Conference PaperPDF Available

Abstract and Figures

In this paper we show early evidence of the feasibility of detecting labour during pregnancy, non-invasively and in free-living. In particular, we present machine learning models aiming at dealing with the challenges of unsupervised, free-living data collection, such as identifying periods of high quality data and detecting physiological changes as labour approaches. During a first phase, physiological data including electrohysterography (EHG, the electrical activity of the uterus), heart rate (HR) and gestational age (GA) were collected in laboratory conditions for model development. In particular, data were collected 1) during simulated activities of daily living, aiming at eliciting artifacts and developing diagnostic models for free-living data 2) during pregnancy, including labour, aiming at developing labour probability models from clean, supervised physiological recordings. Machine learning models using datasets 1) and 2) were deployed in free-living, longitudinally, in 142 pregnant women, between week 22 of pregnancy and delivery. A total of 1014 hours of data and an average of 7 hours per person were collected. Output of the developed models was analyzed to determine the feasibility of detecting labour non-invasively using physiological data, acquired with a single sensor placed on the abdomen. Results showed that the probability of being in labour for recordings collected during the last 24 hours of pregnancy was consistently higher than the probability during any other pregnancy week. Thus, non-invasive labour detection from physiological data seems promising.
Content may be subject to copyright.
Towards Non-invasive Labour Detection: A Free-Living Evaluation
Marco Altini1, Elisa Rossetti1, Michiel J. Rooijakkers1and Julien Penders2
Abstract In this paper we show early evidence of the
feasibility of detecting labour during pregnancy, non-invasively
and in free-living. In particular, we present machine learning
models aiming at dealing with the challenges of unsupervised,
free-living data collection, such as identifying periods of high
quality data and detecting physiological changes as labour
approaches. During a first phase, physiological data includ-
ing electrohysterography (EHG, the electrical activity of the
uterus), heart rate (HR) and gestational age (GA) were collected
in laboratory conditions for model development. In particular,
data were collected 1) during simulated activities of daily
living, aiming at eliciting artifacts and developing diagnostic
models for free-living data 2) during pregnancy, including
labour, aiming at developing labour probability models from
clean, supervised physiological recordings. Machine learning
models using datasets 1) and 2) were deployed in free-living,
longitudinally, in 142 pregnant women, between week 22 of
pregnancy and delivery. A total of 1014 hours of data and
an average of 7hours per person were collected. Output of
the developed models was analyzed to determine the feasibility
of detecting labour non-invasively using physiological data,
acquired with a single sensor placed on the abdomen. Results
showed that the probability of being in labour for recordings
collected during the last 24 hours of pregnancy was consistently
higher than the probability during any other pregnancy week.
Thus, non-invasive labour detection from physiological data
seems promising.
I. INTRODUCTION AND RELATED WORKS
Labour is the process of delivering a baby and placenta,
membranes and umbilical cord. The process is normally di-
vided into three phases, going from more regular contractions
to delivery [1]. The ability to detect labour non-invasively,
outside of the hospital environment, could help expecting
mothers avoiding unnecessary visits, receive better care, as
well as improve detection and management of pregnancy
complications such as preterm birth. Unfortunately, no such
method exists today, and labour is often diagnosed once at
the hospital, using subjective methods [2].
A. Predicting labour
One of the earliest signs of labour is a change in uterine
activity, typically reflected as an increase in frequency and
regularity of uterine contractions. Recent developments in
wearable sensors technology as well as signal processing
and machine learning techniques made it possible to further
investigate changes in uterine activity and contractions non-
invasively. Analysis of the electrical activity of the uterus,
or electrohysterography (EHG) is promising as it reflects the
This work was funded by Bloomlife
1M. Altini, M. J. Rooijakkers and E. Rossetti are with Bloom Technolo-
gies, Genk, BE email: altini.marco@gmail.com
2J. Penders is with Bloomlife, San Francisco, USA
source of the contractions. Uterine contractions are generated
by the electrical activity originating from the depolarization-
repolarization of smooth muscle myometrial cells, creating
intermittent bursts of spike-like action potentials [3]. To-
gether with changes in EHG, our group as well as others
showed consistent changes in cardiac activity during labour,
making maternal heart rate (HR) a useful predictor in the
detection of labour from physiological data [4]. However,
to the best of our knowledge, no study today was able
to investigate such changes in EHG and HR outside of
supervised laboratory settings.
B. Challenges in real-life deployment of wearable sensors
Measuring and collecting data in unsupervised free-living
conditions is finally becoming more common as wearable
sensors are entering the lives of millions of individuals
worldwide. While many of these devices are consumer gad-
gets, several companies have developed clinically validated
tools and released them on the market. Even for devices and
sensors that have been validated rigorously under supervised
laboratory conditions, it can be challenging to trust data
acquired in free-living, as use (and misuse) of the system is
outside of the researchers supervision. However, diagnostics
and signal quality are hardly discussed in the development
of new technologies.
C. Approach and contributions
In this paper, we highlight the steps taken while in-
vestigating the ability of a single sensor device able to
measure physiological data during pregnancy, to estimate the
probability of labour in unsupervised free-living settings. To
address the above mentioned challenges on signal quality
and trusting data acquired outside standard research settings,
we structured our work as follows:
First, we developed artifact and labour probability es-
timation models using data collected under supervised
laboratory settings. We combined EHG and HR data
acquired at different gestational ages in a sample of
58 and 55 women respectively. We show that our
models are able to identify artifacts and labour with
high accuracy.
Secondly, we deployed artifact and labour probability
estimation models in free-living, on a group of 142
pregnant women for whom no data was collected during
model development. We highlight how high quality data
could be identified and show that the probability of
being in labour for recordings collected during the last
24 hours of pregnancy is consistently higher than the
probability during any other pregnancy week.
Supervised*Laboratory*
Se1ngs*
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
Ac6vi6es*of*daily*living*
aiming!at!genera*ng!
ar*facts!
Delivery!room:*labour*
recordings.!High!uterine!
and!cardiac!ac*vity,!clean!
data!
Pregnancy*recordings.!
Low!uterine!and!cardiac!
ac*vity,!clean!data!
Models*
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
Ar6facts*probability:!
combines!EHG,!HR!and!
accelerometer!data!to!
determine!data!quality!
every!second!(short!
windows)!
Labour*probability:!
combines!EHG,!HR!and!
GA!data!to!determine!
the!probability!of!
imminent!labour!every!
20!minutes!(long!
windows)!
Unsupervised*Free-Living*
Se1ngs*
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
!
Discard*poor*quality*
recordings:*at!least!85%!
of!the!data!for!each!20!
minutes!segment!needs!
to!be!high!quality!to!be!
included!
Compute*probability*of*
being*in*labour:*high!
quality!recordings!
collected!in!free-living!
are!analyzed!to!
determine!the!likelihood!
of!imminent!labour!
Fig. 1. Overview of the approach and analysis described in this paper. From
left to right, we first collected data under supervised laboratory settings,
aiming at being able not only to determine labour probability, but also
to analyze data quality when collected unsupervisedly. Then, models were
deployed in free-living and high quality data was used to determine the
models’ ability to detect labour probability outside of laboratory settings.
II. DATA ACQUISITION
A. Laboratory studies
Measurements for all laboratory studies were performed
using a research version of the Bloomlife wearable device,
configured to acquire two channels ExG at 4096 Hz and
triaxial accelerometer data at 128 Hz (see Fig. 4). Measure-
ments were assigned to the labour class retrospectively based
on delivery within 24 hours from the measurement.
1) Artifact detection: Fifty-eight recordings were col-
lected in pregnant women under two conditions. First, data
was acquired during rest while not experiencing contractions
as well as during activities aiming at generating artifacts such
as: talking, coughing, turning to a side, stretching, walking,
standing up, playing with a toddler, bending, rubbing the
abdomen and contracting the abdomen and rectum. These
activities were chosen as common during daily life and
typically interfering with EHG and HR data. Secondly, 18
recordings were collected during labour, so that EHG and
HR data could be collected as a true representation of labour,
without artifacts. An example is shown in Fig. 2.
2) Labour detection: Thirty-seven pregnancy recordings
collected at different time points during pregnancy (GA =
37.5±4.4weeks) were added to labour recordings described
in the previous Sec. to complete the training set used for
labour probability estimation. An example is shown in Fig. 3.
The studies were approved by the ethics committee of
Ziekenhuis Oost-Limburg.
B. Free living
142 women used the Bloomlife device in free-living
between week 22 of pregnancy and delivery. Instructions
to properly apply the patch and sensor were provided. Data
collected included EHG, HR and GA as well as the output of
artifact and labour probability estimation models. Similarly
to laboratory recordings, measurements were assigned to the
labour class retrospectively based on delivery within 24 hours
from the measurement. Delivery was self-reported by the
users. All women agreed to their data being used for research
as part of Bloomlife’s Terms of Services and Privacy Policy.
0
25
50
75
2400 2700 3000 3300 3600
Time (s)
RMS
Reference artifacts
FALSE
Labour recording used for artifact detection model training
70
80
90
100
110
2400 2700 3000 3300 3600
Time (s)
Heart rate (bpm)
0
250
500
750
1000
0 400 800 1200 1600
Time (s)
RMS
Reference artifacts
FALSE
TRUE
Activities recording used for artifact detection model training
80
100
120
0 400 800 1200 1600
Time (s)
Heart rate (bpm)
Fig. 2. Example of data used to build artifact detection models. Recordings
during labour were included to make sure the classifier would learn that
high EHG and HR activity are not necessarily associated with artifacts and
noise, but also with uterine contractions, as shown in the first two plots. The
bottom two plots show data collected while performing activities aiming
at generating artifacts, such as rubbing the abdomen or walking. Artifact
probability is estimated on a window by window basis as they can be short.
BL_KICK_2016−003−S012_03
BL_CXN_2015−002−S002_06
2400 2700 3000 3300 3600
6000 6250 6500 6750 7000 7250
0
100
200
300
0
50
100
150
Time (s)
RMS
Reference labour class
FALSE
TRUE
Recordings used for labour detection model training
BL_KICK_2016−003−S012_03
BL_CXN_2015−002−S002_06
2400 2700 3000 3300 3600
6000 6250 6500 6750 7000 7250
60
80
100
120
140
80
100
120
140
Time (s)
Heart rate (bpm)
Fig. 3. Example of data used to build labour detection models. Recordings
during labour are shown in lighter gray, highlighting higher uterine activity
and heart rate patterns typical of contractions. Recording collected earlier
during pregnancy, as shown in the second and fourth plots, typically show
less uterine and heart rate activity, when collected at rest under supervised
laboratory conditions. Labour probability is estimated on an entire recording,
and not on a window by window basis as a recording can either be during
labour or not, differently from what we have for artifacts.
Fig. 4. Bloomlife wearable sensor.
III. DATA ANALYS IS
A. Pre-processing
For all models features were derived after isolating the
main frequency bands for the signals of interest. In particular,
EHG data was downsampled to 16 Hz and filtered between
0.1 and 4 Hz [5] while ECG was band pass filtered between
2Hz and 98 Hz to remove all out-of-band noise and
then using a notch filter at 50 Hz to remove powerline
interference. Maternal R-peak detection was performed based
on the algorithm described in [5]. Accelerometer data were
bandpass filtered between 1and 10 Hz to isolate maternal
movement. Recordings were split in 20-minutes chunks and
time and frequency domain features were extracted from
EHG and HR data.
B. Features Extraction
First, features were extracted over 16 seconds windows to
capture EHG and cardiac properties as they evolve over time.
An example of two features can be seen in Fig. 3 where the
Root Mean Square (RMS) of the EHG signal and the mean
HR over consecutive 16 seconds windows is shown. Features
extracted were: RMS of the EHG signal, normalized range
of the EHG signal [6], mean crossing rate of the EHG signal,
power of the EHG signal and mean HR. Features were then
summarized in terms of mean and standard deviation over 20-
minute windows. Additionally, we computed features on the
entire 20-minute segment, with the aim of capturing more
information relative to the rhythmic pattern present during
labour. In particular, we extracted the following features:
power of the EHG and HR signal, frequency and amplitude
of the main peak (EHG and HR), HR quantiles and max
autocorrelation of the HR signal.
C. Laboratory models
Generalized linear models (GLMs) to estimate artifacts
and labour probability were developed using datasets de-
scribed in Sec. II and features listed in Sec. III-B. Models
were derived and validated using leave one participant out
cross-validation and a binary classification problem distin-
guishing artifacts from non-artifacts and labour from non-
labour recordings. According to standard notation, our gen-
eralized linear models can be defined as µi=g(ηi)where g
is the link function, in our case the logit function as errors
follow a binomial distribution, given our two classes prob-
lem. The transformed expected value µis a linear function
of the predictors and can be defined as µi=xiβwhere xiis
the array of EHG, HR and GA features previously described
and βare the model coefficients determined during training.
D. Free-living models
GLMs described in Sec. III-C were deployed in free-
living. As GLMs produce an output probability, a threshold
needs to be selected in order to transform the output into a
discrete class. We set our threshold for artifact probability
estimation models to 0.5, so that an artifact is detected when
the probabilistic output is greater than 0.5. As artifacts are
analyzed on a window by window basis, each 20 minutes
segment is further analyzed to determine the percentage
of detected artifacts. Given the high influence of artifacts
on physiological data such as EHG and HR, we used a
conservative threshold in which only recordings that were
at least 85% artifact-free were considered for our labour
probability estimation. No cut-off threshold was applied to
GLMs for labour probability estimation, as only probabilities
were evaluated at this stage.
IV. RESULTS
A. Laboratory results
1) Artifacts probability estimation: Artifacts probability
estimation results are shown in Fig. 5 for the entire dataset
and a series of activities performed in laboratory settings.
Rest includes both periods of actual rest and clean labour
recordings. The average artifacts probability for reference
artifacts was 63% while it was 32% for reference rest and
labour data (medians were 62% and 31% respectively). We
can see from Fig. 5 that some activities elicit high artifacts
probability (e.g. sit to stand), while others have less impact
on signal quality (e.g. talking).
0.00
0.25
0.50
0.75
1.00
abdom_contr
adl
adl task
bend
coughing
rectal_contr
rest
rubbing
rubbing belly
sit_stand
stretching
talking
toddler
turning
turning to side
walking
Reference artifacts
Probability
Artifacts detection by task − probability
Fig. 5. Output probability of the artifact probability estimation models
showing consistent results with respect to reference data across different
activities.
● ●
0.00
0.25
0.50
0.75
1.00
0 1 2 3 4 5 6 7 8 9 10 11 12
Time to delivery (weeks)
Probability
Labour
FALSE
TRUE
Probability of being in labour (HR + EHG + GA)
Fig. 6. Output probability of the labour detection models when applied to free-living recordings of 142 individual pregnancies, ordered by time to delivery
in weeks.
2) Labour probability estimation: Labour probability es-
timation results are shown in Fig. 7. The average labour
probability for reference labour recordings was 67% while
it was 20% for reference pregnancy recordings collected in
laboratory settings (medians were 94% and 0% respectively).
B. Free-living results
11858 twenty minutes sessions were collected, 3042 of
which were deemed high quality according to our criteria
described in Sec. III-D (26%). 142 women were included,
with an average of 7hours of clean data per person. Labour
probability estimation results are shown in Fig. 6, ordered by
time to delivery, as retrospectively acquired from the study
participants. The average labour probability for reference
labour recordings was 54% while it was 21% for reference
pregnancy recordings (medians were 67% and 0% respec-
tively). Even when analyzing only the last week of data,
labour probability could discriminate between conditions, as
the average labour probability for reference labour recordings
was 54% while it was only 31% for reference pregnancy
recordings (medians were 67% and 11% respectively).
0.00
0.25
0.50
0.75
1.00
FALSE TRUE
Reference labour class
Probability
Labour detection − probability
Fig. 7. Output probability of the labour detection models showing
consistent results with respect to reference data.
V. CONCLUSIONS
During labour, greater physiological changes occur with
respect to pregnancy. As the uterus prepares to expel the
fetus, changes in both EHG and cardiac activity were
highlighted in previous literature in correspondence with
uterine contractions. Additional challenges arise when taking
laboratory-based models to unsupervised free-living settings,
mainly due to lack of context, increased noise, artifacts and
possible misuse of the system. In this paper we proposed an
approach consisting in the development of laboratory based
models aiming at detecting not only the output of interest,
i.e. labour, but also specifically addressing signal quality and
artifacts detection. While challenges linked to signal quality
and proper usage of the device remain, results showed that
non-invasive labour detection in free-living seems promising.
VI. ACKN OWLEDGE MEN TS
This project has received funding from the European
Union’s Horizon 2020 research and innovation programme
under grant agreement No 778503.
REFERENCES
[1] E. R. Norwitz, J. N. Robinson, and J. R. Challis, “The control of labor,
New England Journal of Medicine, vol. 341, no. 9, pp. 660–666, 1999.
[2] C. Urquhart, R. Currell, F. Harlow, and L. Callow, “Home uterine
monitoring for detecting preterm labour,The Cochrane Library, 2012.
[3] J. Marshall, “Regulation of activity in uterine smooth muscle.Physi-
ological Reviews. Supplement, vol. 5, p. 213, 1962.
[4] M. Altini, E. Rossetti, M. Rooijakkers, J. Penders, D. Lanssens,
L. Grieten, and W. Gyselaers, “Combining electrohysterography and
heart rate data to detect labour,” in Biomedical & Health Informatics
(BHI), 2017 IEEE EMBS International Conference on. IEEE, 2017,
pp. 149–152.
[5] M. J. Rooijakkers, C. Rabotti, S. G. Oei, and M. Mischi, “Low-
complexity r-peak detection for ambulatory fetal monitoring,Physi-
ological measurement, vol. 33, no. 7, p. 1135, 2012.
[6] Y. Ye-Lin, J. Garcia-Casado, G. Prats-Boluda, J. Alberola-Rubio, and
A. Perales, “Automatic identification of motion artifacts in ehg record-
ing for robust analysis of uterine contractions,” Computational and
mathematical methods in medicine, vol. 2014, 2014.
... Nonetheless, in contrast to the single-cell techniques that have been the mainstay of uterine electrophysiology for decades, MEA has the potential to significantly advance our understanding of the spatial and temporal dimensions of electrical connectivity in the uterus. Moreover, MEA, in combination with emerging uterine monitoring technologies [22][23][24][25], may provide both physiological and mechanistic ways to understand the basis of uterine electrical activity. Our findings suggest that MEA can reliably differentiate between physiological states in the uterus, and thus, is a novel tool to study normal uterine physiology, uterine pathophysiology, and potential therapeutic strategies. ...
Article
Uterine contractions are important for various functions of the female reproductive cycle. Contractions are generated, in part, by electrical coupling of smooth muscle cells of the myometrium, the main muscle layer of the uterus. Aberrant myometrial electrical activity can lead to uterine dysfunction. To better understand and treat conditions associated with aberrant activity, it is crucial to understand the mechanisms that underlie normal activity. Here, we used microelectrode array (MEA) to simultaneously record and characterize myometrial electrical activities at high spatial and temporal resolution. Mouse myometrial longitudinal muscle tissue was isolated at different stages throughout the estrous cycle and placed on an 8x8 MEA. Electrical activity was recorded for 10 minutes at a sampling rate of 12.5 kHz. We used a spike-tracking algorithm to independently analyze each channel and developed a pipeline to quantify the amplitude, duration, frequency, and synchronicity of the electrical activities. Electrical activities in estrus were more synchronous, shorter duration, higher frequency, and lower amplitude than electrical activities in non-estrus. We conclude that MEA can be used to detect differential patterns of myometrial electrical activity in distinct estrous cycle stages. In the future, this methodology can be used to assess different physiological and pathological states and evaluate therapeutic agents that regulate uterine function.
Article
It is estimated that more than 1 in 10 babies are born prematurely worldwide. Moreover, premature babies can face lifelong health-related disabilities, such as difficulties in learning or hearing and vision loss. By monitoring uterine contractions, physicians can evaluate the health and progress of the pregnancy and determine if the pregnant woman is in labour, thus assisting them to go to the hospital and help reduce the complications associated with premature birth. In this paper, we aim to mitigate the consequences of premature birth for pregnant women and the fetus using machine learning and deep learning approaches to detect and predict labour. The proposed system was tested for reliability and accuracy. The results show that the deep learning approach achieved the best results with a 0.98 accuracy rate.
Article
Full-text available
In this paper we propose a method combining electrohysterography (EHG) and heart rate (HR) data to detect labour. Labour detection may be helpful in providing just in time care and avoiding unnecessary antenatal visits. Given specific changes in physiological data such as EHG and HR highlighted from previous literature in correspondence of uterine contractions, we sought to create a model able to classify between labour and non-labour recordings based on EHG and maternal HR data. In particular, we collected 37 recordings (19 labour and 18 non-labour) from pregnant women at different stages of pregnancy using a wearable sensor designed to be attached to the abdomen using an adhesive patch. We extracted time and frequency domain features from EHG and HR data, as stronger, sinusoidal patterns arise on both data streams in correspondence with uterine contractions during labour. Features were used as input to a random forests classifier, trained to recognize labour and non-labour recordings. The accuracy of the proposed model in classifying labour and non-labour recordings was evaluated using leave one out cross validation. We analyzed results including as predictors; gestational age (GA) only, as reference lower bound (68% accuracy), EHG features only (71% accuracy), HR features only (71% accuracy) and combined EHG and HR data, resulting in 82% accuracy. Inclusion of GA as additional predictor further increased detection accuracy to 79%, 82% and 87% for EHG, HR and combined EHG and HR respectively. Our labour detection model demonstrated a high accuracy in classifying labour and non-labour recordings using EHG and HR data collected using a single wearable device.
Article
Full-text available
Electrohysterography (EHG) is a noninvasive technique for monitoring uterine electrical activity. However, the presence of artifacts in the EHG signal may give rise to erroneous interpretations and make it difficult to extract useful information from these recordings. The aim of this work was to develop an automatic system of segmenting EHG recordings that distinguishes between uterine contractions and artifacts. Firstly, the segmentation is performed using an algorithm that generates the TOCO-like signal derived from the EHG and detects windows with significant changes in amplitude. After that, these segments are classified in two groups: artifacted and nonartifacted signals. To develop a classifier, a total of eleven spectral, temporal, and nonlinear features were calculated from EHG signal windows from 12 women in the first stage of labor that had previously been classified by experts. The combination of characteristics that led to the highest degree of accuracy in detecting artifacts was then determined. The results showed that it is possible to obtain automatic detection of motion artifacts in segmented EHG recordings with a precision of 92.2% using only seven features. The proposed algorithm and classifier together compose a useful tool for analyzing EHG signals and would help to promote clinical applications of this technique.
Article
Full-text available
Non-invasive fetal health monitoring during pregnancy is becoming increasingly important because of the increasing number of high-risk pregnancies. Despite recent advances in signal-processing technology, which have enabled fetal monitoring during pregnancy using abdominal electrocardiogram (ECG) recordings, ubiquitous fetal health monitoring is still unfeasible due to the computational complexity of noise-robust solutions. In this paper, an ECG R-peak detection algorithm for ambulatory R-peak detection is proposed, as part of a fetal ECG detection algorithm. The proposed algorithm is optimized to reduce computational complexity, without reducing the R-peak detection performance compared to the existing R-peak detection schemes. Validation of the algorithm is performed on three manually annotated datasets. With a detection error rate of 0.23%, 1.32% and 9.42% on the MIT/BIH Arrhythmia and in-house maternal and fetal databases, respectively, the detection rate of the proposed algorithm is comparable to the best state-of-the-art algorithms, at a reduced computational complexity.
Article
To reduce the morbidity and mortality associated with preterm birth, home uterine activity monitoring aims for early detection of increased contraction frequency, and early intervention with tocolytic drugs to inhibit labour and prolong pregnancy. However, the effectiveness of such monitoring is disputed. To determine whether home uterine activity monitoring is effective in improving the outcomes for women and their infants considered to be at high risk of preterm birth, when compared with conventional or other care packages that do not include home uterine monitoring. We searched the Cochrane Pregnancy and Childbirth Group's Trials Register (30 November 2011), CENTRAL (The Cochrane Library 2011, Issue 4 of 4), MEDLINE (1966 to 30 November 2011), EMBASE (1974 to 30 November 2011), CINAHL (1982 to 30 November 2011) and scanned reference lists of retrieved studies. Randomised control trials of home uterine activity monitoring, with or without patient education programmes, for women at risk for preterm birth, in comparison to the same care package without home uterine activity monitoring. Two review authors independently assessed trials for inclusion, extracted data and assessed risk of bias. Data were checked for accuracy. We did not attempt to contact authors to resolve queries. There were 15 included studies (total number of enrolled participants 6008); 13 studies contributed data. Women using home uterine monitoring were less likely to experience preterm birth at less than 34 weeks (risk ratio (RR) 0.78; 95% confidence interval (CI) 0.62 to 0.99; three studies, n = 1596; fixed-effect analysis). However, this significant difference was not evident when we carried out a sensitivity analysis, restricting the analysis to studies at low risk of bias based on study quality (RR 0.75; 95% CI 0.57 to 1.00, one study, 1292 women). There was no significant difference in the rate of perinatal mortality (RR 1.22; 95% CI 0.86 to 1.72; two studies, n = 2589).There was no significant difference in the number of preterm births at less than 37 weeks (average RR 0.85; CI 0.72 to 1.01; eight studies, n = 4834; random effects, T(2) = 0.03, I(2) = 68%). Infants born to women using home uterine monitoring were less likely to be admitted to neonatal intensive care unit (average RR 0.77; 95% CI 0.62 to 0.96; five studies, n = 2367; random-effects, T(2) = 0.02, I(2) = 32%). Although this difference was not statistically significant when only high quality studies were included (RR 0.86; 95% CI 0.74 to 1.01; one study, n = 1292). Women using home uterine monitoring made more unscheduled antenatal visits (mean difference (MD) 0.49; 95% CI 0.39 to 0.62; two studies, n = 2807). Women using home uterine monitoring were also more likely to have prophylactic tocolytic drug therapy (average RR 1.21; 95% CI 1.01 to 1.45; seven studies, n = 4316; random-effects. T(2) = 0.03, I(2) = 62%) but this difference was no longer significant when the analysis was restricted to high quality studies (average RR 1.22; 95% CI 0.90 to 1.65, three studies, n = 3749,random effects, T(2) = 0.05, I(2) = 76%). One small study reported that the home uterine monitoring group spent fewer days in hospital antenatally. No data on maternal anxiety or acceptability were found. Home uterine monitoring may result in fewer admissions to a neonatal intensive care unit but more unscheduled antenatal visits and tocolytic treatment. There is no impact on maternal and perinatal outcomes such as perinatal mortality or incidence of preterm birth.
Article
Labor is the physiologic process by which a fetus is expelled from the uterus to the outside world. Labor is defined as an increase in myometrial activity or, more precisely, a switch in the pattern of myometrial contractility from irregular contractures (long-lasting, low-frequency activity) to regular contractions (high-intensity, high-frequency activity),1 resulting in effacement and dilatation of the uterine cervix. In normal labor, there appears to be a time-dependent relation between the biochemical changes in the connective tissue in the cervix that usually precede uterine contractions and cervical dilatation. All these events usually occur before the spontaneous rupture of the fetal . . .