Approved Projects 

Chief Investigator: Nicholas Hamilton

Aim:

The aim of this study is to build a computer vision model that can identify when a complete image of the larynx has been captured in real-time during throat endoscopy. The model will be trained and tested on endoscopy images of the larynx held in an archive at the Royal National ENT Hospital (UCLH).

Lay Summary:

This project is about building a smart computer tool that can help doctors and nurses check the throat, especially the voice box (called the larynx), to see if it looks healthy.

Right now, only special doctors called ENT (ear, nose and throat) specialists usually do this test using a tiny camera. That’s because it takes a lot of training to know what to look for. But with this new tool, we want to make the test easier, so more healthcare workers can do it in local clinics, without needing to send people to hospital.

To build the tool, we’ll use a large collection of videos that show what a healthy throat looks like.

The computer will learn from these videos how to spot a normal, healthy voice box. Later on, we’ll test the tool in real clinics to see how well it works.

In the future, this could help people get checked faster and closer to home. It will also help doctors and nurses who aren’t ENT specialists feel confident when doing the test, with the computer helping to confirm that everything looks normal.

Chief Investigator: Mr Raffaele Napolitano, PhD 

Aim:

To develop and validate a machine learning-based risk prediction model that leverages detailed ultrasound measurements alongside other routine clinical data to enable more personalised prediction of adverse pregnancy outcomes. 

Lay Summary:

What is this research about? 

During pregnancy, healthcare teams regularly check for possible complications that might affect the mother or baby. These include high blood pressure, poor growth of the baby, early (preterm) labour, and stillbirth. A key part of these checks are the detailed measurements taken during routine ultrasound scans. 

Why is this research needed? 

Currently, midwives and doctors collect lots of information during pregnancy check-ups. While blood tests and blood pressure readings are important, ultrasound scans provide particularly valuable measurements of the baby, blood flow, and placenta. We want to make better use of all this information, especially the detailed scan measurements, to more accurately identify who might develop complications, better support those at higher risk, reassure those at lower risk, and avoid unnecessary tests or treatments. 

What will the research do? 

We will use computer technology to analyse information collected during routine pregnancy care, focusing particularly on the measurements taken during standard ultrasound scans. By combining these scan measurements with other routine pregnancy data, we will create a new tool that can give each person a more personalised assessment of their risks, help healthcare teams plan the right level of care, make better use of NHS resources, and support discussions between healthcare staff and pregnant people about their care. 

How might this help? 

This research could help pregnant people and their families better understand their individual risks by making the most of measurements already taken during routine ultrasound scans. It will enable healthcare teams to make more informed decisions about care based on detailed scan information combined with other pregnancy checks. This could reduce unnecessary worry and medical procedures for those at low risk, whilst ensuring those at higher risk get extra support when needed. Importantly, since the tool will use measurements from standard pregnancy scans and check-ups, no extra tests will be needed. 

Chief Investigator: Dr Anoop Shah

Aim:

The aim of this research is to evaluate the current diagnostic approach to secondary hypertension in young adults and develop an evidence-based screening strategy.

A secondary aim is to evaluate the performance of methods for extracting detailed clinical information from free text electronic health records. to improve diagnostic efficiency and alignment with best practice.

Lay Summary:

High blood pressure, known medically as hypertension, isn’t very common in younger adults.

When it does occur, it can sometimes be caused by undiagnosed health problems rather than just lifestyle issues such as diet or exercise. This type of hypertension, called “secondary hypertension,” can be linked to hormonal problems, kidney conditions, issues with glands called the adrenal glands, narrowing of the kidney arteries, or even sleep disorders like sleep apnoea.

Finding out the exact cause of high blood pressure is important, because secondary hypertension often needs treatments that are different from those used to treat regular high blood pressure.

At the moment, doctors aren’t always sure exactly when and how to test younger adults for secondary hypertension. In this study, researchers will look back at medical records from young adults who have visited hospital clinics with high blood pressure. They will assess how effective different medical tests are in detecting secondary hypertension and will try to find the best approach to testing for these conditions. They will also look at tests used to check for any harm caused by high blood pressure to important organs, such as the heart, kidneys, and brain.

The goal of this research is to create clearer, evidence-based guidelines for identifying secondary hypertension in younger people. This could help doctors diagnose these conditions faster and more accurately, leading to quicker treatment, fewer unnecessary tests, and better use of healthcare resources.

Chief Investigator: Dr Steve Harris

Aim:

We aim to develop a predictive model that estimates an individual patient’s risk of antimicrobial resistance. This model will support more accurate and personalised antibiotic prescribing, improving patient outcomes and antimicrobial stewardship.

Lay Summary:

Imagine you are sitting in hospital – again. This is your sixth visit over the last few months with a fever. You have recently had chemotherapy, and you were warned your immune system may be weaker, making you more vulnerable to infection. You feel relieved you are being admitted to hospital for antibiotics.

It is your fifth or sixth course of antibiotics, you have lost count. You hope that, like before, you will get better. However, you have heard about antibacterial resistance, and this worries you as you have had lots of courses of antibiotics. You have had a lots of blood tests, urine samples and swabs taken but do not know the results in detail or what they mean. You do know that sometimes it has taken a few changes of antibiotics before the right one was found- the one that worked to get you better.

It is difficult for the healthcare professional looking after you to access all your previous antibiotics and test results. Time is limited, and they do not always have time to look through everything, especially all the antibiotics you have previously had. So, they follow hospital guidelines and prescribe the antibiotics that work for most people, while they wait for the results from the latest tests – the swabs, blood cultures, and urine samples taken today. These results can take at least 48 hours to come back.

Ideally, antibiotics should target infections as precisely as possible- and fortunately for most people they do. When this happens, treatment is more effective, there are fewer unnecessary side effects, and there is a lower chance of developing antibacterial resistance.

But when taking antibiotics repeatedly, bacteria can become resistant. This means that for some people standard treatment might not work, and delays in finding the right antibiotics can mean people are sicker and have poorer outcomes.

We want to simplify and personalise this process. Our research aims to bring together key information including your past infections and the antibiotics you received to support clinicians in choosing the best antibiotic for you right from the start.

To achieve this, we will use data already available at University College London Hospitals (UCLH), including prescription histories, and laboratory results. By analysing this information, we aim to identify patterns that could help predict which patients are likely to benefit from different, more personalised antibiotics treatment right from the start. Our goal is to help healthcare professionals make faster, more precise decisions to improve care for patients like you.

Chief Investigator: Dr Anand Pandit

Aim:

The goal of this research is to create highly accurate models to improve neurosurgery outcomes by carefully analysing existing patient data, including medical images and clinical information. Using the latest artificial intelligence (AI) methods, including sophisticated computer models trained on large datasets and specialised statistical approaches for more specific questions, we aim to tailor treatments more closely to each patient's individual needs.

We will particularly focus on conditions for which we have extensive data at UCLH, such as brain tumours, complex spine problems, and adult hydrocephalus (excess fluid buildup in the brain). Ultimately, this research aims to provide better, more personalised care and improve results for patients.

Lay Summary:

This research aims to make diagnosing and treating neurosurgical conditions—such as brain tumours, spinal conditions, and hydrocephalus (excess fluid in the brain)—better and safer. We will do this by using artificial intelligence (AI) and advanced data analysis. By studying large amounts of anonymised patient records and medical images, we hope to create AI tools that can help doctors predict surgery results, tailor treatments to individual patients, and quickly spot patients who might face higher risks.

For patients, this means clearer diagnoses, fewer unnecessary hospital visits, shorter stays in hospital, and fewer complications after surgery. For doctors and hospitals, this research will help them manage resources better, making sure each patient gets the best care exactly when they need it.

Chief Investigator: Dr Vasilis Stavrinides

Aim:

The overarching aim of this project will be to systematically describe the histological transitions occurring in men who undergo serial MRI-targeted biopsies for either suspected or actively surveyed early-stage prostate cancer.

Specifically, the primary focus will be to estimate the rates at which either cancer-negative or early cancer MRI lesions progress to clinically significant, high-grade prostate tumours.

This will be achieved by acquiring serial prostate-specific antigen (known as PSA), MRI and targeted biopsy data, which will then be modelled across time using modern statistical techniques.

The potential clinical benefit will be the early identification of baseline or longitudinal features that stratify patients according to their disease progression risk, the timely treatment of those patients, and the optimisation of modern imaging-based protocols for prostate cancer surveillance.

Lay Summary:

The use of MRI to guide prostate biopsies has changed the way doctors find and manage prostate cancer. It helps them spot suspicious areas more accurately than older methods, which were inaccurate and more painful. Using MRI is especially important for finding and treating aggressive cancers early, while also avoiding unnecessary treatment for slow-growing tumours that do not pose an immediate threat.

However, we still don’t fully understand how suspicious MRI areas in the prostate change over time. This knowledge is important, as doctors would like to track MRI changes better and make more informed decisions about treatment.

Our study aims to do just that - by using modern statistical tools to look at repeated prostatespecific antigen tests (known as PSA), MRI scans, and prostate biopsy results over time in UCLH patients. The goal is to better understand how prostate cancer behaves across time and, as a result, help doctors make better decisions and improve care for patients.

Chief Investigator: Dr Peter Odor

Aim:

To evaluate whether elective major surgical patients cared for with a perioperative critical care model have lower morbidity, mortality and health care utilisation compared with control patients who received post-operative care in combined generic critical care.

Lay Summary:

Recovering well after a major operation is very important for a patient’s health. At University College London Hospitals (UCLH), a new 10-bed specialist unit has been set up to care for patients who are having planned (elective) major surgery. This unit is called a perioperative critical care unit, which means it supports patients around the time of surgery—especially during the recovery period after the operation. The unit is staffed by a team with special training in looking after surgical patients.

This study will look at whether patients cared for in this specialist unit have better recovery than patients who were looked after in a general intensive care unit (ICU), where both medical (non-surgical) and surgical patients are treated together. We will compare results such as how long patients stay in hospital, how many complications they have, and how well they recover overall.

We are not collecting any new information from patients. Instead, we will use existing information that is already routinely collected in patient records. This data will be fully anonymised, which means that no names or personal details will be used, and individual patients cannot be identified in any way.

Our goal is to find out whether this kind of specialist care helps people recover better and uses hospital resources more effectively. If we find benefits, the results could help other hospitals consider introducing similar care for patients after major surgery.

Chief Investigator: Professor Sonya Crowe

Aim:

To develop a method for predicting demand for inpatient beds, that incorporates all sources  of patient flow - including emergency and elective admissions into each specialty, and transfers  and discharges out of it - for use by hospital managers. 

Lay Summary:

Picture yourself being wheeled into A&E. Your heart sinks when someone says you need to stay in hospital. It sinks deeper when you discern a queue of people waiting for a bed on a ward. You wait, shifting uncomfortably as time passes. Meanwhile, a bed manager scours their list of free beds, trying to squeeze you in. The clinical specialty you need is pressured, and the only beds due for release today are for the opposite sex. The few staff members with the clout to make more beds free have other pressing concerns. You wait. And wait. 

We aim to give hospital managers an early warning about delays like this. We will use the hospital’s patient record system to identify how many patients will arrive into each clinical specialty within the next 8 or 12 hours, and how many will leave. We will combine these inflows and outflows to predict an overall rise or fall in the numbers needing care under each specialty. We will check whether the predicted numbers match what actually happened. 

If our method works less well for patients of a particular race, sex or age group, or some other protected characteristic, we will investigate whether our results reflect a bias in the care given to patients with those characteristics and whether our method risks reinforcing that bias. We will alert hospital managers to any differences due to bias, to help them address any inequalities. As this is a retrospective study using past data, there would be no immediate changes to address inequalities, but the findings would hopefully influence policy and practice down the line.

Hospital managers with access to this early warning would know which clinical areas are likely to face pressure for beds and could act sooner to make space for the incoming profiles of patients.

We will develop, test and demonstrate our method using UCLH data, and will share the method and the results in a way that makes it usable by other hospitals. By doing so, we intend that there will be a wider benefit for patients and the hospitals that care for them.

Chief Investigator: Dr Paul Expert

Aim:

To describe the interaction between infection control measures and patient flow, and the impact on patient level (e.g. new infections, length of stay) and hospital level (e.g. outbreaks, cancellations) outcomes. 

Lay Summary:

Healthcare professionals must regularly balance competing demands for hospital beds while maintaining infection prevention and control (IPC) standards. For example, a patient ready to leave the ICU had their transfer delayed because the only available bed was in a room previously occupied by patients exposed to COVID-19. Despite the low likelihood of infection, protocols required the patient to remain in the ICU, blocking access for critically ill patients and causing a cascade of delays in bed availability across other wards. This research examines how the whole hospital works together, using real-world data to improve IPC decisions. It focuses on two complementary streams:

i) validating a risk assessment model for bay closure decisions to balance bed utilisation and patient safety;

ii) characterising how infected patient movement patterns affect general patient outcomes to inform hospital-wide IPC strategy.

Chief Investigator: Dr Jonathan Gillham

Aim:

To develop models, using multimodal data (e.g., clinical notes, laboratory results, patient demographics), to assist clinicians in making informed decisions about which routine blood tests to order for inpatients at UCLH.

Lay Summary:

Routine blood tests are an essential part of patient care, but the decision of which tests to order is nuanced.

Over-testing is common, which can lead to unnecessary discomfort for patients, poorer outcomes, and additional strain on healthcare resources. On the other hand, missing an important test can delay diagnosis or management, impacting patient care. This study will explore routine blood testing practices for inpatients at UCLH using retrospective data, uncovering trends and patterns in current practice. The insights will be used to develop a decision-support tool that employs modern machine learning and pattern recognition techniques. By incorporating data from multiple different types of sources, the tool aims to assist clinicians in making the best decisions about which routine blood tests are truly necessary, ultimately improving patient outcomes and optimising resource use.

Chief Investigator: Peter Odor

Aim:

To quantify and explain variation in surgical procedure length.

Lay Summary:

A retrospective analysis of existing dataset to determine variation in surgical procedure duration, and factors contributing to variation​.

The current model for theatre planning using moving average summary duration​ is flawed due to lack of accounting for variation. An alternative approach = “probabilisticscheduling” – uses mean and SD​. It determine factors contributing to variation​.

The primary objective is to improve theatre scheduling efficiency, productivity, reduce day of surgery cancellation​, to quantify and explain variation in surgical procedure length. ​

The secondary objectives are:

• Quantify contribution of potentially explanatory variables to procedure touch time variation​

• Generate algorithmic/probabilistic list model to schedule operating lists based upon expected duration time and variation in procedure time​

• Real world validation of an updated algorithmic/probabilistic list scheduling system by comparing lists designed with the traditional system vs. updated model​

Chief Investigator: Veeru Kasi, Shonit Punwani

Aim:

To build an AI model that can detect prostate cancer on prostate MRI scans. This is a feasibility question, and the evaluation of a deployed model is out of scope.

Lay Summary:

Prostate cancer is a very common disease and the 5th largest cause of death of men globally. Diagnosing prostate cancer has recently changed to a pathway that requires a MRI scan to help find suspicious lesions within the prostate that can then be sampled using a needle. This has much improved the diagnosis in the UK and internationally. About 200,000 men in the UK will currently have an MRI scan of their prostate every year. However, there is a real shortage of trained experts to provide reports for these studies. This results in delays to diagnosis across the country and also sometimes incorrect reporting which can miss important cancers or cause patients to have unnecessary biopsies (with potential for infection and bleeding as two main risks of biopsy). Given the large numbers of men being investigated there is also a large healthcare system cost to diagnosing prostate cancer and avoiding unnecessary procedures would significantly reduce costs.

Based on the wealth of prostate MRI expertise at UCL and UCLH and the number of studies that we have done in men since 2004, we envision sharing prostate MRI images with Microsoft so that we can co-develop an AI based automated reporting system that can address the workforce shortage in the UK and world wide. We also plan to develop along the way AI tools that radiologists can use to improve workflow and avoid incorrect interpretation of MRI studies.

Initially we aim to share images that we hold on EPIC (approx. 5000 prostate MRI studies). This will be used to explore and develop initial automated reporting and tools. These will then need to be further refined and tested before any of them can be used at UCLH and at other sites. This initial request for data sharing will seed the developmental process.

Chief Investigator: Nishchay Mehta

Aim:

To identify hearing loss phenotypes, their progression and associations with otological and non-otological disease using hearing health big data.

Lay Summary:

Hearing loss is the most common sensory disorder in humans with 1.5 billion people affected by this during their lifetime. Some types of hearing loss are more common as we get older and therefore the burden of hearing loss is predicted to rise further as our aging population increases. Hearing loss has a significant impact on quality of life, patient health and safety as well as placing a huge demand on increasingly stretched public health services.

In order to prepare and respond to this rapidly accelerating public health crisis we need to better understand the different types of hearing loss to identify who is most at risk of both worsening hearing loss but also other associated medical conditions and also which patients are most likely to benefit from new treatments. The digitialisation of patient health records offers an exciting opportunity to use the latest advances in computer science methods to look at large amounts of hearing health patient data and answer these questions.

In this project, we will create a store of patient data that has been collected routinely as part of standard NHS clinical care. This will include demographic details, test results, measurements and details of medical conditions. A powerful computer programme will be used to analyse this data and describe different types of hearing loss as well as how these hearing loss types change over time.

We will perform further analysis to identify links between these hearing loss subtypes and other medical conditions including dementia, diabetes, stroke and high blood pressure.

Chief Investigator: Mohammad Eddama, Yassar Qureshi

Aim:

To identify optimal perioperative management pathway for patients undergoing surgical resection for oesophago-gastric cancer.

Lay Summary:

Surgery for oseophago-gastric cancer is associated with high risk of morbidity and mortality. Actions, such as enhanced peri-operative care, centralisation of specialist services and better patient selection, have significantly improved patient outcome in western countries.

The aim of this project is to identify peri-operative factors that can be manipulated to improve the long-term outcome of patients undergoing treatment for oesophago-gastric cancer.

The objectives of this research project are: 1) to describe the relationship between peri-operative patient's factors, such as age, anaemia post-operative outcome; 2) assess the effectiveness of cardiopulmonary exercise (CPEX) testing as a predictive tool for the development of complications following surgery; 3) measuring the impact of complications on both disease-free survival (DFS) and overall survival (OS) after surgery; 4) developing an enhanced recovery (ER) pathway to advise clinical practice.

We will collect anonymised data from patients undergoing oesophago-gastric surgery using our patient data system (EPIC), analyse and present the data using descriptive and inference statistics. We will use statistical packages including SPSS and GraphPad Prism to analyse the data.

Chief Investigator: Matt Wilson

Aim:

To describe current patterns of opiate use for the perioperative population stratified by perioperative stage and surgical specialty.

Lay Summary:

Every year in the UK thousands of people undergo surgery for emergency and chronic conditions with many more people remaining on waiting lists for operations. Providing pain relief during and after surgery is a key component of a patient's journey through this process. Delivering good quality pain relief (or analgesia) is associated with improved recovery, less complications after surgery and shorter stays in hospital.

Pain management during and after surgery is normally achieved using a combination of medications from simple paracetamol and ibuprofen, through to more potent medications like morphine. As part of the anaesthetic for surgery, patients may also be offered specific procedures to numb areas of the body being operated on (known as regional anaesthesia) using local anaesthetic. In some cases a tube or catheter may be left close to the site of surgery or in the back to continue delivering local anaesthetic for several days after surgery when pain is likely to be worst.

As with many treatments, what pain medications are used for surgery and how they are delivered is not a standardised practice - patients may receive different treatments depending on who looks after them. At present we don't know how much variation exists between individual patients and clinicians and whether it makes a difference to how you recover or not.

This research study uses the electronic data that is collected as part of patients normal care in hospital to search for patterns and differences in medication prescribing and whether this variation effects outcomes from surgery like postoperative nausea and vomiting, how patients experience pain following surgery and the time taken to recover key functions like eating and drinking and mobilising postoperatively.

In addition, a small proportion of patients go on to experience severe pain after surgery which can develop into a condition known as Chronic Post Surgical Pain. This study will examine how differences in pain management during and after surgery might contribute to the development of this condition

Chief Investigator: Mathew Grech-Sollars

Aim:

To answer the question "Can quantitative reports on MRI biomarkers assist neuroradiologists in reporting on neurological conditions and help patients by speeding up the time taken to report whilst improving accuracy and lowering subjectivity?”

Lay Summary:

The quantitative neuroradiology initiative aims to deliver a means by which radiologists reporting on MR images have access to quantitative, non-subjective data that reflects the presence or progress of a neurological condition or disease.

To this end, we are applying for DAP-R approval to access both imaging and non-imaging patient data to cover work in the research, validation, and clinical evaluation of the QNI framework.

Chief Investigator: Magdalena Sokolska

Aim:

To assess methods for improvement in image quality, resolution and robustness of clinical fetal brain MRI.

Lay Summary:

Magnetic resonance imaging (MRI) is an important tool for confirming and detailing abnormalities in the developing fetus. Fetal MR imaging can be difficult due to the movement of the fetus especially very early in the pregnancy when MRI for spina bifida surgery indications is needed. Conventionally, successive stacks of images are acquired very fast to try ‘freeze’ the motion. However, this does not prevent problems with the movement of the fetus, which causes anatomy to either be missed or not aligned properly in consecutive images in the stack. Typically, this leeds to the necessity to acquire several stacks to ensure that the whole brain can be evaluated and clinicians need to review all scans to assess the pathology.

An alternative approach is to take all acquired stacks and work out how these images would fit together in 3 dimensions. This enables the generation of higher- imaging of the brain that can be viewed in 3 dimensions. This is called super-resolution slice-to-volume registration (SVR).

This study will assess 3D SVR using standard clinical data in comparison with 2D stacks and evaluate the potential of introducing it to a clinical service. This will be done by measurement of brain structures as well as visual assessment and comparison of clinical findings between 3D SVR and standard of care 2D stacks.

Chief Investigator: Steve Harris

Aim:

To develop ML models for detecting and qualifying nasogastric tube (NGT) placement on chest x-radiographs (CXR) with the prospective aim of facilitating the timely correction and prevention of further complications associated with incorrectly positioned feeding tubes within clinical practice; and thereby increase patient safety.

Lay Summary:

A Nasogastric tube (NGT) is a thin tube that is passed into the stomach via the nose for short- to medium-term nutritional support, medication administration or aspiration of stomach contents. NGTs are amongst the most commonly used catheters in critically ill patients in intensive care units (ICU) and high-dependency units and departments where patients require nutritional-support (i.e., Stroke units). Due to increases in the number of hospitalized patients, it is estimated that approximately 10 million NGTs are used annually in Europe, 1 million of which in the UK (~1.2 million in the US).

Previous research highlights a variety of complications associated with NGT placement, which can range from minor cases of nose bleeds to inhalation of stomach contents into the lung and even death. Instances of unknowingly misplaced NGTs being used for feeding, with the feed entering the patients lungs are classified by the NHS as Never Events: “serious incidents that are entirely preventable because guidance or safety recommendations providing strong systemic protective barriers are available at a national level, and should have been implemented by all healthcare providers” . While all this highlights the importance for feeding tubes in particular to be placed properly and used safely, clinical studies demonstrate that up to 3% of NGTs are reported as misplaced into the airways, causing complications in up to 40% of these cases.

Given the serious complications that can occur from NGT misplacement, UCLH has a detailed policy describing the indications and technique of NGT insertion alongside nationally agreed standards for positioning verification. This includes training and guidelines for doctors or reporting radiographers when checking NGT position radiographically. In this policy, the first line of test in confirming the correct positioning of a feeding tube is by obtaining a sample of fluid from the stomach that shows a level of acidity indicative of the stomach. However, since this cannot be achieved successfully for some patients, and with a large proportion of ICU patients receiving anti-acid medication, the use of CXRs remains the most definitive test for checking NGT placement.

Due to the large number of CXRs obtained each day, especially in intensive and emergency care, and with a limited number of radiologists available, image interpretation can be substantially delayed. Thus, current practices indicate that it is often emergency and ICU doctors who check the CXR to verify the NGT’s correct positioning and suitability for use prior to the radiology report being issued.

Yet, such assessments by non-radiologists working in stressful situations when hospitals are capacity, are prone to both human error and some delays in assessment. This means that sub-optimally positioned NGTs can be missed initially, but are often picked up by the radiologists later. This emphasizes the importance of early detection of misplaced NGTs to allow for more timely correction and prevent any additional complications.

We envision two main use scenarios in which an accurate, instant detection and notification of NGT misplacements from CXRs could benefit clinical practice: (1) As an early alert to ICU doctors or nurses, it will enable prompt, data driven decision-making and NGT adjustment for more effective and safe use. (2) As an early alert to help prioritize the review of most urgent CXRs by local (UCLH) radiologists to reduce delays in notifying ICU doctors of potentially unrecognized NGT misplacement.

Initially, this work focuses on developing a machine learning model to identify misplaced NG tubes on CXR. We will also study ML integration within the ICU at UCLH due to its already established all-digital end-to-end radiology workflow, and to ensure that the sickest, most dependent patients in the hospital will get treatment faster and more safely. In parallel, we will study requirements for future ML system roll outs to any other inpatient area that frequently places NGTs. In a first instance, this will include Stroke Departments within UCLH.

The work can also generate a training opportunity leveraging known cases of misplaced NGTs or cases that were hard to interpret on CXRs. The training datasets can upskill ICU and Stroke ward doctors who often have little experience of assessing such CXRs in routine practice.

Chief Investigator: Dr Jamie McClelland

Aim:

Developing and evaluating models for anatomical changes in H&N cancer patients treated with RT Creating Radiotherapy treatment plans which are robust against anatomical changes.

Lay Summary:

Anatomical uncertainties during proton radiotherapy can lead to a difference in the delivered treatment from the planned treatment. This difference has the potential to reduce the efficacy of treatment; increasing toxicity to organs at risk and healthy tissues along with a reduced dose to the tumour and leads to replanning of the treatment. The proposed project aims to model the anatomical changes which head and neck cancer patients undergo during treatment. The results of such models could be used to predict the anatomical changes of new patients. By anticipating anatomical changes in advance, they can be incorporated into the treatment planning stage of clinical practice.

Chief Investigator:

Tom Lumbers

Aim:

To evaluate the patient characteristics and treatment patterns who received an intensification of diuretic therapy at any time during their admission.

Lay Summary:

Fluid overload is a health issue that can cause unexpected hospital stays or make other serious health problems worse. It can happen due to conditions such as heart failure, kidney and liver disease. Heart failure cases alone cause people to stay in the hospital for a million days each year in the UK. Even though fluid overload affects a lot of patients and puts pressure on the healthcare system, we have a lack of knowledge about the precise medical conditions that cause this condition and the differences in how fluid overload is treated for these patients.

One particular aspect of treatment is fluid restriction. Clinicians will have different preferences to either recommend this or not for patients with fluid overload. At UCLH the THIRST Alert trial is ongoing to determine whether in this context of variation in clinical practice, it is feasible to recruit patients into a research study where fluid restriction is given as part of a randomised controlled trial.

For patients assessed as unsuitable for the study, investigating the patient features and clinical factors which contribute to this assessment will be an important aspect for understanding the feasibility of current and future pragmatic research studies.

Chief Investigator: Rachael Brookes

Aim:

To describe the mental health status of newly arrived asylum seekers to the UK living in Home Office Contingency Accommodation who have used the RESPOND Refugee Integrated Health Service.

Lay Summary:

The use of hotels to house asylum-seekers has increased substantially over the past three years. The use of these hotels has been linked to an increased risk of poor mental health. An outreach service from University College London Hospital has been providing health-screenings for asylum-seekers housed in eight of these hotels in North Central London. A questionnaire looking for the presence of a mental health need was used as part of this screening appointment. The service has now screened over 1,000 individuals during an 18-month period.

This study will look at the number of screened individuals who had a mental health need and whether this was different depending on the age, gender and country of origin of the individuals. It will also look at things that occurred during the migration journey such as experiencing torture or coming alone or with family to see how this affected the mental health needs of this group. This information will be useful to help plan and inform the mental health services required for this group and to explore in which groups services are most needed.

Chief Investigator: Lauren Borg Xuereb

Aim:

Assessing blood loss proportionately as a percentage of blood volume loss (relative blood loss) correlates better with shock index than quantifying the blood loss in volume (ml) (absolute blood loss).

Lay Summary:

The study design is a retrospective cohort study, at a single Tertiary hospital, using electronic health record system (EHRS) data. We will include women with a post partum haemorrhage of >500ml.

Only operative delivery patients will be included, since high frequency haemodynamic data is available from electronic documentation of intraoperative anaesthesia episodes. Based upon eligibility criteria and a EHRS start date of April 2019, approximately 800 women are expected to have available data.

A comprehensive data set including demographics, obstetric, medical and surgical history, laboratory investigations and clinical characteristics will be collected from the electronic patient record. Haemodynamic variables, including shock index, Shock Index to Transfusion Interval (SITI) will be calculated, as will be the estimated blood volume (ml/kg) and relative blood loss as a percentage.

Chief Investigator: Sohail Bampoe

Aim:

To describe the association between hospital length of stay (HLOS) following major surgery and intraoperative time weighted average (TWA) hypotension <65mm Hg during anaesthesia.

Lay Summary:

Previous research has suggested that low blood pressure during surgery may make patients more likely to suffer medical complications (such as stroke, kidney or heart dysfunction) after surgery. We aim to investigate these concerns in patients who have undergone recent major surgery at UCLH.

Data collected for the study is based upon routine process of care data only, almost all of which is currently being used for service evaluation and improving the quality of care. Whilst we monitor and record blood pressure at regular intervals during anaesthesia, we do not currently have a means of systematically analysing intraoperative hypotension in large groups of patients. This study will overcome that barrier, give us a means of evaluating the presence and effects of low blood pressure during surgery in an accurate way at large scale.

We anticipate the first phase of the study to take three months for data analysis (including data from ~3000 patients), which will then be used to develop a clinical study in which anaesthetists are provided with an advanced technology monitor that might help to reduce the chance of intraoperative low blood pressure happening. We will use data collected during the clinical study and compare with data from our retrospective review to identify any improvement in outcomes. The clinical study is expected to last 6 months and involve ~100 patients.

Chief Investigator:

Prof Sallie Baxendale

Aim:

We have a large database of neuropsychological test scores and clinical data stored on 2,500 people with epilepsy who have been referred to our department for a clinical assessment since 2000. We would like to explore this data to examine which factors influence performance on tests of cognitive function and how these factors interact, so that we can take these factors into account when we interpret the scores of the patients who attend for clinical assessments.

Lay Summary:

We have a large database of neuropsychological test scores and clinical data stored on 2,500 people with epilepsy who have been referred to our department for a clinical assessment since 2000.

We would like to explore this data to examine which factors influence performance on tests of cognitive function and how these factors interact with each other, so that we can take these factors into account when we interpret the scores of the patients who attend for clinical assessments. We are requesting permission to use our data to do identify the important factors that influence performance on neuropsychological tests. We know about some of these factors already, such as the location and type of brain pathology and age, but we don't have a clear idea how these factors interact with educational and cultural factors, particularly when English in not someone's first language . We now have enough data on enough people to look at this.

The results from this research will provide a direct benefit for all patients who are subsequently referred for a neuropsychological assessment both in our department and elsewhere, when the results are disseminated to the wider community of clinicians working with people with epilepsy. The results of these analyses will improve our ability to determine which factors are responsible for cognitive difficulties in this condition. This knowledge will help to direct the clinical team to the provide the correct diagnosis, prognosis and treatment for people with epilepsy who are referred for an evaluation of their cognitive complaints.

Although we will be looking at the impact of some personal and clinical characteristics, such as age and seizure type, the data will be completely anonymised at the time of the analyses and it will not be possible to identify any individual patients in the analyses or any subsequent presentations of the results of our study.

Chief Investigator: Elliott Ridgeon

Aim:

How old must an RCT be before its study population demographics become significantly different from the demographics of the patients we currently treat?

Lay Summary:

Treatments for patients undergoing major abdominal surgery should be based on best evidence, often from clinical trials. Some of the trials we use to inform our practice are older than others. These older trials may have taken place when population demographics were different from those we currently see. We wish to find out 'how old is too old?' for clinical trials, so that we can ensure we provide treatments that have been studied in people who are sufficiently similar to those in our hospitals nowadays.

Chief Investigator: Pinkie Chambers

Aim:

A single-site evaluation of a risk-prediction model enabling the reduction of renal and hepatic function tests that are performed in cancer patients receiving chemotherapy. The overall aim is to evaluate the performance of a risk prediction model using UCLH data.

Lay Summary:

Chemotherapy is the treatment of disease using chemical substances, commonly used to treat cancer. Chemotherapy is often cytotoxic and two common risks relate to kidney and liver damage. Affected patients may require delays or changes to treatment and even treatment suspension.

Whilst only <10% of patients will encounter kidney and/or liver damage, current best practice requires monitoring for all patients with regular blood testing. There is currently no way to stratify risk in individual patients.

The vision for this project is to accurately predict the risk of kidney and liver damage in patients receiving chemotherapy. The result of this prediction will mean low risk patients can be saved unnecessary trips to hospital for blood tests and monitoring whilst high risk patients can receive more appropriate management.

To date, computer scientists at Durham University, working closely with pharmacists at University College London Hospital (UCLH) have developed a machine learning algorithm to accurately predict liver and kidney function. Results have been shared at a number of key conferences and stakeholder events and there is strong clinical interest in utilising this algorithm in clinical practice.

Chief Investigator: Anoop Shah

Aim:

To investigate associations between chronic kidney disease and hyperkalaemia and medication prescribed on discharge; and to develop a prognostic risk score for patients with acute heart failure.

Lay Summary:

In heart failure, the patient’s heart is not able to pump blood around sufficiently. This leads to symptoms of shortness of breath and fluid retention. If not adequately treated or if progression occurs, this can lead to hospitalization, in which the patients have to be admitted for diuretic therapy to get rid of the accumulated fluid. Patients with heart failure often have chronic kidney disease, as inadequate blood supply to the kidneys can lead to failure impairment, and the kidneys are crucial for regulating the fluid balance. In this situation, it is difficult to find the best dose of medication to use, as patients are at risk of complications such as electrolyte abnormalities. It is also difficult to know which patients can be safely treated at home and which patients need to be admitted to hospital for closer monitoring, and these decisions are currently made in an inconsistent way.

In this study, we aim to use real word clinical data to evaluate the prescription of medication in patients with chronic kidney disease who are hospitalized with heart failure. This will lead to improved knowledge of the current implementation of clinical treatment guidelines. Secondly, we aim to develop a risk score for patients admitted with heart failure. This score can provide an estimate of the risk of adverse events during and following the admission. This may help clinicians may help the clinician to decide if a patient should be admitted or managed at home, and what level of monitoring is required. The risk score could also help to predict and optimise the use of health care resources.

For both aims, we will use clinical data extracted from the electronic health record. To develop the risk score, we will use data on patient characteristics at the time that they are admitted to hospital with heart failure, and we will use statistical models to try to predict outcomes such as mortality and readmission (at 30 days and other time periods). The statistical prediction model will be developed by data scientists using the assistance of computer methods, and different types of model will be tested in order to find out which performs best for this task.

The study is being carried out in collaboration with a number of other European centres, so that the findings from all the sites can be combined to produce results that are more accurate and include patients from a broader range of populations. First, we will perform a local analysis using only UCLH data. Secondly, using a federated learning approach, we will combine the results of different hospitals in Europe. We will do this without any data leaving the hospital. Instead, we will carry out a similar analysis in each hospital. Anonymised summary data from each hospital will be combined in order to generate results that include information from all the participating hospitals.

Chief Investigator: Danny Alexander

Aim:

To develop a tool to perform robust, accurate segmentation of the hippocampi from a 3D MR image series using the Microsoft InnerEye platform.

Lay Summary:

The hippocampus plays an important role in neurodegenerative diseases and other neurological conditions such as epilepsy. Reductions in hippocampal volume can provide an early indication of disease but existing, openly available tools for automatically measuring hippocampal volume lack robustness, particularly in patients with conditions that affect hippocampal size and shape. The National Hospital for Neurology and Neurosurgery (NHNN) currently uses an in-house developed software to automatically measure the volume of the hippocampus to support diagnosis of epilepsy patients ,however, although it has performed well in the majority of cases, the existing method suffers from the issues mentioned above and hence it is not always reliable, particularly in older patients and those with advanced or varying disease (e.g. dementia).

Microsoft have developed an artificial intelligence (AI) tool called InnerEye, that can be trained on previously acquired image data to automatically identify and mark out specific regions of the human body, a technique referred to segmentation. A joint project between UCL, UCLH and Microsoft has produced a model using InnerEye that can identify and segment the left and right hippocampal regions from 3D MRI brain scans. This was done by training the software using image data from a publicly available dataset (The Alzheimer’s Disease Neuroimaging Initiative - ADNI). The next step of this project is to setup the model on a UCLH trusted research environment and assess its performance using a reserved set of the ADNI data and then UCLH data. Please note that, Microsoft have made InnerEye open source and the model we are testing is trained on ADNI data only, a UCLH model (trained on UCLH data) will be UCLH’s.

In the long term, it is hoped that InnerEye will prove a more robust method for identifying brain regions such as the hippocampus to improve patient diagnosis and follow-up. If the InnerEye model proves more effective at segmentation compared to our current method, we will look towards using it for Hippocampal segmentation in preference to our existing in-house tool and then look at whether

InnerEye can be similarly trained to produce models for segmentation of other brain regions.

Chief Investigator: Talisa Ross

Aim:

To assist in operative planning through the use of automated image segmentation; and to identify abnormalities in middle ear structures through the use of machine learning.

Lay Summary:

Care for patients with hearing loss is inconsistent and fails to reach those who struggle the most and fails to meet patient expectations in those we reach. Research for hearing loss, whilst recently ground breaking, is rarely translated into patient benefit. Our hospital is in the unique position to help.

We have the largest store of hearing healthcare data in the country, collected over many visits on hearing loss patients for more than 20 years. By leveraging the UCLH data science expertise to standardise and unify hearing health related datasets we will be able to use machine learning techniques to draw new insights on the identification of hearing loss types. This will help us better understand the underlying reasons for an individual's hearing loss.

This knowledge will be used to implement monitoring and following up strategies as well as target traditional treatments more effectively and novel treatment more precisely for their specific type of hearing loss. Additionally, we will use our algorithms to improve the healthcare documentation around hearing loss so that safety and effectiveness of treatments can be monitored more rigorously. We will initially build a dataset of scans of the ear for patients who we look after with hearing loss.

After converting these scans into a standardised format, we will build a computer vision algorithm which can analyse the bones of hearing. This will generate important information to diagnose and treat our patients with hearing loss. The models that we build will be fine-tuned and validated on our large dataset.

Chief Investigator: Edward Palmer

Aim:

To develop and evaluate statistical methodology to enable estimation of the effects of different treatment strategies on risk of severe outcomes for seriously ill patients using routinely collected intensive care data, and to apply the methods to investigate the effects of different ventilation strategies on risk of death for patients hospitalised with Acute Respiratory Distress Syndrome (ARDS), COVID-19 and respiratory failure.

Lay Summary:

Over 470,000 people have been hospitalised due to COVID-19 in the UK to date. 36,000 of these required intensive care, of which the majority received breathing support, known as ventilation.

Ventilation can be “invasive” or “non-invasive”. Invasive ventilation is where a machine does the breathing for a patient, who must be sedated and have a breathing tube put into their windpipe. Non-invasive ventilation is where a patient receives air and oxygen via a tightly-fitting mask at high pressure. Around 40% of COVID-19 patients admitted to intensive care have received invasive ventilation, however, there is debate over whether and when to introduce invasive ventilation in COVID-19 patients.

The ideal way to study whether one treatment is more effective than another is through a randomised controlled trial (RCT). This involves randomly choosing half of a group of patients to receive one treatment and half to receive another, enabling a fair comparison between treatments.

However, RCTs are limited because they are generally not designed to look at multiple treatments or treatments that vary over time. Furthermore, RCTs may not possible due to financial, logistical or ethical constraints. Meanwhile, there is a growing availability of “observational” data routinely collected when patients are admitted to hospital or attend their GP, known as “electronic health records”. These data are often available on a large number of people and reflect how treatments are used in day-to-day care, providing a useful resource for examining the effects of different treatments.

However, using such data has challenges since the treatments are not randomly allocated. The data have complicated features that must be dealt with using tailored statistical methods, and in recent years there have been many advances in such methods.

In this project, we will use and compare existing statistical methods and develop new methods to evaluate the effects of these ventilation based treatments using routinely collected data. We will focus on two key statistical challenges. First, the fact that measurements on patients’ clinical status are collected at irregular times and on differing schedules for each patient. This presents a challenge as most of the statistical methods assume measurements are made at regular time intervals.

Second, we will examine the effects of both “dynamic” and “static” ventilation treatment strategies. A “static” treatment strategy is where the treatment received is the same for all patients meeting defined criteria, e.g. “what would happen if all patients receive invasive ventilation within 24 hours of admission to the intensive care unit”. A “dynamic” treatment strategy is where there are rules that adapt the treatment to the health status of the patient, e.g. “what would happen if patients receive invasive ventilation only after they deteriorate to a particular point”.

We will apply these methods to investigate the effectiveness of invasive and non-invasive ventilation on outcomes of patients with COVID-19 and other forms of severe breathing problems that may require the use of ventilation. The data source is routinely collected hospital data of all patients with COVID-19 or a lung condition that places them at risk of needing ventilation admitted to University College London Hospital (UCLH) since January 2020. These data are collected as part of the normal delivery of care of patients, for example: blood pressure measurements and vital signs taken by nurses, the results of daily blood tests and diagnoses made by doctors. Our group is ideally positioned to tackle with research question, as a collaboration between UCLH, UCL and the LSHTM, representing world leading experts from the fields of intensive care medicine and causal inference.

We will seek involvement from patients who have spent time in intensive care on ventilation to understand experiences of invasive vs non-invasive ventilation, and seek advice on communicating the outcomes of this research. This work will give patients, doctors and policy makers more information about ventilation strategies, allowing them to make more informed decisions about the best ventilation strategy tailored to the patient. It will impact future research by providing methods that enable best possible use of routinely collected hospital data to benefit patients with other diseases.

Chief Investigator: Joseph Jacob, Steve Harris

Aim:

To develop a machine learning model for detecting and qualifying nasogastric tube (NGT) placement on chest x radiographs (CXR) with the prospective aim of facilitating the timely correction and prevention of further complications associated with incorrectly positioned feeding tubes within clinical practice thereby increasing patient safety.

Lay Summary:

A Nasogastric tube (NGT) is a thin tube that is passed into the stomach via the nose for short- to medium-term nutritional support, medication administration or aspiration of stomach contents. NGTs are amongst the most commonly used catheters in critically ill patients in intensive care units (ICU) and high-dependency units and departments where patients require nutritional-support (i.e., Stroke units). Due to increases in the number of hospitalized patients, it is estimated that approximately 10 million NGTs are used annually in Europe, 1 million of which in the UK (~1.2 million in the US) [15].

Previous research highlights a variety of complications associated with NGT placement, which can range from minor cases of nose bleeds to inhalation of stomach contents into the lung and even death. Instances of unknowingly misplaced NGTs being used for feeding, with the feed entering the patients lungs are classified by the NHS as Never Events: “serious incidents that are entirely preventable because guidance or safety recommendations providing strong systemic protective barriers are available at a national level, and should have been implemented by all healthcare providers” .

While all this highlights the importance for feeding tubes in particular to be placed properly and used safely [16], clinical studies demonstrate that up to 3% of NGTs are reported as misplaced into the airways, causing complications in up to 40% of these cases [6]. Given the serious complications that can occur from NGT misplacement, UCLH has a detailed policy describing the indications and technique of NGT insertion alongside nationally agreed standards for positioning verification. This includes training and guidelines for doctors or reporting radiographers when checking NGT position radiographically. In this policy, the first line of test in confirming the correct positioning of a feeding tube is by obtaining a sample of fluid from the stomach that shows a level of acidity indicative of the stomach. However, since this cannot be achieved successfully for some patients, and with a large proportion of ICU patients receiving anti-acid medication, the use of CXRs remains the most definitive test for checking NGT placement (see Appendix B for NHNN workflow example).

Due to the large number of CXRs obtained each day, especially in intensive and emergency care, and with a limited number of radiologists available, image interpretation can be substantially delayed [2][16]. Thus, current practices indicate that it is often emergency and ICU doctors who check the CXR to verify the NGT’s correct positioning and suitability for use [13] prior to the radiology report being issued [14]. Yet, such assessments by non-radiologists working in stressful situations when hospitals are capacity, are prone to both human error and some delays in assessment [3]. This means that sub-optimally positioned NGTs can be missed initially, but are often picked up by the radiologists later. This emphasizes the importance of early detection of misplaced NGTs to allow for more timely correction and prevent any additional complications [14].

We envision two main use scenarios in which an accurate, instant detection and notification of NGT misplacements from CXRs could benefit clinical practice:

(1) As an early alert to ICU doctors or nurses, it will enable prompt, data driven decision-making and NGT adjustment for more effective and safe use.

(2) As an early alert to help prioritize the review of most urgent CXRs by local (UCLH) radiologists to reduce delays in notifying ICU doctors of potentially unrecognized NGT misplacement.

Our study, initially focussed on ICU, will develop computer based models that can detect the correct placement of an NGT and alert radiologists to examine CXRs where the NGT appears incorrectly placed. Once an abnormally positioned NGT is detected on a CXR an automatic alert will also be sent out to the ICU teams informing them to cross-check the expedited radiology report and that the CXR likely needs replacing. In parallel, we will study requirements for expanding this system to other inpatient areas that frequently place NGTs. This may include Stroke Departments, all High Dependency Units with post-surgery patients (particularly gastrointestinal surgery patients), Geriatric Wards, as well as other UCLH hospitals such as the National Hospital for Neurology and Neurosurgery.

Chief Investigator: Professor Harry Hemingway

Aim:

The DOME initiative aims to develop and deliver data infrastructure for UCLH to understand all the diseases it diagnoses and manages, and to map disease definitions in the electronic health record, with NICE and other clinical practice guidelines. Because UCLH is a tertiary hospital DOME spans common and rare diseases. We envision using this data framework as a means to connect doctors at UCLH with academics at UCL – who often have highly specialized disease and disease mechanism interests. It is anticipated this work will help drive innovation and translation for patient benefit not only locally but also nationally and internationally.

Lay Summary:

One of UCLH's central objectives is to provide high quality care for all the patients diagnosed and managed as in patients and out patients. This proposal seeks to take initial steps in fulfilling this objective by analysing anonymised UCLH electronic health records of all patients. Importantly, we consider this a ‘democratic’ approach whereby every patient and every diagnosis matters. The steps that this proposal aims to address are as follows:

1. We seek to define each disease (disorder / syndrome) that a patient attending UCLH between April 2019 onwards has been diagnosed or presented with. For this purpose, initially we seek to use the coded (structured) electronic record data. But because most patients seen at UCLH are outpatients (never admitted) and because most out-patients do not have a diagnosis we will explore the use of NLP of unstructured data to ‘close the gap’ in diagnosis coding. TFor every disease defined in data (we anticipate many hundreds), we will identify national standards of high quality care. Often these clinical practice guidelines will come from NICE, but because UCLH diagnoses and manages many patients with uncommon and rare diseases, the definitions of high quality care may also come from European and American professional bodies. We think this step to be feasible because proof of concept work using national data has already been conducted.

2. If we are successful at UCLH with defining disease and guidelines, a next step would be to involve clinicians and patients to identify ‘beacon recommendations’ from each disease guideline which are ‘Epic ready’. We believe that sometimes health systems create digital divides in which patients with some diseases are more likely to be included in quality of care research and initiatives than patients with other diseases. Clinical judgement is required to evaluate NICE clinical practice recommendations: for example ‘risk rating’ (impact on patient outcomes of non-compliance), ‘digital readiness’ (extent to which Epic can implement today) and measurable practice variation. DOME investigators have used formal methods (Delphi expert ratings) to capture clinical judgements and assess the extent of any consensus. If the above two steps are addressed, we believe that UCLH has the opportunity to be the first hospital (nationally and possibly internationally) to develop a system-wide ‘all patients, all diseases’ approach to quality care and treatment.