour (1‐12, AM/PM). For the purpose of this paper, these times have been converted to the 24‐hour clock.
There were 118 explanatory variables. Variables relevant to this report are as follows: age (days, categorised 0‐59, 60‐120 and 121+), birthweight, gestation, multiple birth, primary caregiver (parent vs. not parent), infant race (white vs. non‐white), sex of infant, prenatal care provided, cigarette smoking before pregnancy, cigarette smoking during pregnancy, infant breastfed, place of death (home, relative or friend's home, day‐care centre or home, foster care home, hospital, other = hotel, baby sitter, shelter, outside, Indian reservation, public buildings, other specified categories, not specified and unknown), death related to sleep environment (specific question answered by child death review: ‘Was the death related to sleeping or the sleep environment?’), sleeping place (crib, bassinet, adult bed, other), bed sharing (sleeping on same surface with person or animal), state (included in model, but not reported because of privacy concerns), position infant placed to sleep (back, stomach, side or unknown), position found (back, stomach, side or unknown) and sleeping on the floor.
We only used data in the model where sample size was at least
Those cases with missing time of death were compared to those with this data.
The data were used to determine which two 12‐hour periods of time most differentiated SUID cases, using statistical models and the feature set described above. To ensure that results were consistent, we developed both logistic regression models and gradient boosted decision trees models (XGBoost). Gradient boosted decision trees are an ensemble model of decision trees whose optimisation algorithm often results in better model performance compared to logistic regression, as it does in this case. However, the model is inherently non‐linear, which prevents us from using it to calculate adjusted odds ratios. Having a logistic regression model that agrees with the XGboost model on the daytime/nighttime split and mostly agrees on the list of predictive variables gives us greater confidence in the results and allows us to calculate reliable adjusted odds ratios. All logistic regression and XGBoost models were adjusted for covariates, and model performance was evaluated using macro‐average F1 score and 5‐fold cross‐validation.
To understand the features most correlated with daytime and nighttime SUID, we used the SHAP TreeExplainer method
In the NCFRP case registry, there were 20 005 SUID deaths. Of these, 12 191 (60.9%) deaths had a recorded nearest hour of discovery of the infant. These deaths are the subject of this study. Distribution pattern of the reported time of death/found is shown in Figure
Distribution of time of death
Cases with a reported time of death are less likely to be bed sharing (63% vs. 72%,
All 12 possible divisions of the day into distinct 12‐hour intervals were analysed, and the time periods which were most differentiable using our statistical models were 22:00 to 09:00 and 10:00 to 21:00 (hereafter referred to as ‘nighttime’ and ‘daytime’, respectively). The 5‐fold f1 score for the models predicting 22:00‐09:00 and 10:00‐21:00 was 0.67 for the XGboost model and 0.66 for the logistic regression model. This is compared to f1 scores of 0.41 and 0.42, respectively, for the pair of times which were least statistically differentiable (05:00‐16:00 vs. 17:00‐04:00). There were 4752 (39.0%) deaths in daytime and 7439 (61.0%) at night.
After controlling for all features in the dataset, both XGBoost and logistic regression models showed that the factors that were most associated with nighttime SUID included bed sharing, younger infants, non‐white infants, placed supine to sleep and found in supine position, and caregiver was the parent (Table
Prevalence of risk factors that were significant and unadjusted and adjusted odds (OR) ratios and their 95% confidence intervals (CI)
Daytime | Nighttime | OR (95% CI) | AOR (95% CI) | |
---|---|---|---|---|
Infant age | ||||
0‐59 days | 1052 (22.3%) | 2891 (39.1%) | 0.45 (0.41, 0.49) | 0.56 (0.51, 0.61) |
60‐120 days | 1754 (37.2%) | 2656 (35.9%) | 1.06 (0.98, 1.14) | 0.99 (0.91, 1.07) |
121+ days | 1914 (40.6%) | 1854 (25.1%) | 2.04 (1.88, 2.20) | 1.72 (1.58, 1.88) |
Race | ||||
White | 3050 (64.5%) | 4084 (55.1%) | 1.48 (1.37, 1.59) | 1.34 (1.22, 1.46) |
Non‐white | 1681 (36.5%) | 3329 (44.9%) | 0.68 (0.63, 0.73) | 0.75 (0.68, 0.82) |
Primary caregiver | ||||
Parent | 4368 (92.3%) | 6973 (94.1%) | 0.76 (0.66, 0.88) | 0.83 (0.69, 0.99) |
Not parent | 363 (7.7%) | 440 (5.9%) | 1.32 (1.14, 1.52) | 1.21 (1.01, 1.45) |
Death related to sleep environment | ||||
Yes | 4241 (89.7%) | 7078 (95.5%) | 0.41 (0.35, 0.47) | 0.49 (0.33, 0.72) |
No | 490 (10.3%) | 335 (4.5%) | 2.44 (2.11, 2.82) | 2.05 (1.39, 3.00) |
Other place of death | ||||
Yes | 213 (4.5%) | 222 (3.0%) | 1.53 (1.26, 1.85) | 1.57 (1.27, 1.94) |
No | 4518 (95.5%) | 7191 (97.0%) | 0.65 (0.54, 0.79) | 0.64 (0.52, 0.79) |
Day care | ||||
Yes | 460 (9.7%) | 51 (0.7%) | 15.55 (11.6, 20.8) | 8.79 (6.48, 11.9) |
No | 4271 (90.3%) | 7362 (99.3%) | 0.06 (0.05, 0.09) | 0.11 (0.08, 0.15) |
In adult bed | ||||
Yes | 1685 (35.6%) | 3693 (49.8%) | 0.56 (0.52, 0.60) | 2.15 (1.11, 4.19) |
No | 3046 (64.4%) | 3720 (50.2%) | 1.79 (1.67, 1.93) | 0.46 (0.24, 0.90) |
Bed sharing | ||||
Yes | 1401 (29.6%) | 4593 (62.0%) | 0.26 (0.24, 0.28) | 0.28 (0.25, 0.31) |
No | 3330 (70.4%) | 2820 (38.0%) | 3.87 (3.58, 4.19) | 3.57 (3.18, 4.02) |
Sleeping on floor | ||||
Yes | 72 (1.5%) | 100 (1.3%) | 1.13 (0.83, 1.53) | 2.18 (1.04, 4.60) |
No | 4659 (98.5%) | 7313 (98.7%) | 0.88 (0.65, 1.20) | 0.46 (0.22, 0.96) |
Position placed to sleep | ||||
Stomach | 1045 (22.1%) | 1272 (17.2%) | 1.37 (1.25, 1.50) | 1.27 (1.12, 1.45) |
Back | 1598 (33.8%) | 2918 (39.4%) | 0.79 (0.73, 0.85) | 0.81 (0.71, 0.92) |
Side | 496 (10.5%) | 845 (11.4%) | 0.91 (0.81, 1.02) | 1.05 (0.90, 1.22) |
Unknown | 1592 (33.7%) | 2378 (32.1%) | 1.07 (0.99, 1.16) | 0.86 (0.71, 1.04) |
Position found | ||||
Stomach | 1851 (39.1%) | 2238 (30.2%) | 1.49 (1.38, 1.60) | 1.28 (1.13, 1.45) |
Back | 877 (18.5%) | 1987 (26.8%) | 0.62 (0.57, 0.68) | 0.76 (0.66, 0.86) |
Side | 438 (9.3%) | 895 (12.1%) | 0.74 (0.66, 0.84) | 1.05 (0.90, 1.23) |
Unknown | 1565 (33.1%) | 2293 (30.9%) | 1.10 (1.02, 1.19) | 1.07 (0.87, 1.31) |
Distribution of time of death for infants aged 0‐59 days and 120+ days of age
Distribution of time of death categorised as daytime and nighttime and infant age categorised as 0‐59, 60‐120 and 121+ days of age
In contrast, factors that were most correlated with daytime SUID included older infants, day care, white infants, sleeping in an adult bed, put to sleep in prone position, found in prone position, death not related to sleep environment, sleeping on the floor and other places of death (Table
Bed sharing is the strongest variable associated with nighttime SUID and explains part of the age difference between nighttime and daytime SUID. However, even after restricting our analysis to non‐bed sharing infants only, we see very similar statistical results. 10:00‐21:00 and 22:00‐09:00 continue to be highly statistically differentiable, and the variables for nighttime or daytime death remain very similar, with older infants, day care and sleeping in an adult bed being most correlated with daytime SUID (data not shown).
Several features were not associated with time of death including maternal smoking, sex of infant, breastfeeding and caregiver's income.
Consistent with previous reports, the majority of SUID cases occur during nighttime sleep. A common reported scenario is that an apparent healthy infant is placed to sleep in the evening and found dead in the morning. Thus, the exact time of death is unknown, and this is indeed likely for most cases found around the time when the parents wake up. At the younger ages (0‐2 months), at which the risk is especially high for a nighttime death, babies are generally sleeping in short stretches due to needing to be fed and changed regularly, the anticipation of frequent waking increases the chance of finding the infant within a shorter interval of actual death. In this study, there are deaths reported at all hours from the late evening to morning with a clear pattern of progressively increasing number of SUID cases discovered each hour between 00:00 and 07:00 (Figure
Perhaps the biggest strength of this study is the large population size. Most studies analysing time of death in association with sudden infant death have been limited to several hundred cases in contrast to the >12 000 SUID cases reported here. We were thus able to bring much higher granularity to the comparative analysis of risk factors between the nighttime and daytime sleep environments. Another strength is the advanced statistical modelling techniques used in the analysis.
There were differences between those with time of death recorded (60.9%) and those without (39.1%). In most instances, the differences were small; for example, the mean age of caregiver where time found was recorded was 25.7 years versus 25.4 years for those without time recorded. However, cases with a reported time of death were much less likely to be bed sharing than those without a recorded time of death (63% vs. 72%,
Much of our findings agreed with previous account
Even restricting to only those infants who were at home and not bed sharing at the time of death, we still observe two separate distribution patterns that peak in the early morning hours for younger infants and during the daytime for older infants (Figure
Maternal cigarette smoking is strongly associated with SUID,
In addition to smoking, other features that were also not associated with time of death after controlling for the full feature set included, sex of infant, breastfeeding and caregiver's income and age. These observations suggest that the diurnal variation in SUID risk is not related to socioeconomic status.
This is one of the first studies to report a racial difference in time of death (aside from a reported increased risk of nighttime death in the Maori population
Similar to the New Zealand
After controlling for all features in the dataset, the factors that are associated with the greatest risk of a nighttime SUID death includes bed sharing, young infants, non‐white infants and supine sleep position. In contrast, factors associated with daytime SUID included older infants, day‐care attendance, white infants and prone sleep position. The circadian variations in SUID risk factors described here suggest different underlying causal mechanisms of death, and these data should provide clues towards future physiological and genetic research dedicated to uncovering the mechanistic differences.
The authors have no conflicts of interest relevant to this article to disclose.
This work was supported by the National Institutes of Health (grant numbers P01HL0906654 and R01HL126523 awarded to JMR), and by the Aaron Matthew SIDS Research Guild. EAM was supported by Cure Kids. The funder did not participate in the work.
The content of this paper is solely the responsibility of the authors and should not be construed as the official position or policy of, nor should any endorsements be inferred by HRSA, HHS or the US Government.
FigS1
Click here for additional data file.
This data set was provided by the National Fatality Review and Prevention. The National Center is funded in part by Cooperative Agreement Number UG7MC28482 from the US Department of Health and Human Services (HHS), Health Resources and Services Administration (HRSA), Maternal and Child Health Bureau (MCHB) as part of an award totalling $799 999 annually with 0% financed with non‐governmental sources.
We would like to thank Juan Lavista, Urszula Chajewska, Ricky Johnston and Xiaohan Yan for statistical guidance and useful discussion. This research was inspired by the results of a survey of SIDS parents run by Darci and Robert Torres. A special thank you to John and Heather Kahan for sparking this collaboration.