6 early career researchers were selected to give a short 4-minute presentation on their exciting new research, on Day 1 of our HDR UK Conference 2024 at 12.15pm. Speakers included:
– Patrick Bidulka (LSHTM): Addressing confounding in analyses of comparative effectiveness using electronic health records: a case study in type 2 diabetes using an instrumental variable.
– Claudia Lindner (University of Manchester): Using AI to enhance efficiency and equality of hip surveillance in children with cerebral palsy.
– Hayley Lowther-Payne (Lancaster University): Understanding access to NHS adult mental health services for sexual minority groups in North West England: an exploratory study using routinely collected data.
– Matthew Watson (Durham University): From Prediction to Practice: Addressing Bias and Data Shift in Machine Learning Models for Chemotherapy-Induced Organ Dysfunction.
– Georgina Ireland (UCL): Linking administrative data from hospitals and family courts for England to estimate the cumulative incidence of involvement in public law family court proceedings and inform healthcare intervention.
Erum Masood (University of Dundee): Mapping Scottish Population Scale COVID-19 and relevant clinical data Health Data to Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM).
This session was chaired by Elliot McClenaghan at London School of Hygiene & Tropical Medicine.
#HDRUKConference
Hi everyone so welcome to the early career researcher lightning talk session so every year hduk wants to shine a spotlight on some of the work of early career researchers in the community and this year is no exception so I’m joined on stage by Patrick Claudia uh haly
Matthew uh Georgina and arum all of whom are early career researchers coming from a really nicely diverse range of disciplines roles uh research top tocs and geographies and all of their research topics fit really nicely in with hdr’s ethos so they beat out quite a tough um group of compet competitors
To get to this stage um there were um many more applications and there were speaking slots um decided by an independent panel um and so let me congratulate them for getting this far but their challenge today and as any of you in this room who have presented research might understand that it really
Is real challenge is to present their work in less than four minutes so it’ll be then it’ll be over to you the audience to uh vote for who you think was the best lightning Talk of the session um there’ll be no uh Q&A set as part of this session but please do um
Grab the researchers and um spark up any conversations or ideas you might have for them over the lunch break um and with that um I’ll introduce our first Speaker Patrick vulka uh so thanks very much for having me uh my name is Patrick Bolan I’m a research
Fellow at the London School of hygiene tropical medicine and I’m also a PhD student I’m submitting my PhD in three weeks so it’s very calm at the office um so uh very quickly I’m going to be talking about type 2 diabetes uh anti-diabetic treatments so people who
Uh have type 2 diabetes often start with meformin monotherapy to manage their blood glucose and then when metformin isn’t enough they often or usually add uh a second line treatment to metformin monotherapy and there are three main Alternatives in the UK um and the problem is that there’s still some
Clinical uncertainty as to which treatment is best for particular patients and so what we end up seeing because of this clinical uncertainty is substantial variation in the proportion of people who are prescribed these treatments across groups of General practices in the UK or in England and so
Each bar in this graph is a group of General practices and you see that the proportion of people prescrib each treatment varies quite substantially and so in this study design uh what we try and do is use this variation in an instrumental variable analysis to try and get as at the treatment
Effect of these three alternative anti-diabetic treatments and so you can think of it kind of like a randomized controlled trial although it’s not a randomized controlled trial but instead of looking at the treatment people actually receive you look at an instrument that’s Upstream of this treatment assignment um
And that is unrelated to the confounders that normally bias uh our treatment effects and observational studies and so here we’re using that variation as the prescribing history of the group of the GPS um to then estimate this treatment effect in dependent of measured and unmeasured confounders and of course
We’re making very strong assumptions in this analysis uh which I don’t have time to go into today and so I’m cutting straight to the results here um but we had 75,000 people from UK primary and secondary care we use clinical practice research data link uh link with Hospital episode statistics
We looked at a variety of outcomes that are important to patients with diabetes and so uh to just summarize this Forest plot we we did find that the newer type of drug sodium glucose co-transporter 2 Inhibitors were better at reducing blood blood sugar body mass index and blood pressure compared to the
Alternatives so we had good evidence of this um sorry these graphs are hard to read and that’s why I’m sum mizing them in these tables but when we look at cardiovascular and kidney outcomes and death we observe sort of uh less less strong evidence um in terms of these
Outcomes we see that sglt2i were better than the Alternatives at preventing heart failure hospitalization we saw some evidence that these drugs were better at prevent slowing the decline of kidney function but we didn’t see much evidence to suggest that they were better than the alternatives for other cardiovascular
And kidney outcomes and all cause mortality so to very quickly conclude um using the instrumental variable analysis we did get some insight into um the comparative effectiveness of these drugs and seeing that sglt2i were better than the alternatives for some treatment or for some outcomes among a general
Population of people with anti with diabetes when we compare these results to other studies which use more traditional methods but assume no unobserved confounding and other studied designs like Placebo controlled rcts we do see um similar but less strong benefits of these drugs and so this study design can be a
Useful complement to other study designs and randomized control trials using routinely collected Health Data to get at sort of understanding treatment effects and informing clinical practice so just thank you everybody for your attention and thanks to my study team who um I couldn’t do the work without no go
Ahead my name is Claudia I’m on from Manchester and I’d like to thank hduk for the opportunity to talk about our work on using artificial intelligence to improve the hip surveillance for children with C poy why is this important this Advance this little boy here has got sery which is the most common
Physical disability in children and having C poy this means this Bo will have all kind of Health implications one of which being that he’s got a very high risk that hip Bor will dislocate now if this happened he would have severe pain he will have problems sitting and he
Also will require very complex invasive and costly surgery now hip dislocation means that the hip bone has moved out migrated out of the hip socket and insur C pory this happens because abnormal massive forces keep pulling on that bone now if the doctors were able to identify early that a bone
Is moving out of the socket they can could then offer those children less less complex less invasive and less costly treatments so our goal is to enable all children with C pory to be regularly monitored for hip migration such that doctors can identify a migrating hip early and can provide timely
Treatment the National Institute for health and caul actually recommends that those children have regular hip excise image being taken and that they’re monitored forp migration and the framework to achieve this is a national surveillance program called the C poliy integrated pathway or cips so that means that all the images taken as
Part of seips they need to be analyzed and the bones need to be measured because it’s those measurements that help the doctors to identify how much change there has been compared to the last image now in areas where CIP is working well this is really successful and it leads to improved patient outcome
By identifying those migrating hips unfortunately though in a lot of areas there’s a lack of capacity and resource to do this analysis because it’s time consuming and the quality of it is affected by who is doing the analysis so in some areas there might be months between the image being taken and the
Image being analyzed and because of that there’s a range of variation in the terms quality of care that should and re py get with regards to HP migration so we’ve developed an artificial intelligence BAS setem to automatically analyze images so our system will outlines the bone of
Interest in the hip xray image and it will then calculate What’s called the rhus migration percentage which is the key clinical measurement to assess hip migration we’ve validated our system on a large clinical data set of 1650 hip X-rays of children with Civ pory and we do the analysis by comparing
Our automatic measurements to that of five clinical expert who have all manually measured the very same images now in terms of results we found that the variation amongst those five clinical experts was the same as the variation between our automatic system and the manual measurements and also
That when we look at the average difference between the automatic measurements and the manual measurements that that was within the range of what would be expected if we were to ask a clinic expert to measure the same image twice so overall our results suggest that our system performs equally well to clinical
Experts our ultimate goal is to integrate our system into cpip such that every image being taken is automatically measured that data would then fit into the electronic health record for that patient to inform clinical care but it will also feed into what might become the world’s largest database of hip
Measurements of children with sorty informing important future research so in conclusion our system will have help spot doctors to identify which certain re policy might have hip problems it will enable less complex less invasive less costly treatment it will save doctors about 10 minutes analysis time per image it will reduce measurement
Inconsistency by automating the process and overall our system will enable that every child with civil posy will receive the same high quality of care thank you if we’re not counted we do not count a statement used by an LGBT Foundation report published in 2021 he’s one that really resonates with my
Research the idea that we have groups in societ we know little about and whose data we don’t have and feel they don’t count this is particularly the case in the design and delivery of healthc care services we have very little data on sexual orientation and so we have groups that
Don’t um that we have groups that don’t think that they don’t count and so my mix methods PhD research um has looked at aiming to understand uh access to Mental Health Services for sexual minority groups um I initially my systematic mapping review looked at existing evidence of inequalities in Access and
Found that there was very little very little evidence on access to um Mental Health Services for um sexual minority groups I wanted to rectify this by working with the local NHS trust um to look at their data to see whether there was any sex orientation data that we
Could look at um to understand access to sexual minority groups I gained access to four and a half years worth of data um from two adult Mental Health Services improving access to psychological therapist also known as IOP services and Community Mental Health teams um also known as
Cmhts to understand whether we could see whether that we can understand access to Mental Health Services for sexual minority groups so I looked at representation of sexual minority groups across the data sets and compared that with census 21 um data finding that there are um such mons are um accessing
Mental Health Services um with more almost double the representation um compared with the data we’ve got for census population however we can’t make any um assumptions about the size of mental health need for these groups I then looked at the extent of missing data um sex for sexual orientation
Across the um data sets um to assess um whether we can look up predict of missingness are certain things associated with missing sexual orientation data um and we found that missing sexual dator is quite prevalent across the um data sets but also that it’s associated with various different service and service user characteristics
Such as age ethnicity deprivation and referral source indicating that there is something we could do about addressing those um data recording practices patterns of access over time differ by sexual orientation um and also by services so particularly during the covid-19 pandemic um referrals to IOP Services um increased for sexual minority groups
Whereas they decreased for cmht services during the pandemic when compared with heterosexual groups and contacts um for non-attendance for contacts with mental health services um doesn’t appear to differ by sexual orientation and I’m just looking at um doing some regression analysis at the moment to see whether there are
Other predictors of non-attendance in the data sets so I draw your attention to this so to highlight that we can use this data to look at groups and consider um addressing the missingness um and considering the Serv service user themselves so I’m currently conducting sem structured interviews for
L with ltq plus people about their experiences of mental health services and there just just a few quotes there um around um their perspectives of mental health services and their use um and their experiences of disclosing their sexual orientation I’ve also embedded um patient and public involvement and um
Stakeholder engagement in my PhD um from the outset to embed their perspectives into the design of the research but also interpret the findings My Hope Is that this research can illustrate that you can that we should collect more data for these groups so that sexual minority groups feel that they are counted and
That they no longer feel like they don’t count thank you next we have Matthew so hello I’m Matthew Watson from der University and I’m going to talk a bit about uh chemotherapy risk stratification model and how we’ve taken it from Theory to being used in practice
So for those of you who aren’t aware chemotherapy is given in Cycles this cycle consists of a blood test which amongst other things looks at your kidney and liver function to make sure you’re well enough to continue with treatment if you are well enough you’re given your next dose of chemotherapy and
Then you have a couple of weeks to rest before the cycle starts again now these blood tests are a real block for patients that mean you have to go into a clinic to have the blood test taken they put a lot of strain on photomy services to get the results in
And quite often the blood tests are delayed which results in your treatment being delayed which is linked to worse outcomes and we know that actually only around 10% of chemotherapy patients have any dysfunction that’s picked up by these blood tests so maybe we can use some sort of machine learning uh
Instead so that’s what we did and on a small dayis from a single hospital we trained and validated a machine learning model that took in your demographics and some blood test results from your previous two cycles of treatment to stratify you into a high risk or lowrisk group the idea being
That if you predicted high risk you have the blood test as normal if you predicted low risk it’s not quite as important to get that blood test result in before you continue with treatment we knew this worked on data from one hospital so we went to some clinicians and said
What do you need from us to be able to use it in practice we’ve got two main themes one needs to be very well validated and two it needs to be easy to use so we extended our validation to three different hospitals in the UK each with very different patient populations
Um and this is the important bit because if your model isn’t trained on distinct very distinct patient populations when it’s deployed in practice it might not not work as well as you expect and so this is a bit of a complex slide but just very quickly higher is
Better the blue bar is a model trained on data from two hospitals whereas the orange bar was only trained on data from hospital one and the green bar was only trained on data from hospital 3 and in general what you see is that the blue bar which was trained on data from
Multiple hospitals performs well on data from multiple hospitals whereas models trained on data from only one Hospital tends to only work on data from that hospital so now we have a well validated model what about the ease of use of it well originally on the left we had a
Model which takes in a lot of complex data and it was difficult for cl conditions to input it into the model so we simplified this made it easier to use and that’s the results we have on the right and as you can see performance is pretty consistent um and in some cases
Even better than the more complex model at the same time we looked at the bias of the model so we evaluated uh performance across different patient subgroups so does it perform better on males versus females for example or older people versus young people and in general you see that it’s fairly well
Balanced and not massively biased towards a significant population so now we have clinician buying what about patients are they happy with it being used well actually in general when we talk to chemotherapy patients yes they are as long as you your clinicians are happy what we got
From patients is that there’s actually a massive information overload for them and they given a lot of information that maybe isn’t always relevant to them so we actually chain some of our project to develop a patient facing app which can uh provide patient specific cancer specific information for them and I’ll
Leave you with a couple of conclusions and that’s my time so Georgina hi my name is Georgina Ryland um and I’m going to talk you through our project where we’ve linked Family Court data on public uh law proceedings to Health Data in England so what are public law Family Court
Proceedings these are cases where children’s Social Services uh believe the child’s at risk of Mal treatment or serious harm and in these cases they may go to court or to intervene and this can end up in the children being removed from the family home so we’ve taken this
Data on the mothers specifically and linked them to hospital data and the admissions associated with the delivery of children and this has enabled us to create a cohort of one uh 3.1 million mothers in England who have delivered their first child from 2007 onwards and we were able to follow them up over
Their parenting and childbearing years and see how many um end up in court proceedings within 10 years so we’ve ident the F this slide is um looking at the profile of mothers um who have been involved in care proceedings within 10 years and what they look like at the delivery of their
Child and we see that they’re more deprived and they’re younger at first delivery than women who were not involved in court proceedings so 79% of the women who were involved in court proceedings were aged under 25 at first delivery compared to 30% of women who
Were not and 72% of women were living in the two most deprived quintiles um in comparison to 30 uh 50% of other women not only this but we can look at their health profiles before delivery and we see that the women who are involved in care proceedings um have much higher
Prevalence of of chronic health conditions intellectual or multiple disability mental health and adversity related injuries and these are using their Hospital admissions in the three years prior to their first child um and adversity related injury admissions are associated with drug and alcohol use self harm and violence so using this cohort we’ve then
Looked at them for the first 10 years and estimated cumulative incidence of care proceedings with Kaplan Meer analysis um and we and the chart is showing the cumulative prevalence by um 10 years for each of the social and demographic uh indicators and overall we see that 1.3% of women were involved in
Care proceedings within 10 years of their first child but unsurprisingly from the previous slide we’re seeing much higher prevalence um for teenage mothers and women who are um living in the most appro areas of England and not only that when we look at their health burden at Umi at first
Child we see this Stark Difference by um Health profiles so almost 23% of women with an intellectual multiple disability will be involved in care proceedings within 10 years that’s 5.7% if they had an admission associated with mental health um diagnosis and 133% if they were had an admission associated with Verity related
Injury now we think the take-home messages from this is that by profiling women and using this information at first birth we can Target services at the most vulnerable women at the start of their parenting journey and this doesn’t have to just in uh involve their health it could also be looking at
Parenting classes help getting back into Education and Training and those wider social um and social uh factors um we also this is the first time we’ve linked this data in England um and so this is the start of our analysis journey and we are hoping that
Um by digging more into the profiles and the birth trajectories and the mortality different and we’re going to be looking at mortality in these women we’re hoping that over the coming years we’ll be able to better Target services and help these families involved uh in care proceedings and this will improve the health
Outcomes not only of the mother but also of their children uh thank you and hi I’m arum I’m data engineer at University of dundy and today I’ll be talking about uh mapping the Scottish population Health scale data to omop to support Fair data principles so we are aware that data
Saves lives but the biggest obstacle in achieving this is uh the inability of the researchers to uh easily find and access uh the data that fits their research needs um as we know that the volume of the data is increasing we need to be mindful that we create this data
With longevity in mind and uh the researchers are able to reuse this data this way we would be able to use the data to its true potential and also we’ll be able to generate uh quick feasibility studies um uh And Timely research outputs without duplicating the effort of recollecting the data or
Wasting time in finding where it exists Fair data principles explain how the data and the outputs could be represented and organized such that the data is findable accessible understandable and um exchangeable and reusable so we know that uh the data could be present in uh variable formats
In silos um and in different structures and there could be various data governance issues which could make it difficult for the researcher to find and access the data um so we are trying to transform this data into um uniform standard uh common data model which is called omop
Which you’ve heard before so it’s called Uh observational model outcomes partnership which was developed by Odyssey observational Health uh data science and informatics the main concept of om is to transform all the data sets to a standard schema terminologies units and measurements uh this has enabled us
To um so in public health Scotland um is also actively participating in transforming their data to omop and this has uh allowed Public Health Scotland to collaborate with HDR and their covid-19 project called cocon where they uh published their metadata on uh the Health Data uh research UK Innovation
Gateway and their um om of transform data sets are available on the cohort Discovery tool this tool allows the researchers to um query their research questions in a Federated Fashion on uh variable data sets because they’re standardized also Odyssey um and Eden collaborated to create a u multinational study uh which involved 26
Databases across 11 country’s Health inform informatic Center was also part of it including data from NHS um tayside and5 and um the database included uh data of 24 patients the aim of the study was to find the incident rates um involving um Co vaccinations um uh around the aesis and compared them with
The uh with the patients um population of preco uh levels uh the only way this uh huge large scale study was uh possible was because um the analytical code was shared across and the data was transformed to omop so it was fairly easy to do this study in a
Shorter time across uh 11 countries and 26 databases uh so we uh during the co cocon project we uh developed tools which are called carrot tools uh which will uh enable the OM of transformation the main idea behind this was to standardize the uh process of transformation um and to ensure that um
The organizations can transform the data with minimal omop um knowledge so we are trying to make sure that the data mapping uh exists outside the data transformation so the organizations can use external help uh who can map the data and since the mapping process does not include any real data and it’s based
On the metadata of the data um uh they can easily reuse the mapping rules and they can exchange the mapping rules across different organizations so as discussed earlier omop could be labor intensive but using our tools um it has a very low learning curve and even if
You’re not familiar with omop you can still transform by using um external expertise who can map your data and you can transform in your secure network um without giving your data away and these are some of the results of the co data that we have transformed in the public health Scotland I’m over
Time so thank you