Context
Starting point from Mike’s email (25th January):
Focussed on avoidable harm, like hypoxic ischaemic encephalopathy (HIE)
Ultimate target: to build a machine learning algorithm that monitors live data during childbirth, accounts for pre-existing risk factors, and alerts clinicians when infant health is at increased risk if intervention is delayed or missed
How well can we predict poor neonatal outcomes?
Can we predict the likely effect of intervention (ceasarean) when predicted outcome is poor?
Can we predict the likely effect of intervention (ceasarean) dependent on timing during labour, when predicted outcome is poor?
Concern: Treatment paradox - that predictors of poor outcome (e.g. FHR) will trigger an effective intervention (caesarean), impacting relationship with outcome, and making it hard to understand what was associated with poor outcomes or not.
Why do we want to do causal inference? To understand true cases where they needed a ceasarean and it prevented poor outcomes, or cases where it wasn’t needed. We’re not interested in just being able to predict what happens - we’re interested in the causal relationship, and ultimately a counterfactual of when should we and when should we not do caesareans: Should we have done a caesarean? If we had done one, or if we had done one earlier, would it have prevented HIE?
So what’s our research question? In causal inference, you start with drawing a DAG, which beings with your research question, which should be composed of an exposure/treatment and an outcome, for which you are interested in their causal relationship.
Proposed DAG
Research question
If we’re interested in whether we should do a caesarean to prevent HIE, then proposed research question is:
flowchart LR;
A("Treatment<br><b>Caesarean</b>"):::green;
Y("Outcome<br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
A --> Y;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Add confounders
We want to add confounders (which cause both the treatment and the outcome). These can be measured or unmeasured.
Hypoxia is an unmeasured confounder, but it is indicated by abnormal fetal heart rate (FHR), which then triggers a decision to do a caesarean
Sentinal events - since they cause hypoxia and can also cause decision to do a caesarean - not certain if these are “measured” or indicated by something else
flowchart LR;
L_sen("Confounder (L)<br><b>Sentinal event</b>"):::white;
L_fhr("Confounder (L)<br><b>Abnormal FHR</b>"):::white;
U("Unmeasured confounder (U)<br><b>Hypoxia/Asphyxia</b>"):::white;
A("Treatment (A)<br><b>Caesarean</b>"):::green;
Y("Outcome (Y)<br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
L_sen --> U; L_sen --> A;
A --> Y;
U --> L_fhr; L_fhr --> A; U --> Y;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Add other causes of at least two things in the DAG
flowchart LR;
L_sen("Confounder (L)<br><b>Sentinal event</b>"):::white;
L_fhr("Confounder (L)<br><b>Abnormal FHR</b>"):::white;
U("Unmeasured confounder (U)<br><b>Hypoxia/Asphyxia</b>"):::white;
A("Treatment (A)<br><b>Caesarean</b>"):::green;
Y("Outcome (Y)<br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
age("Gestational age<br>BMI"):::white;
age --> L_sen; age --> A;
L_sen --> U; L_sen --> A;
A --> Y;
U --> L_fhr; L_fhr --> A; U --> Y;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Timing of the caesarean
I think there might be two ways to view timing of the caesarean:
As a moderator
As a time-varying treatment
Since this is about timing of the treatment, rather than some other factors that moderates effectiveness of treatment, I think the latter is a more appropriate way of handling this.
flowchart LR;
age("Gestational age<br>BMI"):::white;
AK("A<sub>K</sub><br><b>Caesarean</b>"):::green;
AK1("A<sub>K+1</sub><br><b>Caesarean</b>"):::green;
YK1("Y<sub>K+1</sub><br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
YK2("Y<sub>K+2</sub><br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
LK("Confounder (L<sub>K</sub>)<br>Abnormal FHR"):::white;
LK1("Confounder (L<sub>K+1</sub>)<br>Abnormal FHR"):::white;
sen_K("Sentinal event<br>at timepoint K"):::white;
sen_K1("Sentinal event<br>at timepoint K+1"):::white;
UK("U<sub>K</sub><br><b>Hypoxia/Asphyxia</b>"):::white;
UK1("U<sub>K+1</sub><br><b>Hypoxia/Asphyxia</b>"):::white;
age --> sen_K; age --> sen_K1; age --> AK; age --> AK1;
AK --> YK1;
LK --> AK;
sen_K --> UK;
UK --> LK;
UK --> YK1;
sen_K1 --> UK1;
UK1 --> LK1;
LK1 --> AK1;
AK1 --> YK2;
UK1 --> YK2;
sen_K --> sen_K1;
UK --> UK1;
YK1 --> YK2;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Is there treatment-confounder feedback?
Not sure.
Here, its about whether or not they did a caesarean.
If they don’t do a ceasarean, we expect things to get worse, FHR more abnormal, so no caesarean impacts later heart rate
HOWEVER if they did do a caesarean, this will have no impact on later FHR, as there is no later FHR - it has ended
So is it treatment-confounder feedback or not?
flowchart LR;
age("Gestational age<br>BMI"):::white;
AK("A<sub>K</sub><br><b>Caesarean</b>"):::green;
AK1("A<sub>K+1</sub><br><b>Caesarean</b>"):::green;
YK1("Y<sub>K+1</sub><br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
YK2("Y<sub>K+2</sub><br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::green;
LK("Confounder (L<sub>K</sub>)<br>Abnormal FHR"):::white;
LK1("Confounder (L<sub>K+1</sub>)<br>Abnormal FHR"):::white;
sen_K("Sentinal event<br>at timepoint K"):::white;
sen_K1("Sentinal event<br>at timepoint K+1"):::white;
UK("U<sub>K</sub><br><b>Hypoxia/Asphyxia</b>"):::white;
UK1("U<sub>K+1</sub><br><b>Hypoxia/Asphyxia</b>"):::white;
age --> sen_K; age --> sen_K1; age --> AK; age --> AK1;
AK --> YK1;
LK --> AK;
sen_K --> UK;
UK --> LK;
UK --> YK1;
sen_K1 --> UK1;
UK1 --> LK1;
LK1 --> AK1;
AK1 --> YK2;
UK1 --> YK2;
sen_K --> sen_K1;
UK --> UK1;
YK1 --> YK2;
AK --> LK1;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
What about all our other risk factors we’re interested in?
In a DAG, we should only include common causes. However, there are lots of risk factors that we often consider that I’m not sure if they are common causes e.g.
Parity
Meconium
Analgesia
Ethnicity
What do we do about those? Do we care? Should it be that I just haven’t managed to fit them on the DAG? If they don’t fit on, do we not care?
Diagnostic error
You could also include representation of the diagnostic error in HIE - ignoring the rest of the DAG, would look like…
Another representation with diagnostic error in HIE…
flowchart LR;
Y("Outcome (Y)<br><b>Hypoxic ischaemic<br>encephalopathy (HIE)</b>"):::important;
Y*("Y* Mismeasurement of HIE"):::white;
Yerr("Measurement error in HIE"):::white;
Y --> Y*;
Yerr --> Y*;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef empty width:0px,height:0px;
classDef important fill:#DDF2D1, stroke: #FFFFFF;
Another (less certain) proposal
Another angle (not sure if this is right)…
We know gestational age is a risk factor for HIE - but does it have a causal relationship with HIE?
flowchart LR;
%% Define the nodes and subgraphs
exp("Gestational age"):::green;
out("HIE"):::green;
%% Produce the figure
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Would we then be interested in the direct causal effect of age on HIE that is not mediated by having a caesarean? (i.e. treating ceasarean as a mediator of the relationship between age and HIE)
flowchart LR;
%% Define the nodes and subgraphs
exp("Gestational age"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Expanding on that, there are many risk factors of interest…
flowchart LR;
%% Define the nodes and subgraphs
exp("Gestational age"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Hypoxia"):::green;
fhr("Abnormal FHR"):::white;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> fhr;
fhr --> treat;
treat --> out;
fhr --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Parity"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Previous caesarean"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("BMI"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Ethnicity"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Analgesia"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Meconium"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
exp("Parity"):::green;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
exp --> treat;
treat --> out;
exp --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
But as they are all causes of two things on the DAG, you’d need to add them all to each DAG….
flowchart LR;
%% Define the nodes and subgraphs
age("Gestational age"):::green;
par("Parity"):::white;
hypo("Hypoxia"):::white;
fhr("Abnormal FHR"):::white;
mec("Meconium"):::white;
ana("Analgesia"):::white;
eth("Ethnicity"):::white;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
age --> treat; age --> out;
par --> treat; par --> out;
hypo--> fhr; fhr --> treat; fhr --> out;
mec --> treat; mec --> out;
ana --> treat; ana --> out;
eth --> treat; eth --> out;
treat --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
flowchart LR;
%% Define the nodes and subgraphs
age("Gestational age"):::white;
par("Parity"):::green;
hypo("Hypoxia"):::white;
fhr("Abnormal FHR"):::white;
mec("Meconium"):::white;
ana("Analgesia"):::white;
eth("Ethnicity"):::white;
treat("Caesarean"):::white;
out("HIE"):::green;
%% Produce the figure
age --> treat; age --> out;
par --> treat; par --> out;
hypo--> fhr; fhr --> treat; fhr --> out;
mec --> treat; mec --> out;
ana --> treat; ana --> out;
eth --> treat; eth --> out;
treat --> out;
classDef white fill:#FFFFFF, stroke:#FFFFFF;
classDef black fill:#FFFFFF, stroke:#000000;
classDef green fill:#DDF2D1, stroke: #FFFFFF;
Context: Reasons for doing a caesarean
A key thing to know is what causes need for caesarean - and whether those are also things that cause HIE (similar to hypoxia).
Panda et al. 2018 - Clinicians’ views of factors influencing decision-making for caesarean section: A systematic review and metasynthesis of qualitative, quantitative and mixed methods studies - https://doi.org/10.1371%2Fjournal.pone.0200941 - some of the mentioned reasons across the world…
Perceived risk of CS v.s. vaginal birth - e.g. concerns related to risk of urinary and fecal incontinence and pelvic floor collapse following vaginal births
Perceived safety of CS - e.g. believing CS could reduce risks of and prevent complications for women living in isolated areas
Women requesting CS (which itself is influenced by socio-cultural perspectives, women’s preferences, demands, obstetricians’ beliefs in women’s right and autonomy to choose a CS, and their perception of women’s anxiety and fear)
Clinical reasons -
previous CS
risk of anorectal trauma
preventing perineal injury, urinary and anal incontinence
maternal age
obesity
previous birth complications
risk of pelvic prolapse
uterine rupture
medical conditions like mypoia and previous abortions
breech presentation
previous classical CS
fetal distress
malpresentation
dystocia
placenta previa
umbilical cord collapse
Fear of litigation / medico-legal problems
Lack of resources (e.g. not enough experienced clinicians to facilitate natural birth, availability of personal for emergency CS, availability of anaesthesia, access to basic infrastructure, access to emergency care facilities)
Type of health care coverage (private/public)
Hospital policies
Clinician characteristics - personal convenience, clinician demographics, confidence and skills
