# Focus of the research :::{toctree} :hidden: :caption: Things to consider self 2_treatment_paradox ::: :::{toctree} :hidden: :caption: Previous studies 3_examples_risk 4_examples_intervene ::: :::{toctree} :hidden: :caption: Example research proposals 5_proposal_1 6_proposal_2 references ::: `````{admonition} Executive summary :class: info The aim of this research is currently quite broad. There are several things to consider when clarifying the focus of this work. **What we hope to learn** - e.g. timing of caesarean, where caesarean was needed, where caesarean was not needed, where forceps or ventouse deliveries would be unsuccessful, understanding inequalities. **Outcome** - Assuming our aim is still a bit broad, need to choose a suitable outcome, and thinking about why we perform caesareans can help with that. For the outcome itself, things to consider include: neonatal and/or maternal; conditions, interventions and/or clinical observations; number of outcomes and whether seperate or composite; how outcomes are measured; expert feedback. **Treatment/intervention** - which interventions, how many you compare, whether you also include treatments after birth. **Methods** - considering the timing of events; which predictors we include; how we analyse cardiotocograph data; how we identify risk factors; how we assess intervention effectiveness. ````` ## Stated aim of this research As stated in the protocol for the ethics application, the ultimate aim of this project is to produce a tool that provides **real-time, data-driven risk assessment during labour, using machine learning to avoid adverse neonatal and maternal outcomes**. Key components of this work are: 1. Identifying an appropriate method of analysis for the **cardiotocograph data** 2. Identifying **risk factors** (maternal characteristics during labour, and risk factors that develop during labour) that influence maternal and neonatal outcomes 3. Identifying when **intervention** during labour would beneficial (as that is the natural recommendation you would be making off the back of such a tool) Supporting this is the fact that: * Reliable methods of interpreting CTG in labour have not yet been developed * Current studies of risk factors for adverse outcomes do not use live data to predict outcomes during labour, and to modify the pre-labour risk [Protocol] ## What do we hope to learn? The stated aim above is quite general and, as such, there have been a wide range of suggestions regarding what we might hope to learn from this analysis (such as in the expression of interest, and ethics protocol). This have included reference to: * Comparing outcomes from labours with similar characteristics except that there was an **earlier or later** decision to do a caesarean * Identifying infants who **did not receive caesareans** but who would have benefited * Reducing risk of **unecessary caesarean** sections - with inherent maternal complications like infection, bleeding and bladder damage. * Identifying women in whom **forceps or ventouse deliveries would be unsuccessful** and reduce the risk of traumatic assisted vaginal deliveries which lead to adverse neonatal outcomes * Focus on **inequalities** in a particular treatment/outcome. Death and disability in pregnancy are much higher in disadvantaged groups. For example, stillbirth is up to four times higher in black than white women in the UK. Depending on what we hope to learn, we will find a more specific outcome, as well as relevant methods and treatments to consider. ## Outcome ### Outcome: Why we do perform caesareans? Given the broad stated aim of this project, there are a range of possibile choices for the outcome. Thinking about the range of possible outcomes is therefore a good starting point. As this research revolves around when we should use caesareans to prevent poor neonatal outcomes, it would be good to understand the various reasons **why caesareans are performed** (and so, **what they are trying to prevent**). To be clear, I'm not referring to risk factors that mean a poor outcome is more likely - examples from [Panda et al. 2018](https://doi.org/10.1371%2Fjournal.pone.0200941) * Before labour: * Previous caesarean * Maternal age * Obesity * Previous birth complications * Medical conditions like mypoia and previous abortions * During labour: * Uterine rupture * Breech presentation * Fetal distess * Malpresentation * Dystocia (obstructed labour - pelvis physically blocked) [[Panda et al. 2018]](https://doi.org/10.1371%2Fjournal.pone.0200941) Instead, I'm referring to the outcomes themselves, that they are trying to prevent. Some examples I've compiled (which may or may not be outcomes that relate to caesarean - wasn't always 100% sure, but [lots more that may or may not be here](https://www.birthinjuryhelpcenter.org/birth-injuries.html))... > Maternal outcomes: > * Perineal injury or anorectal trauma > * Urinary and fecal incontinence > * Pelvic floor collapse > * Pelvic prolapse [[Panda et al. 2018]](https://doi.org/10.1371%2Fjournal.pone.0200941) > * Post-partum haemorrhage (PPH) [[source]](https://www.marchofdimes.org/find-support/topics/pregnancy/preeclampsia) > * Mortality > > Neonatal outcomes: > * Transfer of a chorioamnionitis to the infant [[source]](https://www.birthinjuryhelpcenter.org/chorioamnionitis.html) > * Hypoxic ischaemic encephalopathy (HIE) [no specific source] > * Neonatal seizures > * Mortality due to complications during labour (e.g. cord prolapse, placental abruption) [[Penn et al. 2001]](https://doi.org/10.1053/beog.2000.0146) > * Erb's palsy / brachial pleus birth palsy (damage to nerves that supply arms and hands, e.g. due to shoulder dystocia) [[source]](https://my.clevelandclinic.org/health/diseases/21986-erbs-palsy) [[source]](https://www.stanfordchildrens.org/en/topic/default?id=birth-injuries-90-P02687) > * Brachial palsy > * Caput succedaneum (swelling of scalp soft tissues) > * Cephalohematoma (bleeding between skull bone and fibrous covering) > * Facial paralysis > * Fractures > * Subconjunctival hemorrhage [[source]](https://www.stanfordchildrens.org/en/topic/default?id=birth-injuries-90-P02687) > > Clinical outcomes: > * Transfer to neonatal care services (*not sure this example is right, but my thinking was along the lines of, examples like infants where they have a stressful birth and poor clinical signs in the first instance and so are watched, but with a little intervention they are shortly discharged? so wouldn't fall under above outcomes but is an outcome?*) ### Outcome: Things to consider > Whether you focus on **neonatal and/or maternal** outcomes > > The **type of outcome** you focus on... > * **Conditions** - e.g. NE, HIE, pre-eclampsia > * **Interventions** - e.g. transfer to neonatal care services, resuscitation, mode of delivery (spontaneous, vacuum, forceps, caesarean), whether rotation is needed > * **Clinical observations** - e.g. chorioamnionitis, perinatal death (antenatal and intrapartum), Apgar (1-, 5-, 10-minute), meconium-stained amniotic fluid, vaginal tearing or episiotomy, blood clots, and urinary or anal incontinence > > **How many** outcomes you focus on - [Cheong-See et al. 2016](https://doi.org/10.1111/1471-0528.13859) explain that obstetric models may often have more than one relevant outcome (e.g. eclampsia, abruption). As (in their example) very few mothers will develop early-onset pre-eclampsia, these outcomes will occur at very low rates. This would not meet the suggested level of having 10-15 outcome events per predictor - and so, they recommend using composite outcomes and competing risk models.[[Cheong-See et al. 2016]](https://doi.org/10.1111/1471-0528.13859) Options: > * **One** outcome > * **Several** seperate outcomes > * A **composite** of several outcomes (which could even combine neonatal + maternal outcomes, and different types of outcome) > > How the chosen outcome/s are **measured** - as some will be quite simple, but others may be harder to define, and require proxy measures or a combination of indicators for that one outcome. This requires careful thought for each outcome. > > These decisions could also involve **expert feedback** which could be quite structured (such as in previous studies, with steering committees and Delphi studies to form consensus opinions) In this book, I have often focussed on HIE as an example. This is simply because of our meeting in January (which largely had revolved around HIE). However, we might not be interested in HIE if - * **We are actually interested in a broader range of outcomes**, or * **There are too few HIE cases** as HIE is a fairly rare outcome. Depending on the criteria used to identify cases, and depending on the size of our dataset, there is a possibility that we have too few cases of HIE for us to use it as an endpoint To contextualise this... * The reported incidence of NE ranges from about 2 to 6 per 1000 live births, and the incidence of HIE ranges from about **1 to 8 per 1000** live births, although these figures have several limitations.[[Kurinczuk et a. 2010]](https://doi.org/10.1016/j.earlhumdev.2010.05.010) * [[Törn et al. 2023]](https://doi.org/10.1111/1471-0528.17533) - using data from Swedish national health and quality registers from **2009 to 2015, 0.0008% had mild HIE diagnosis(n=414, whilst n=504661 had no HIE)** * [[Leith et al. 2024]](https://doi.org/10.1016/j.annepidem.2023.11.011) - using US dataset of **836,216 births, of which 376 (0.00045%) had a diagnosis of HIE** * [[Odd et al. 2017]](https://doi.org/10.3233/NPM-16152) - using Bristol dataset of 14,000 births. In the training set, had **6712 births, of which 130 (0.01%) had HIE** If looking at a broader outcome (and not with the intention of it being an indicator for HIE), this has **implications** for the predictors you should include, and any DAGs, and ultimately what your research question and focus is all about. For example: * If a particular treatment/intervention/process is used as the outcome, then you'd need to think carefully about what that outcome represents - i.e. **all the possible reasons for receiving that intervention** - so they are appropriately included and considered in the model. * If a particular condition is used as the outcome, then you'd need to think carefully about what is influencing the relationships between predictors and the outcome - this might often be about the influence of intervention via caesarean section, but **other treatments** might commonly be used that influence relationships as well. I have not done alot of research into broader outcomes. However, whilst looking into HIE, I have come across examples of papers looking more broadly at outcomes, and examples include: * **Maternal and perinatal complications** - Maternal complication = maternal death, neurological, hepatic, cardiorespiratory, renal or haematological complications, or delivery before 34 weeks. Perinatal complications by discharge = perinatal or infant mortality, bronchopulmonary dysplasia, necrotising enterecolitis, intraventricular haemorrhage, cystic periventricular leukomalacia, retionpathy of prematury, or HIE. [[Thangaratinam et al. 2017]](https://doi.org/10.1186/s12916-017-0827-3) * **Adverse pregnancy outcome** - Composite of: hypertensive disorders of pregnancy, LGA, neonatal hypoglycaemia, shoulder dystocia, fetal death, neonatal death, bone fracture, nerve palsy. [[Cooray et al. 2020]](http://dx.doi.org/10.1136/bmjopen-2020-038845) * **Adverse neonatal outcome** - 5-min Apgar <7, or composite of 5-minute Apgar score <7, resuscitation by intubation and/or perinatal death. [[Steer et al. 2023]](https://doi.org/10.1111/1471-0528.17531) * **Adverse pregnancy outcomes** - delivery by elective or emergency caesarean section,pre-eclampsia, gestational hypertension, small-for-gestational-age, stillbirth, neonatal death and neonatal unit admission for at least 48 hours. [[Syngelaki et al. 2022]](https://doi.org/10.1002/uog.26036) ## Treatment/Intervention We are often concerned about interventions - either in the context of accounting for treatment use when identifying risk factors for adverse outcomes, or in making recommendations about changes in risk when an intervention occurs. Things to consider here include... > **Which interventions** you choose to focus on. These could be specific **interventions during labour or delivery types**. > * Examples of type of delivery include: > * Unassisted vaginal delivery > * Assisted vaginal delivery - ventouse (vacuum cup) or forceps > * C-section - elective or emergency > * Vaginal birth after C-section > * Examples of treatments or other interventions during labour include: > * Gas and air (Enotox) > * Pethidine injections > * Epidural > * Remifentanil > * Water birth > * TENS machine > * Artifical rupture of the membranes > * Oxytocin drip/syntocinon > * Episiotomy > > How you compare these - e.g. compare **two** delivery pathways, or compare **multiple** delivery pathways > > If the outcome is not immediately after birth, whether you focus on **subsequent treatments** - for example, magnesium sulfate and cooling for infants with HIE. ## Methods **Timing** will be very important in this research. Things to consider will include... > How you **model** the evolving risks during labour, and subsequently how you **analyse** data feeding into the model (e.g. analysing CTG data in 15 minute windows, and rerunning predictions) > > How you account for timing when assessing **intervention** effectiveness. The risk-benefit ratio is dynamic and changes during labour depending on how the labour progresses and how events unfold (e.g. caesareans are safer earlier in labour than later). > > How you **measure** the timing during labour. Some possibilities (with limitations): > * Length of labour (Steve: time from start is really variable and inaccurate) > * Contraction rate/force (Steve: Unfortunately, we cannot determine the force of contraction, only the timing which is really important for any changes in the fetal heart rate) > > Some other (perhaps stronger) suggestions: > * **Cervical dilation** (Steve: This is the standard method of determining the length of labour) > * **Length of Stage-II labour** - this would eliminate any caesareans prior to stage 2 (although from talking with Rob, the majority happen later, after a prolonged labour) > * **Presence of multiple (𝑘) risk factors** - this is slightly different to the others, and assumes that time itself is not of interest, but instead, what risk factors accumulate over time, are. Another methodological consideration is **what variables** are included. This will depend on our chosen focus. Things to consider here include - > **Which variables** are considered, which depends on the relationship you focus on, and your analysis method. Examples include: > * Suspected fetal growth restriction (FGR proxy is number of antenatal scans) > * Previous CS > * Parity (number, or nulliparity vs multiparity) > * Antepartum haemorrhage > * Body mass index > * Maternal age > * Maternal ethnicity > * Gestational age at delivery > * Induction of labour > * Analgesia - epidural use > * Oxytocin use > * Maternal pyrexia > * Abnormal FHR > * Meconium-stained amniotic fluid > * Single unsupported mother > * Highest antenatal blood pressure > * Duration of labour > * Fetal size on scan (if done) How **cardiotocograph data** is analysed and incorporated as a predictor is also complex, with a wide range of methods used in the literature, as explored in [CTG exploratory repository](https://github.com/birth-outcomes/ctg_exploratory) and [FHRMA Python repository](https://github.com/birth-outcomes/fhrma_python). For example: > * Generating a set of parameters that describe the signal (clinically relevant or others), and the chosen method for that analysis > * Transforming the signal, such as into an image using continuous wavelet transformation, and the chosen parameters How we account for **treatment use** when identifying risk factors is another important factor [explored on the subsequent page](./2_treatment_paradox.md) Finally, how we assess **intervention effectiveness**. When making recommendations about whether or not to intervene, we are focussed on the true causal relationship between the intervention and the outcome, which necessitates that we consider the use of [causal inference techniques](../causal_concepts/1_predict_vs_causal.md) ## Working with clinicians and parents For any chosen study design and focus, involvement of clinicians, as well as public and patient involvement, will be important. * Public and patient involvement - e.g. parents of infants who experienced adverse outcomes (short or long-term) * Clinicians - e.g. how acceptable a live tool would be, and how co-production could be used to inform design ## Other things to consider ### Sample [Cheong-See et al. 2016](https://doi.org/10.1111/1471-0528.13859) note that models are population-specific. For the example of pre-eclampsia, it is multisystemic, different predictor groups associated with different outcomes, 'case mix impacts on the distribution of the predictors and the prevalence of the outcomes, and this in turn impacts on predictor–outcome associations, thereby influencing the accuracy of the model.' [[Cheong-See et al. 2016]](https://doi.org/10.1111/1471-0528.13859) ### External validation It has been commented that - for obstetric models - there is a lack of external validation (many obstetric models perform well when internally validated but few have been externally validated using a population seperate from development).[[Cheong-See et al. 2016]](http://dx.doi.org/10.1111/1471-0528.13860) * 'Historically, the field of obstetrics has been successful in developing prediction models but has been poor in fully validating and thus implementing them effectively... Only two thirds of the papers [62.4%, 164/263] in a large systematic review of prognostic models in obstetrics were found to have presented their models in such a way that external validation would be feasible. This has been highlighted as a concern given the importance of validity in the development of such models.'[[Murphy et al. 2019]](https://doi.org/10.5772/intechopen.87311) * 'Certain models can be too complex for routine clinical usage and this may lead to a reluctance on the part of the clinicians to accept them... It is also important that models which have been developed are also validated in a new population as otherwise it may not be possible to generalise them to a different cohort of patients. This is also known as impact analysis and this paper by Reilly et al. highlights that very few prediction models have undergone formal impact analysis or validation. This is essential in order for clinicians to know if the usage of such a model will have a positive or negative effect, i.e. is there a possibility that it will cause harm. The authors highlighted the benefit of having clinicians involved in the development and validation of such models before, during and after implementation.'[[Murphy et al. 2019]](https://doi.org/10.5772/intechopen.87311) ### Other statistical challenges As from [Cheong-See et al. 2016](http://dx.doi.org/10.1111/1471-0528.13860): * The ascertainment of a suitable sample size - this is likely to be a problem, particularly if we choose a specific endpoint like HIE * The choice of candidate predictors * Reliable measurement of the outcome and predictors * The identification of important predictors and their functional form * Internal validation, potentially including bootstrap resampling and cross-validation, as well as shrinkage for potential over-optimism in model performance. * Lack of external validation (many obstetric models perform well when internally validated but few have been externally validated using a population seperate from development) * Lack of evaluation of the impact of models on clinician behaviour and patient outcomes * **Handling of interventions in the prediction model** [[Cheong-See et al. 2016]](http://dx.doi.org/10.1111/1471-0528.13860)