# HIE Proposal (risk factors)

This is just **an example** of how we could design this study. It requires further thought, with regards to whether this is the chosen focus of the study, and whether these are the methods we would to use, as summarised on the page ['Focus of the research'](./1_finding_focus.md)

In this proposal, I am assuming a focus of **retrospective cross-sectional identification of risk factors for HIE (whilst accounting for the impact of caesarean section on the observed relationships)**.

## Introduction

Neonatal encephalopathy (NE) refers to brain disease, damage or malfunction in late pre-term or term infants (born at or beyond 35 weeks). It is characterised by 'challenges initiating and maintaining respiration, reduced tone and reflexes, seizures, and impaired levels of consciousness'.{cite}`molloy_neonatal_2023,quirke_cohesion_2024` The cause of NE is often difficult to identify, with careful phenotyping using a range of measurements and examinations required to differentiate between possible causes. Morover, there may not just be one cause, but several factors that lead to NE.{cite}`molloy_neonatal_2023` The possible causes of NE include: hypoxia ischaemia, neonatal stroke, infection, intracranial haemorrhage, congenital brain anomalies, as well as neurometabolic, genetic and epigenetic factors.{cite}`quirke_cohesion_2024,austin_guidelines_2021`

Hypoxia ischaemia is the most common cause of NE. This is when there has been restricted blood flow to tissues ("ischaemia"), meaning that tissues that have low/insufficient oxygen levels ("hypoxia").{cite}`klabunde_ischemia_2023` This type of NE is referred to as hypoxic ischaemic encephalopathy (HIE), but was often previously referred to as birth/perinatal asphyxia.{cite}`bliss_hypoxic-ischaemic_nodate` The current treatment for infants with suspected HIE is therapeutic hypothermia. This must be started within six hours of birth,{cite}`parmentier_magnetic_2022` and involves cooling of the infant's head or whole body to reduce their body temperature to 33-36.5°C for 48 to 72 hours.{cite}`allen_hypoxic_2011` Potential consequences of moderate/severe HIE include mortality, and the development of cerebral palsy, epilepsy, and intellectual disability.{cite}`shim_which_2021,torn_outcomes_2023`

There are several risk factors associated with HIE. These include maternal and pregnancy characteristics which indicate that a delivery is at higher risk of HIE occuring, as well as observations during labour which indicate HIE may have occured/is more likely to have occurred. Risk factors for HIE during labour include: decreased fetal movement; abnormal fetal heart rate or contractions patterns; severe maternal cramping, maternal high blood pressure, or vaginal bleeding. However, in some cases (particularly infants with mild to moderate HIE), there may not be any obvious signs or symptoms of HIE during labour or at the time of birth.{cite}`hie_help_center_hypoxic-ischemic_nodate`

If a clinician suspects that HIE is likely, they may choose to perform an emergency caesarean section, based on the rationale that if an infant is truly experiencing HIE, sooner delivery would reduce the severity of HIE by reducing the length of time that the infant experiences hypoxia ischaemia. However, in the United States, a third of malpractice claims in obstetrics are related to neurological injuries in neonates, with factors leading to these including that clinicians have failed to account for risk factors known before birth, and failure to identify or act on (or delays in acting on) identified risk factors during birth like abnormal fetal heart rate, fetal distress, and malpresentation.{cite}`leith_predictive_2024`

### Existing studies

There have been some studies that have created predictive models for HIE which incorporate a wide range of known risk factors. The rationale/hope of these studies is that the development of models that can combined information on risk to generate an overall estimate of risk, can then help clinicians identify high-risk pregnancies and plan delivery accordingly to hopefully prevent HIE.

Odd et al. 2017{cite}`odd_hypoxic-ischemic_2017` use a logistic regression model to predict cases of HIE (i.e. infants who need resuscitation and had NE symptoms). It is based on data from over 14,000 infants in a Bristol cohort study, with half used for training and half for validation. In the training set (n=6712), 130 infants had HIE (0.01%). The model includes a range of maternal, antepartum and intrapartum characteristics - although it does not include the delivery method. However, they do estimate the number of elective caesareans that would be required to prevent one case of HIE in different antenatal risk groups.{cite}`odd_hypoxic-ischemic_2017`

Leith et al. 2024{cite}`leith_predictive_2024` use a logistic regression model to predict cases of HIE (i.e. presence of HIE diagnostic code) in a US dataset of 836,216 births - of which, 376 (0.00045%) had a diagnosis of HIE. The model includes a range of maternal, antepartum and intrapartum characteristics - including the delivery type (with predictors including forceps, vacuum, elective C-section, emergency C-section, and intrapartum C-section).{cite}`leith_predictive_2024`

Abnormal fetal heart rate is an important indicator which we'd hope to include in our analysis, but is not included in Odd et al. 2017{cite}`odd_hypoxic-ischemic_2017`, and is included via identification of an ICD code for fetal heart rate abnormality in Leith et al. 2024{cite}`leith_predictive_2024`. There is a study by Eden et al. 2018{cite}`eden_fetal_2017` which includes more detailed information on fetal heart rate. They used fetal heart rate, baseline variability, accelerations, decelerations and uterine activity - in combination with other maternal, obstetrical and fetal risk factors - to produce a risk score. This score was to distinguish between infants with intrapartum neurological injury and subsequent cerebral palsy, and control term infants who had a normal fetal tracing on admission.{cite}`eden_fetal_2017`

### Treatment use

When predicting HIE, it is important to account for treatment use. This is because we have a treatment paradox, where classical prognostic factors for HIE (like abnormal fetal heart rate) trigger an effective intervention (caesarean section).

As such, a predictive model from this partially treated population would actually be comparing outcomes between:
1. Infants without HIE - composed of:
    * (a) infants who never would have had HIE, and
    * (b) infants who would have received HIE but received a caesarean section (whether due to a classical prognostic factor or otherwise).
2. Infants with HIE - composed of:
    * (a) infants with classical prognostic factor for whom there was no or a delayed caesarean, and
    * (b) infants without classical prognostic factors.

This is a problem as we want to apply the model to an untreated population, and be simply comparing infants who would have HIE and infants who would not.

The consequence of this treatment paradox is that **good predictors in the untreated population will look like poor predictors in the partially treated population**. As such, it is important we account for treatment use when making predictions.

### Aim

Predict HIE using explainable machine learning models (so we can see what factors increased risk for HIE in the predictive model) - but account for treatment use (caesarean) whilst analysing those relationships.

## Methods

### Inclusion and exclusion criteria

Inclusion criteria:
* Infants born at or beyond 35 weeks of gestation - *as that's the minimum age for diagnosis of neonatal encephalopathy*
* Singleton pregnancies - *only recommending as that's a common criteria similar studies set*
* HIE or no NE

Exclusion criteria:
* NE that is not caused by hypoxia ischaemia

### Measures

#### Outcome: HIE

HIE identified based on the method defined by the Neonatal Data Analysis Unit and the Department of Health, which is the presence of either:
* Diagnosis of NE or HIE (severe, moderate, grade 2 or grade 3)
* Therapeutic hypothermia for 2 or more consecutive days{cite}`gale_brain_2017`

#### Timing of caesarean

As [explored here](./1_finding_focus.md), timing of caesarean related to either:
* **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.

### Analysis

A sensible first step might be to demonstrate the existence of a treatment paradox, as in Uffen et al. 2021.{cite}`uffen_retrospective_2021`

It would be beneficial to consider a range of methods that can be used to answer the same problem. On the [treatment paradox](./2_treatment_paradox.md), I outline some of the suggested methods for accounting for treatment use. These are commented on for this study below...
* **Standardisation of treatment** - potentially suitable, although only through statistical techniques that achieve this, as it is not true in practice - the decision to perform a caesarean and the timing of that decision varies between clinicians
* **Predictor substitution** - unsuitable
* **Treatment as predictor** - potentially suitable
* **Treatment as outcome** - potentially unsuitable, assuming I am correct in understanding that there are many possible reasons why a clinician might perform an emergency caesarean (and it is not just due to suspect HIE).
* **Propensity scores** - potentially suitable
* **Exclusion of treated individuals** - potentially suitable (although may be unsuitable if sample size is too small)

### Limitations

* There may be insufficient HIE cases in the dataset
* This analysis is retrospective and not designed in a way that allows "live" predictions
* Haven't specified how the CTG data would be incorporated
* Limitations of particular methods (e.g. as on [treatment paradox](./2_treatment_paradox.md) page)
* Although we've attempted to account for treatment use, we (a) may not be successful, and (b) will still have other sources of bias such as omitted variable bias

### Other analyses

Given the concern here about accounting for treatment use, it might be wise to consider demonstrating that a paradox does indeed exist, as they do in [Uffen et al. 2021](https://doi.org/10.1136%2Fbmjopen-2020-046518).