**BACKGROUND** Every year over **40% of the UK population take 770 million regular medications** to manage and prevent disease. Their General Practitioner (GP) has the vast task of ensuring they are on the right medications, **with 15% of the population taking more than 5 medications regularly, and 25% having multiple long-term conditions.** **PROBLEM** In an over-stretched health system, medications can be... \***Overused**: where unnecessary, potentially harmful medicines **should be** **stopped** because of side effects, interactions or ineffectiveness \***Underused**: where vital, disease-preventing medicines **should be** **started** Both have health consequences and cost implications for the NHS that are entirely avoidable. **CURRENT PRACTICE** Doctors/Pharmacists are expected to ... 1. look through large volumes of medical data, with information 'hidden' in the free-text notes 2.Collect additional information from patients 3.Maintain up-to-date knowledge of ever-changing clinical guidelines for each medicine and disease 4.Identify potential side effects and contra-indications 5.Escalate therapy when the treatment is not working **NEED** Regular reviews for all patients, putting into practice already existing, clear, guidelines would identify these errors with **continuous, clinical, decisions individualised to each patient**. The **potential benefits of improved care** are enormous, an estimated **£2.5 billion per year to the NHS** across the UK. The **NHS lacks capacity** to do this safely due to: \*Critical staffing shortages exacerbated by COVID-19 \*Requirements to repeat the whole process for each person, for every change in health data, regularly throughout the year to capture any safety issues quickly **SOLUTION** This **_"Auto-Pharmacist"_** will continuously, automatically extract and rationalise multiple data sources including existing patient EHR records against up-to-date clinical guidelines, organising automated patient questionnaires, flag potential adverse side effects and further testing where required. This solution is an **opportunity to deliver population health delivered efficiently and effectively** whilst reducing medication-related harms.

More than a thousand patients are diagnosed with cancer daily in the UK - **1 in 2 people will get this diagnosis in their lifetime**. Many cancers being diagnosed at an advanced stage when treatments are less effective. The UK has poorer cancer survival rates compared to other high-income countries. The long-term plan aims to be diagnosing 75% of cancer at an early stage -- currently **less than half of all cancers diagnosed at this stage when treatment could be more effective and less costly**. Delayed diagnosis has a significant impact on survival. In bowel cancer as an example, it can reduce survival rates from 91% to 10%. **Primary care (general practices) play an important role in raising suspicion of a hidden cancer** and referring a patient for diagnostic testing as early as possible. To be able to make an appropriate referral, the clinician needs to be able to estimate the risk for each patient based on their specific symptoms. Once identified, patients require further confirmatory tests including blood tests, scans and biopsies. **Improving this individual risk assessment would improve both the pickup rate and the speed through this pathway meaning more patients are identified, fewer patients are unnecessarily tested and diagnosis is made as soon as possible.** Electronic Health Records store patients' health conditions, tests requested, results and referrals. Information about early signals of cancer can be found in structured variables often stored in tables (vital signs, lab results, measurements). However**, key information is more often in clinical notes**, which contain narrative text used by General Practitioners (GPs) to record patients' symptoms. Recent advancements in statistical techniques and machine learning have improved the ability of **machines to understand data stored as language**. These Large Language Models can "read" large volumes of text rapidly, potentially for every patient, every day, to look for signs which might suggest they have developed a disease like cancer. This grant will fund a health-tech company, as well as GP practices and experts in evaluating the clinical safety and impact of medical algorithms. This group will use machine learning techniques including Large Language Models to detect early signs of cancer months in advance by continuously scanning patient records and warning their doctors when concerning signs are identified. The resulting software and new medical approach is **expected deliver 60,000 cancer diagnoses earlier in the UK annually. The technology will then be expanded to other disease types.**

The **NHS and health systems around the world are under great pressure** with rising A&E attendances and hospital admissions alongside **budgetary pressures,** backlogs and an increasingly **ageing, less healthy population**. Hospitals are overflowing. However, **1 out of 5** emergency admissions are **avoidable** and responsible for taking up 13% of all NHS hospital beds, costing **£2.5 billion each year** with **40% of all deaths** having **preventable** causes. Given the current stresses on the healthcare budget and wider economy, increasing staff cannot be the only solution - **the equivalent of 8800 full-time GPs and 6400 Nurse posts will be vacant by 2030**. In this context, digital solutions are recognised to be essential by NHS England's Long-term Plan. **GP surgeries are responsible for monitoring patients**, reviewing them once or more each year if they have certain risk factors. However, in between these appointments, patients often have **subtle, gradual, often unrecognised worsening** over weeks or months before becoming unwell enough to require urgent care. A **wearable device could have picked up these signals** early, allowing assessment, treatment and avoiding deterioration. Remote Monitoring data, processed through machine learning based technology data creates an opportunity to build **an early-detection system for** patients who are at home with silent signals of **worsening health**. This project would allow **better targeting of _existing_ prevention and monitoring** activities and uses existing clinical capacity more effectively and efficiently, **enabling health systems _prevent_ hospitalisations,** patient harm whilst reducing health spending. **Wearable and remote monitoring devices are increasingly commonly used by patients including smartwatches (15%), fitness trackers (21%), and smartphones with step counters (90%)**. Furthermore, 6 out of 10 patients would agree to share this data with their doctor. All relevant information about a patient's health condition and their planned appointments will be in electronic health records (EHRs). This project will deliver a system which **combines data currently available to the GP** with additional **data collected from patients at home** using existing devices they might already have to. The system will identify and prioritise patients, suggesting modifications to their routine monitoring schedule allowing the GP to react early which a patient is suspected to be deteriorating by bringing them in for a more urgent appointment. The overall number of annual appointments is expected to be the same, but patients will be brought in at the right time for them, guided by rich data sources and advanced data processing techniques.

In addition to _treating_ disease, healthcare staff in primary care try to _prevent_ disease. Patients who meet clinical risk criteria can benefit from actions like starting medications before they develop a disease. These _preventative interventions_ reduce the risk of disease or delay it. **40% of all yearly deaths are thought to be due to preventable causes**\[1\]**,** signifying the immense potential to improve life span and quality of life. Prevention is the most cost-effective approach compared to the costs and burden on the healthcare system of managing and treating established disease. Clear guidelines are made available by institutions like the National Institute for Health and Care Excellence (NICE). **Delivering prevention is simple and inexpensive -- finding which patients, need which intervention, at what time is a huge and expensive task**. For example, approximately 800,000 out of 18 million patients who meet initial criteria for risk assessment\[2\] would benefit from bisphosphonates, an affordable, easily available drug that strengthens bones and reduces the risk of hip fractures -- a disabling, costly condition. Finding these individuals at risk is currently a complex, repetitive and manual task that involves asking the right patients the right questions and processing this information through complex clinical guidelines. This task is the responsibility of General Practices who are also responsible for managing existing disease and struggling with workload and staffing issues made worse by the COVID pandemic. Current digital solutions only support fragments of this process such as running searches or sending text messages making automation impossible and still requiring multiple touchpoints from clinicians. As a result, coverage is poor -- for example, only 8.2% of patients who fractured their hip were on protective medications\[3\]. This is just one of 33 primary care guidelines promoted by NICE on prevention\[4\] which could be delivered more effectively, uniformly and efficiently. The grant will fund a health-tech company, as well as GPs, a Clinical Commissioning Group, and experts in measuring economic benefits and clinical safety, will deliver the platform to 250,000 patients over the next two years. This project will deliver a software that will continuously find patients who might benefit from interventions that prevent hip fractures, heart attacks and strokes. It will automatically contact those patients to collect specific information required to identify those who need preventative treatment. The evidence collected will allow the Clinical Intelligent Automation platform to be delivered throughout the UK and internationally within 5 years.

**Challenge to address** At every stage of the Patient inflow and outflow process within the NHS, there are delays which impact on patient care, staff, hospital resources and efficiency. The vast majority of delays are attributable to hospitals and staff not having access to information or that information being incomplete or out-of-date, which in turn leads to the mis-allocation of scarce resources (staff, beds, materials). With no tools or mechanisms to predict admission level or the expected length of stay - current leading platforms only provide time-delayed data - hospitals and staff need to make reactive, subjective decisions regarding beds, staffing and resources. Whilst clinician experience will always be important role, AI has the ability to harness, analyse and support decision-making and resource allocation in a quick, accurate and standardised way (analysing years of big data). **Solution** We are proposing to develop the first highly accurate AI real-time predictor of hospital admission and patient length of stay. Projected benefits include: 1.Predict resourcing requirements (bed, staff, equipment, etc) based upon large historical health datasets. 2.Help better manager whole-hospital bed occupancy status and resources in the short, medium and long term. 3.Speed up the discharge of patients by providing accurate real-time information to clinicians. 4.Predict peaks and troughs in demand **Innovation** The base information collected will be derived from local electronic health record datasets with the predictive core of the platform based on machine learning models that can accurately analyse both structured (numerical and categorical values) and unstructured (text-based information like triage and physicians' notes) large datasets. Such clinical algorithms will be trained on millions of data points. This process leads to accurate, actionable intelligence for clinicians and management teams to act on. For example, the engine will learn that for a given number of patients presenting with a high NEWS score or low oxygen saturations, a proportion will be admitted. With accumulation of clinical information on COVID-19 patients, increasingly accurate predictions for admission and length of stay will be generated. Such information will then be relayed to clinicians and hospital leads in real-time so that resources can be accurately ordered/allocated, discharge assessments planned and patients aren't kept in any longer than needed. **Impact** With delays to discharge purely from untimely information costing the NHS 625,942 bed-days/£92.6m per annum alone, the technology is timely and urgently needed.

UK cancer survival rates lag behind other developed countries(1), up to 10,000 excess deaths occur annually in comparison(2). Delayed diagnosis is thought to contribute to this(3). Currently nearly half of all cancers are diagnosed at a late stage(4). Approximately 20% of patients see their GP 3 or more times before their cancer diagnosis(7) with approximately 12% of avoidable diagnosis delay occurring in in this setting(6). To diagnose cancer, multiple features ranging from subtle findings to non-specific symptoms must be considered - these are recorded across different sections of the electronic health record. In addition, in up to 59% of records this data is 'hidden': doctors often write them in the free-text records(10) outside of rigid coding frameworks, where they cannot be identified by routine means. With increasing demand and complexity, primary care clinicians struggle in a resource-strained NHS: GPs have on average 9 minutes per patient(9). In this time, identifying this relevant data accurately and with speed has potential for error: indeed, often tell-tale patterns predictive of cancer are already present in the records long before a patient is finally diagnosed(8) There is a national drive towards earlier detection of cancer, ranging from new diagnostic centres, incentives for primary care and regional Cancer Alliances. Discussions with these stakeholders have shown us that there is a strong appetite and need for an intelligent method of supporting primary care cancer detection. Clinical decision support systems (CDSS) have been identified as a key tool for early diagnosis by major national cancer stakeholders in a Cancer Research UK 2020 report(11). Our aim is to develop a CDSS to diagnose cancer earlier, and reduce delayed or missed diagnoses, by using key information in health records. We will machine learning and natural language processing, AI methods which enable analysis of large, complex datasets. We will evaluate the CDSS to ensure it is able to detect cancers accurately, and then assess its performance in real-world clinical practice to assess its impact on cancer diagnosis by GPs. By leveraging powerful AI methods and the type of data unreachable by existing tools, we believe we can make a difference to cancer diagnosis whilst supporting clinicians and the health service. The end goal is for patients to be diagnoses earlier, treated more effectively, live longer, healthier lives with downstream impacts on the economy through reduced morbidity and pressures on the healthcare system.

**Challenge to address** At every stage of the Patient inflow and outflow process within the NHS, there are delays which impact on patient care, staff, hospital resources and efficiency. The vast majority of delays are attributable to hospitals and staff not having access to information or that information being incomplete or out-of-date, which in turn leads to the mis-allocation of scarce resources (staff, beds, materials). With no tools or mechanisms to predict admission level or the expected length of stay - current leading platforms only provide time-delayed data - hospitals and staff need to make reactive, subjective decisions regarding beds, staffing and resources. Whilst clinician experience will always be important role, AI has the ability to harness, analyse and support decision-making and resource allocation in a quick, accurate and standardised way (analysing years of big data). **Solution** We are proposing to develop the first highly accurate AI real-time predictor of hospital admission and patient length of stay (AUC\>0.9). Projected benefits include: 1. Predict resourcing requirements (bed, staff, equipment, etc) based upon large historical health datasets. 2. Predict critical resource supply gaps such as PPE and oxygen (COVID-19 identified failure-point); 3. Speed up the discharge of patients by providing accurate real-time information to clinicians. 4. Help better manager whole-hospital bed occupancy status and resources in the short, medium and long term. **Innovation** The base information collected will be derived from local electronic health record datasets with the predictive core of the platform based on machine learning models that can accurately analyse both structured (numerical and categorical values) and unstructured (text-based information like triage and physicians' notes) large datasets. Such clinical algorithms will be trained on millions of data points. This process leads to accurate, actionable intelligence for clinicians and management teams to act on. For example, the engine will learn that for a given number of patients presenting with a high NEWS score or low oxygen saturations, a proportion will be admitted. With accumulation of clinical information on COVID-19 patients, increasingly accurate predictions for admission and length of stay will be generated. Such information will then be relayed to clinicians and hospital leads in real-time so that resources can be accurately ordered/allocated, discharge assessments planned and patient's aren't kept in any longer than needed. With delays to discharge purely from untimely information costing the NHS 625,942 bed-days/£92.6m per annum alone, the technology is timely and urgently needed.