Search

Abstract Background Heart failure (HF) with preserved or mildly reduced ejection fraction includes a heterogenous group of patients. Reclassification into distinct phenogroups to enable targeted interventions is a priority. This study aimed to identify distinct phenogroups, and compare phenogroup characteristics and outcomes, from electronic health record data. Methods 2,187 patients admitted to five UK hospitals with a diagnosis of HF and a left ventricular ejection fraction ≥ 40% were identified from the NIHR Health Informatics Collaborative database. Partition-based, model-based, and density-based machine learning clustering techniques were applied. Cox Proportional Hazards and Fine-Gray competing risks models were used to compare outcomes (all-cause mortality and hospitalisation for HF) across phenogroups. Results Three phenogroups were identified: (1) Younger, predominantly female patients with high prevalence of cardiometabolic and coronary disease; (2) More frail patients, with higher rates of lung disease and atrial fibrillation; (3) Patients characterised by systemic inflammation and high rates of diabetes and renal dysfunction. Survival profiles were distinct, with an increasing risk of all-cause mortality from phenogroups 1 to 3 (p

Abstract Background Pregnancy acts as a cardiovascular stress test. Although many complications resolve following birth, women with hypertensive disorder of pregnancy have an increased risk of developing cardiovascular disease (CVD) long-term. Monitoring postnatal health can reduce this risk but requires better methods to identity high-risk women for timely interventions. Methods Employing a qualitative descriptive study design, focus groups and/or interviews were conducted, separately engaging public contributors and clinical professionals. Diverse participants were recruited through social media convenience sampling. Semi-structured, facilitator-led discussions explored perspectives of current postnatal assessment and attitudes towards linking patient electronic healthcare data to develop digital tools for identifying postpartum women at risk of CVD. Participant perspectives were gathered using post-it notes or a facilitator scribe and analysed thematically. Results From 27 public and seven clinical contributors, five themes regarding postnatal check expectations versus reality were developed, including ‘limited resources’, ‘low maternal health priority’, ‘lack of knowledge’, ‘ineffective systems’ and ‘new mum syndrome’. Despite some concerns, all supported data linkage to identify women postnatally, targeting intervention to those at greater risk of CVD. Participants outlined potential benefits of digitalisation and risk prediction, highlighting design and communication needs for diverse communities. Conclusions Current health system constraints in England contribute to suboptimal postnatal care. Integrating data linkage and improving education on data and digital tools for maternal healthcare shows promise for enhanced monitoring and improved future health. Recognised for streamlining processes and risk prediction, digital tools may enable more person-centred care plans, addressing the gaps in current postnatal care practice.

BackgroundThere has been a substantial increase in the development of artificial intelligence (AI) tools for clinical decision support. Historically, these were mostly knowledge-based systems, but recent advances include non–knowledge-based systems using some form of machine learning. The ability of health care professionals to trust technology and understand how it benefits patients or improves care delivery is known to be important for their adoption of that technology. For non–knowledge-based AI tools for clinical decision support, these issues are poorly understood. ObjectiveThe aim of this study is to qualitatively synthesize evidence on the experiences of health care professionals in routinely using non–knowledge-based AI tools to support their clinical decision-making. MethodsIn June 2023, we searched 4 electronic databases, MEDLINE, Embase, CINAHL, and Web of Science, with no language or date limit. We also contacted relevant experts and searched reference lists of the included studies. We included studies of any design that reported the experiences of health care professionals using non–knowledge-based systems for clinical decision support in their work settings. We completed double independent quality assessment for all included studies using the Mixed Methods Appraisal Tool. We used a theoretically informed thematic approach to synthesize the findings. ResultsAfter screening 7552 titles and 182 full-text articles, we included 25 studies conducted in 9 different countries. Most of the included studies were qualitative (n=13), and the remaining were quantitative (n=9) and mixed methods (n=3). Overall, we identified 7 themes: health care professionals’ understanding of AI applications, level of trust and confidence in AI tools, judging the value added by AI, data availability and limitations of AI, time and competing priorities, concern about governance, and collaboration to facilitate the implementation and use of AI. The most frequently occurring are the first 3 themes. For example, many studies reported that health care professionals were concerned about not understanding the AI outputs or the rationale behind them. There were issues with confidence in the accuracy of the AI applications and their recommendations. Some health care professionals believed that AI provided added value and improved decision-making, and some reported that it only served as a confirmation of their clinical judgment, while others did not find it useful at all. ConclusionsOur review identified several important issues documented in various studies on health care professionals’ use of AI tools in real-world health care settings. Opinions of health care professionals regarding the added value of AI tools for supporting clinical decision-making varied widely, and many professionals had concerns about their understanding of and trust in this technology. The findings of this review emphasize the need for concerted efforts to optimize the integration of AI tools in real-world health care settings. Trial RegistrationPROSPERO CRD42022336359; https://tinyurl.com/2yunvkmb

Abstract Aims Population‐wide, person‐level, linked electronic health record data are increasingly used to estimate epidemiology, guide resource allocation, and identify events in clinical trials. The accuracy of data from NHS Digital (now part of NHS England) for identifying hospitalization for heart failure (HHF), a key HF standard, is not clear. This study aimed to evaluate the accuracy of NHS Digital data for identifying HHF. Methods and results Patients experiencing at least one HHF, as determined by NHS Digital data, and age‐ and sex‐matched patients not experiencing HHF, were identified from a prospective cohort study and underwent expert adjudication. Three code sets commonly used to identify HHF were applied to the data and compared with expert adjudication (I50: International Classification of Diseases‐10 codes beginning I50; OIS: Clinical Commissioning Groups Outcomes Indicator Set; and NICOR: National Institute for Cardiovascular Outcomes Research, used as the basis for the National Heart Failure Audit in England and Wales). Five hundred four patients underwent expert adjudication, of which 10 (2%) were adjudicated to have experienced HHF. Specificity was high across all three code sets in the first diagnosis position {I50: 96.2% [95% confidence interval (CI) 94.1–97.7%]; NICOR: 93.3% [CI 90.8–95.4%]; OIS: 95.6% [CI 93.3–97.2%]} but decreased substantially as the number of diagnosis positions expanded. Sensitivity [40.0% (CI 12.2–73.8%)] and positive predictive value (PPV) [highest with I50: 17.4% (CI 8.1–33.6%)] were low in the first diagnosis position for all coding sets. PPV was higher for the National Heart Failure Audit criteria, albeit modestly [36.4% (CI 16.6–62.2%)]. Conclusions NHS Digital data were not able to accurately identify HHF and should not be used in isolation for this purpose.

Objectives To compare the patterns of multimorbidity between people with and without rheumatic and musculoskeletal diseases (RMDs) and to describe how these patterns change by age and sex over time, between 2010 and 2019.Participants 103 426 people with RMDs and 2.9 million comparators registered in 395 Wales general practices (GPs). Each patient with an RMD aged 0–100 years between January 2010 and December 2019 registered in Clinical Practice Research Welsh practices was matched with up to five comparators without an RMD, based on age, gender and GP code.Primary outcome measures The prevalence of 29 Elixhauser-defined comorbidities in people with RMDs and comparators categorised by age, gender and GP practices. Conditional logistic regression models were fitted to calculate differences (OR, 95% CI) in associations with comorbidities between cohorts.Results The most prevalent comorbidities were cardiovascular risk factors, hypertension and diabetes. Having an RMD diagnosis was associated with a significantly higher odds for many conditions including deficiency anaemia (OR 1.39, 95% CI (1.32 to 1.46)), hypothyroidism (OR 1.34, 95% CI (1.19 to 1.50)), pulmonary circulation disorders (OR 1.39, 95% CI 1.12 to 1.73) diabetes (OR 1.17, 95% CI (1.11 to 1.23)) and fluid and electrolyte disorders (OR 1.27, 95% CI (1.17 to 1.38)). RMDs have a higher proportion of multimorbidity (two or more conditions in addition to the RMD) compared with non-RMD group (81% and 73%, respectively in 2019) and the mean number of comorbidities was higher in women from the age of 25 and 50 in men than in non-RMDs group.Conclusion People with RMDs are approximately 1.5 times as likely to have multimorbidity as the general population and provide a high-risk group for targeted intervention studies. The individuals with RMDs experience a greater load of coexisting health conditions, which tend to manifest at earlier ages. This phenomenon is particularly pronounced among women. Additionally, there is an under-reporting of comorbidities in individuals with RMDs.

BackgroundMost existing clinical prediction models do not allow predictions under interventions. Such predictions allow predicted risk under different proposed strategies to be compared and are therefore useful to support clinical decision making. We aimed to compare methodological approaches for predicting individual level cardiovascular risk under three interventions: smoking cessation, reducing blood pressure, and reducing cholesterol.MethodsWe used data from the PREDICT prospective cohort study in New Zealand to calculate cardiovascular risk in a primary care setting. We compared three strategies to estimate absolute risk under intervention: (a) conditioning on hypothetical interventions in non-causal models; (b) combining existing prediction models with causal effects estimated using observational causal inference methods; and (c) combining existing prediction models with causal effects reported in published literature.ResultsThe median absolute cardiovascular risk among smokers was 3.9%; our approaches predicted that smoking cessation reduced this to a median between a non-causal estimate of 2.5% and a causal estimate of 2.8%, depending on estimation methods. For reducing blood pressure, the proposed approaches estimated a reduction of absolute risk from a median of 4.9% to a median between 3.2% and 4.5% (both derived from causal estimation). Reducing cholesterol was estimated to reduce median absolute risk from 3.1% to between 2.2% (non-causal estimate) and 2.8% (causal estimate).ConclusionsEstimated absolute risk reductions based on non-causal methods were different to those based on causal methods, and there was substantial variation in estimates within the causal methods. Researchers wishing to estimate risk under intervention should be explicit about their causal modelling assumptions and conduct sensitivity analysis by considering a range of possible approaches.

Summary: Background: Understanding and quantifying the differences in disease development in different socioeconomic groups of people across the lifespan is important for planning healthcare and preventive services. The study aimed to measure chronic disease accrual, and examine the differences in time to individual morbidities, multimorbidity, and mortality between socioeconomic groups in Wales, UK. Methods: Population-wide electronic linked cohort study, following Welsh residents for up to 20 years (2000–2019). Chronic disease diagnoses were obtained from general practice and hospitalisation records using the CALIBER disease phenotype register. Multi-state models were used to examine trajectories of accrual of 132 diseases and mortality, adjusted for sex, age and area-level deprivation. Restricted mean survival time was calculated to measure time spent free of chronic disease(s) or mortality between socioeconomic groups. Findings: In total, 965,905 individuals aged 5–104 were included, from a possible 2.9 m individuals following a 5-year clearance period, with an average follow-up of 13.2 years (12.7 million person-years). Some 673,189 (69.7%) individuals developed at least one chronic disease or died within the study period. From ages 10 years upwards, the individuals living in the most deprived areas consistently experienced reduced time between health states, demonstrating accelerated transitions to first and subsequent morbidities and death compared to their demographic equivalent living in the least deprived areas. The largest difference were observed in 10 and 20 year old males developing multimorbidity (−0.45 years (99% CI: −0.45, −0.44)) and in 70 year old males dying after developing multimorbidity (−1.98 years (99% CI: −2.01, −1.95)). Interpretation: This study adds to the existing literature on health inequalities by demonstrating that individuals living in more deprived areas consistently experience accelerated time to diagnosis of chronic disease and death across all ages, accounting for competing risks. Funding: UK Medical Research Council, Health Data Research UK, and Administrative Data Research Wales.

In 2020, we published an editorial about the massive disruption of health and care services caused by the COVID-19 pandemic and the rapid changes in digital service delivery, artificial intelligence and data sharing that were taking place at the time. Now, 3 years later, we describe how these developments have progressed since, reflect on lessons learnt and consider key challenges and opportunities ahead by reviewing significant developments reported in the literature. As before, the three key areas we consider are digital transformation of services, realising the potential of artificial intelligence and wise data sharing to facilitate learning health systems. We conclude that the field of digital health has rapidly matured during the pandemic, but there are still major sociotechnical, evaluation and trust challenges in the development and deployment of new digital services.

Abstract Backgrounds We aimed to develop and validate machine learning (ML) models for 30-day stroke mortality for mortality risk stratification and as benchmarking models for quality improvement in stroke care. Methods Data from the UK Sentinel Stroke National Audit Program between 2013 to 2019 were used. Models were developed using XGBoost, Logistic Regression (LR), LR with elastic net with/without interaction terms using 80% randomly selected admissions from 2013 to 2018, validated on the 20% remaining admissions, and temporally validated on 2019 admissions. The models were developed with 30 variables. A reference model was developed using LR and 4 variables. Performances of all models was evaluated in terms of discrimination, calibration, reclassification, Brier scores and Decision-curves. Results In total, 488,497 stroke patients with a 12.3% 30-day mortality rate were included in the analysis. In 2019 temporal validation set, XGBoost model obtained the lowest Brier score (0.069 (95% CI: 0.068–0.071)) and the highest area under the ROC curve (AUC) (0.895 (95% CI: 0.891–0.900)) which outperformed LR reference model by 0.04 AUC (p 15%). The XGBoost model reclassified 1648 (8.1%) low-risk cases by the LR reference model as being moderate or high-risk and gained the most net benefit in decision curve analysis. Conclusions All models with 30 variables are potentially useful as benchmarking models in stroke-care quality improvement with ML slightly outperforming others.

Rates of Multimorbidity (also called Multiple Long Term Conditions, MLTC) are increasing in many developed nations. People with multimorbidity experience poorer outcomes and require more healthcare intervention. Grouping of conditions by health service utilisation is poorly researched. The study population consisted of a cohort of people living in Wales, UK aged 20 years or older in 2000 who were followed up until the end of 2017. Multimorbidity clusters by prevalence and healthcare resource use (HRU) were modelled using hypergraphs, mathematical objects relating diseases via links which can connect any number of diseases, thus capturing information about sets of diseases of any size. The cohort included 2,178,938 people. The most prevalent diseases were hypertension (13.3%), diabetes (6.9%), depression (6.7%) and chronic obstructive pulmonary disease (5.9%). The most important sets of diseases when considering prevalence generally contained a small number of diseases, while the most important sets of diseases when considering HRU were sets containing many diseases. The most important set of diseases taking prevalence and HRU into account was diabetes & hypertension and this combined measure of importance featured hypertension most often in the most important sets of diseases. We have used a single approach to find the most important sets of diseases based on co-occurrence and HRU measures, demonstrating the flexibility of the hypergraph approach. Hypertension, the most important single disease, is silent, underdiagnosed and increases the risk of life threatening co-morbidities. Co-occurrence of endocrine and cardiovascular diseases was common in the most important sets. Combining measures of prevalence with HRU provides insights which would be helpful for those planning and delivering services.

Objective Colorectal cancer is a common cause of death and morbidity. A significant amount of data are routinely collected during patient treatment, but they are not generally available for research. The National Institute for Health Research Health Informatics Collaborative in the UK is developing infrastructure to enable routinely collected data to be used for collaborative, cross-centre research. This paper presents an overview of the process for collating colorectal cancer data and explores the potential of using this data source.Methods Clinical data were collected from three pilot Trusts, standardised and collated. Not all data were collected in a readily extractable format for research. Natural language processing (NLP) was used to extract relevant information from pseudonymised imaging and histopathology reports. Combining data from many sources allowed reconstruction of longitudinal histories for each patient that could be presented graphically.Results Three pilot Trusts submitted data, covering 12 903 patients with a diagnosis of colorectal cancer since 2012, with NLP implemented for 4150 patients. Timelines showing individual patient longitudinal history can be grouped into common treatment patterns, visually presenting clusters and outliers for analysis. Difficulties and gaps in data sources have been identified and addressed.Discussion Algorithms for analysing routinely collected data from a wide range of sites and sources have been developed and refined to provide a rich data set that will be used to better understand the natural history, treatment variation and optimal management of colorectal cancer.Conclusion The data set has great potential to facilitate research into colorectal cancer.

BackgroundOnline consultations (OCs) allow patients to contact their care providers on the web. Worldwide, OCs have been rolled out in primary care rapidly owing to policy initiatives and COVID-19. There is a lack of evidence regarding how OC design and implementation influence care quality. ObjectiveWe aimed to synthesize research on the impacts of OCs on primary care quality, and how these are influenced by system design and implementation. MethodsWe searched databases from January 2010 to February 2022. We included quantitative and qualitative studies of real-world OC use in primary care. Quantitative data were transformed into qualitative themes. We used thematic synthesis informed by the Institute of Medicine domains of health care quality, and framework analysis informed by the nonadoption, abandonment, scale-up, spread, and sustainability framework. Strength of evidence was judged using the GRADE-CERQual approach. ResultsWe synthesized 63 studies from 9 countries covering 31 OC systems, 14 (22%) of which used artificial intelligence; 41% (26/63) of studies were published from 2020 onward, and 17% (11/63) were published after the COVID-19 pandemic. There was no quantitative evidence for negative impacts of OCs on patient safety, and qualitative studies suggested varied perceptions of their safety. Some participants believed OCs improved safety, particularly when patients could describe their queries using free text. Staff workload decreased when sufficient resources were allocated to implement OCs and patients used them for simple problems or could describe their queries using free text. Staff workload increased when OCs were not integrated with other software or organizational workflows and patients used them for complex queries. OC systems that required patients to describe their queries using multiple-choice questionnaires increased workload for patients and staff. Health costs decreased when patients used OCs for simple queries and increased when patients used them for complex queries. Patients using OCs were more likely to be female, younger, and native speakers, with higher socioeconomic status. OCs increased primary care access for patients with mental health conditions, verbal communication difficulties, and barriers to attending in-person appointments. Access also increased by providing a timely response to patients’ queries. Patient satisfaction increased when using OCs owing to better primary care access, although it decreased when using multiple-choice questionnaire formats. ConclusionsThis is the first theoretically informed synthesis of research on OCs in primary care and includes studies conducted during the COVID-19 pandemic. It contributes new knowledge that, in addition to having positive impacts on care quality such as increased access, OCs also have negative impacts such as increased workload. Negative impacts can be mitigated through appropriate OC system design (eg, free text format), incorporation of advanced technologies (eg, artificial intelligence), and integration into technical infrastructure (eg, software) and organizational workflows (eg, timely responses). Trial RegistrationPROSPERO CRD42020191802; https://tinyurl.com/2p84ezjy

Objectives Having multiple long term health conditions (MLTCs), also known as multimorbidity, is becoming increasingly common as populations age. Understanding how clusters of diseases are likely to lead to other diseases and the effect of multimorbidity on healthcare resource use (HRU) will be of great importance as this trend continues. Approach Graph-based approaches, also called network analysis in the literature, have been used previously to study multimorbidity. The use of hypergraphs, which are generalisations of graphs where edges can connect to any number of nodes, and their application to the problem of understanding multimorbidity will be discussed. Analysis using hypergraphs was carried out using a population-scale cohort of people in the Secure Anonymised Information Linkage (SAIL) Databank to find the diseases and disease sets which are most important based on a measure of prevalence and measures of healthcare resource utilisation in secondary care. Results The most important sets of diseases based on the centrality of a hypergraph weighted by a measure of prevalence featured hypertension, and the most important was hypertension and diabetes. The most important sets of diseases based on the centrality of a hypergraph weighted by a measure of unplanned inpatient HRU were arrhythmia, heart failure and hypertension while for a measure of outpatient HRU the most important set of diseases was diabetes and hypertension. Conclusion Hypergraphs are very flexible and general mathematical objects and there is still a great deal of development that can be done to make them more useful in epidemiological settings and beyond.

Objectives To identify predictors of overall lupus and lupus nephritis (LN) responses in patients with LN.Methods Data from the Aspreva Lupus Management Study (ALMS) trial cohort was used to identify baseline predictors of response at 6 months. Endpoints were major clinical response (MCR), improvement, complete renal response (CRR) and partial renal response (PRR). Univariate and multivariate logistic regressions with least absolute shrinkage and selection operator (LASSO) and cross-validation in randomly split samples were utilised. Predictors were ranked by the percentage of times selected by LASSO and prediction performance was assessed by the area under the receiver operating characteristics (AUROC) curve.Results We studied 370 patients in the ALMS induction trial. Improvement at 6 months was associated with older age (OR=1.03 (95% CI: 1.01 to 1.05) per year), normal haemoglobin (1.85 (1.16 to 2.95) vs low haemoglobin), active lupus (British Isles Lupus Assessment Group A or B) in haematological and mucocutaneous domains (0.61 (0.39 to 0.97) and 0.50 (0.31 to 0.81)), baseline damage (SDI>1 vs =0) (0.38 (0.16 to 0.91)) and 24-hour urine protein (0.63 (0.50 to 0.80)). LN duration 2–4 years (0.43 (0.19 to 0.97) vs

Abstract Background The methods with which prediction models are usually developed mean that neither the parameters nor the predictions should be interpreted causally. For many applications, this is perfectly acceptable. However, when prediction models are used to support decision making, there is often a need for predicting outcomes under hypothetical interventions. Aims We aimed to identify published methods for developing and validating prediction models that enable risk estimation of outcomes under hypothetical interventions, utilizing causal inference. We aimed to identify the main methodological approaches, their underlying assumptions, targeted estimands, and potential pitfalls and challenges with using the method. Finally, we aimed to highlight unresolved methodological challenges. Methods We systematically reviewed literature published by December 2019, considering papers in the health domain that used causal considerations to enable prediction models to be used for predictions under hypothetical interventions. We included both methodologies proposed in statistical/machine learning literature and methodologies used in applied studies. Results We identified 4919 papers through database searches and a further 115 papers through manual searches. Of these, 87 papers were retained for full-text screening, of which 13 were selected for inclusion. We found papers from both the statistical and the machine learning literature. Most of the identified methods for causal inference from observational data were based on marginal structural models and g-estimation. Conclusions There exist two broad methodological approaches for allowing prediction under hypothetical intervention into clinical prediction models: (1) enriching prediction models derived from observational studies with estimated causal effects from clinical trials and meta-analyses and (2) estimating prediction models and causal effects directly from observational data. These methods require extending to dynamic treatment regimes, and consideration of multiple interventions to operationalise a clinical decision support system. Techniques for validating ‘causal prediction models’ are still in their infancy.

Abstract Clinical prediction models (CPMs) have become fundamental for risk stratification across healthcare. The CPM pipeline (development, validation, deployment, and impact assessment) is commonly viewed as a one-time activity, with model updating rarely considered and done in a somewhat ad hoc manner. This fails to address the fact that the performance of a CPM worsens over time as natural changes in populations and care pathways occur. CPMs need constant surveillance to maintain adequate predictive performance. Rather than reactively updating a developed CPM once evidence of deteriorated performance accumulates, it is possible to proactively adapt CPMs whenever new data becomes available. Approaches for validation then need to be changed accordingly, making validation a continuous rather than a discrete effort. As such, “living” (dynamic) CPMs represent a paradigm shift, where the analytical methods dynamically generate updated versions of a model through time; one then needs to validate the system rather than each subsequent model revision.

Abstract Background Improving medication safety is a major concern in primary care settings worldwide. The Salford Medication safety dASHboard (SMASH) intervention provided general practices in Salford (Greater Manchester, UK) with feedback on their safe prescribing and monitoring of medications through an online dashboard, and input from practice-based trained clinical pharmacists. In this study we explored how staff working in general practices used the SMASH dashboard to improve medication safety, through interactions with the dashboard to identify potential medication safety hazards and their workflow to resolve identified hazards. Methods We used a mixed-methods study design involving quantitative data from dashboard user interaction logs from 43 general practices during the first year of receiving the SMASH intervention, and qualitative data from semi-structured interviews with 22 pharmacists and physicians from 18 practices in Salford. Results Practices interacted with the dashboard a median of 12.0 (interquartile range, 5.0–15.2) times per month during the first quarter of use to identify and resolve potential medication safety hazards, typically starting with the most prevalent hazards or those they perceived to be most serious. Having observed a potential hazard, pharmacists and practice staff worked together to resolve that in a sequence of steps (1) verifying the dashboard information, (2) reviewing the patient’s clinical records, and (3) deciding potential changes to the patient’s medicines. Over time, dashboard use transitioned towards regular but less frequent (median of 5.5 [3.5–7.9] times per month) checks to identify and resolve new cases. The frequency of dashboard use was higher in practices with a larger number of at-risk patients. In 24 (56%) practices only pharmacists used the dashboard; in 12 (28%) use by other practice staff increased as pharmacist use declined after the initial intervention period; and in 7 (16%) there was mixed use by both pharmacists and practice staff over time. Conclusions An online medication safety dashboard enabled pharmacists to identify patients at risk of potentially hazardous prescribing. They subsequently worked with GPs to resolve risks on a case-by-case basis, but there were marked variations in processes between some practices. Workload diminished over time as it shifted towards resolving new cases of hazardous prescribing.

Summary: Background: Surveillance of temporal trends in clinically treated self-harm is an important component of suicide prevention in the dynamic context of COVID-19. There is little evidence beyond the initial months following the onset of the pandemic, despite national and regional restrictions persisting to mid-2021. Methods: Descriptive time series analysis utilizing de-identified, primary care health records of 2.8 million patients from the Greater Manchester Care Record. Frequencies of self-harm episodes between 1st January 2019 and 31st May 2021 were examined, including stratification by sex, age group, ethnicity, and index of multiple deprivation quintile. Findings: There were 33,444 episodes of self-harm by 13,148 individuals recorded during the study period. Frequency ratios of incident and all episodes of self-harm were 0.59 (95% CI 0.51 to 0.69) and 0.69 (CI 0.63 to 0.75) respectively in April 2020 compared to February 2020. Between August 2020 and May 2021 frequency ratios were 0.92 (CI 0.88 to 0.96) for incident episodes and 0.86 (CI 0.84 to 0.88) for all episodes compared to the same months in 2019. Reductions were largest among men and people living in the most deprived neighbourhoods, while an increase in all-episode self-harm was observed for adolescents aged 10–17. Interpretation: Reductions in primary care-recorded self-harm persisted to May 2021, though they were less marked than in April 2020 during the first national lockdown. The observed reductions could represent longer term reluctance to seek help from health services. Our findings have implications for the ability for services to offer recommended care for patients who have harmed themselves.

Abstract Background Audit and feedback (A&F) is more effective if it facilitates action planning, but little is known about how best to do this. We developed an electronic A&F intervention with an action implementation toolbox to improve pain management in intensive care units (ICUs); the toolbox contained suggested actions for improvement. A head-to-head randomised trial demonstrated that the toolbox moderately increased the intervention’s effectiveness when compared with A&F only. Objective To understand the mechanisms through which A&F with action implementation toolbox facilitates action planning by ICUs to increase A&F effectiveness. Methods We extracted all individual actions from action plans developed by ICUs that received A&F with (n = 10) and without (n = 11) toolbox for 6 months and classified them using Clinical Performance Feedback Intervention Theory. We held semi-structured interviews with participants during the trial. We compared the number and type of planned and completed actions between study groups and explored barriers and facilitators to effective action planning. Results ICUs with toolbox planned more actions directly aimed at improving practice (p = 0.037) and targeted a wider range of practice determinants compared to ICUs without toolbox. ICUs with toolbox also completed more actions during the study period, but not significantly (p = 0.142). ICUs without toolbox reported more difficulties in identifying what actions they could take. Regardless of the toolbox, all ICUs still experienced barriers relating to the feedback (low controllability, accuracy) and organisational context (competing priorities, resources, cost). Conclusions The toolbox helped health professionals to broaden their mindset about actions they could take to change clinical practice. Without the toolbox, professionals tended to focus more on feedback verification and exploring solutions without developing intentions for actual change. All feedback recipients experienced organisational barriers that inhibited eventual completion of actions. Trial registration ClinicalTrials.gov, NCT02922101. Registered on 26 September 2016.

Abstract There are many acknowledged benefits for the reuse of clinical trial data; from independent verification of published results to the evaluation of new hypotheses. However, the reuse of shared clinical trial data is not without obstacles. Here we present some of the issues and lessons learned from our own experiences in accessing and analyzing trial data; specifically, where we aim to combine and pool data from multiple different trials. In addition to issues around missing annotation and incomplete datasets, we identify trial-design complexity as a potential hurdle that may complicate downstream analyses. We address potential solutions and emphasize the need for benefits of transparent sharing and analysis of participant-level clinical trial data with appropriate risk mitigation, a matter important to efficient clinical research.

Abstract Background The presence of additional chronic conditions has a significant impact on the treatment and management of type 2 diabetes (T2DM). Little is known about the patterns of comorbidities in this population. The aims of this study are to quantify comorbidity patterns in people with T2DM, to estimate the prevalence of six chronic conditions in 2027 and to identify clusters of similar conditions. Methods We used the Clinical Practice Research Datalink (CPRD) linked with the Index of Multiple Deprivation (IMD) data to identify patients diagnosed with T2DM between 2007 and 2017. 102,394 people met the study inclusion criteria. We calculated the crude and age-standardised prevalence of 18 chronic conditions present at and after the T2DM diagnosis. We analysed longitudinally the 6 most common conditions and forecasted their prevalence in 2027 using linear regression. We used agglomerative hierarchical clustering to identify comorbidity clusters. These analyses were repeated on subgroups stratified by gender and deprivation. Results More people living in the most deprived areas had ≥ 1 comorbidities present at the time of diagnosis (72% of females; 64% of males) compared to the most affluent areas (67% of females; 59% of males). Depression prevalence increased in all strata and was more common in the most deprived areas. Depression was predicted to affect 33% of females and 15% of males diagnosed with T2DM in 2027. Moderate clustering tendencies were observed, with concordant conditions grouped together and some variations between groups of different demographics. Conclusions Comorbidities are common in this population, and high between-patient variability in comorbidity patterns emphasises the need for patient-centred healthcare. Mental health is a growing concern, and there is a need for interventions that target both physical and mental health in this population.

Abstract Background Analysis of competing risks is commonly achieved through a cause specific or a subdistribution framework using Cox or Fine & Gray models, respectively. The estimation of treatment effects in observational data is prone to unmeasured confounding which causes bias. There has been limited research into such biases in a competing risks framework. Methods We designed simulations to examine bias in the estimated treatment effect under Cox and Fine & Gray models with unmeasured confounding present. We varied the strength of the unmeasured confounding (i.e. the unmeasured variable’s effect on the probability of treatment and both outcome events) in different scenarios. Results In both the Cox and Fine & Gray models, correlation between the unmeasured confounder and the probability of treatment created biases in the same direction (upward/downward) as the effect of the unmeasured confounder on the event-of-interest. The association between correlation and bias is reversed if the unmeasured confounder affects the competing event. These effects are reversed for the bias on the treatment effect of the competing event and are amplified when there are uneven treatment arms. Conclusion The effect of unmeasured confounding on an event-of-interest or a competing event should not be overlooked in observational studies as strong correlations can lead to bias in treatment effect estimates and therefore cause inaccurate results to lead to false conclusions. This is true for cause specific perspective, but moreso for a subdistribution perspective. This can have ramifications if real-world treatment decisions rely on conclusions from these biased results. Graphical visualisation to aid in understanding the systems involved and potential confounders/events leading to sensitivity analyses that assumes unmeasured confounders exists should be performed to assess the robustness of results.

Abstract Background Providing health professionals with quantitative summaries of their clinical performance when treating specific groups of patients (“feedback”) is a widely used quality improvement strategy, yet systematic reviews show it has varying success. Theory could help explain what factors influence feedback success, and guide approaches to enhance effectiveness. However, existing theories lack comprehensiveness and specificity to health care. To address this problem, we conducted the first systematic review and synthesis of qualitative evaluations of feedback interventions, using findings to develop a comprehensive new health care-specific feedback theory. Methods We searched MEDLINE, EMBASE, CINAHL, Web of Science, and Google Scholar from inception until 2016 inclusive. Data were synthesised by coding individual papers, building on pre-existing theories to formulate hypotheses, iteratively testing and improving hypotheses, assessing confidence in hypotheses using the GRADE-CERQual method, and summarising high-confidence hypotheses into a set of propositions. Results We synthesised 65 papers evaluating 73 feedback interventions from countries spanning five continents. From our synthesis we developed Clinical Performance Feedback Intervention Theory (CP-FIT), which builds on 30 pre-existing theories and has 42 high-confidence hypotheses. CP-FIT states that effective feedback works in a cycle of sequential processes; it becomes less effective if any individual process fails, thus halting progress round the cycle. Feedback’s success is influenced by several factors operating via a set of common explanatory mechanisms: the feedback method used, health professional receiving feedback, and context in which feedback takes place. CP-FIT summarises these effects in three propositions: (1) health care professionals and organisations have a finite capacity to engage with feedback, (2) these parties have strong beliefs regarding how patient care should be provided that influence their interactions with feedback, and (3) feedback that directly supports clinical behaviours is most effective. Conclusions This is the first qualitative meta-synthesis of feedback interventions, and the first comprehensive theory of feedback designed specifically for health care. Our findings contribute new knowledge about how feedback works and factors that influence its effectiveness. Internationally, practitioners, researchers, and policy-makers can use CP-FIT to design, implement, and evaluate feedback. Doing so could improve care for large numbers of patients, reduce opportunity costs, and improve returns on financial investments. Trial registration PROSPERO, CRD42015017541

Abstract Background Audit and feedback (A&F) is a common quality improvement strategy with highly variable effects on patient care. It is unclear how A&F effectiveness can be maximised. Since the core mechanism of action of A&F depends on drawing attention to a discrepancy between actual and desired performance, we aimed to understand current and best practices in the choice of performance comparator. Methods We described current choices for performance comparators by conducting a secondary review of randomised trials of A&F interventions and identifying the associated mechanisms that might have implications for effective A&F by reviewing theories and empirical studies from a recent qualitative evidence synthesis. Results We found across 146 trials that feedback recipients’ performance was most frequently compared against the performance of others (benchmarks; 60.3%). Other comparators included recipients’ own performance over time (trends; 9.6%) and target standards (explicit targets; 11.0%), and 13% of trials used a combination of these options. In studies featuring benchmarks, 42% compared against mean performance. Eight (5.5%) trials provided a rationale for using a specific comparator. We distilled mechanisms of each comparator from 12 behavioural theories, 5 randomised trials, and 42 qualitative A&F studies. Conclusion Clinical performance comparators in published literature were poorly informed by theory and did not explicitly account for mechanisms reported in qualitative studies. Based on our review, we argue that there is considerable opportunity to improve the design of performance comparators by (1) providing tailored comparisons rather than benchmarking everyone against the mean, (2) limiting the amount of comparators being displayed while providing more comparative information upon request to balance the feedback’s credibility and actionability, (3) providing performance trends but not trends alone, and (4) encouraging feedback recipients to set personal, explicit targets guided by relevant information.

Edge computing aims to integrate computing into everyday settings, enabling the system to be context-aware and private to the user. With the increasing success and popularity of deep learning methods, there is an increased demand to leverage these techniques in mobile and wearable computing scenarios. In this paper, we present an assessment of a deep human activity recognition system's memory and execution time requirements, when implemented on a mid-range smartphone class hardware and the memory implications for embedded hardware. This paper presents the design of a convolutional neural network (CNN) in the context of human activity recognition scenario. Here, layers of CNN automate the feature learning and the influence of various hyper-parameters such as the number of filters and filter size on the performance of CNN. The proposed CNN showed increased robustness with better capability of detecting activities with temporal dependence compared to models using statistical machine learning techniques. The model obtained an accuracy of 96.4% in a five-class static and dynamic activity recognition scenario. We calculated the proposed model memory consumption and execution time requirements needed for using it on a mid-range smartphone. Per-channel quantization of weights and per-layer quantization of activation to 8-bits of precision post-training produces classification accuracy within 2% of floating-point networks for dense, convolutional neural network architecture. Almost all the size and execution time reduction in the optimized model was achieved due to weight quantization. We achieved more than four times reduction in model size when optimized to 8-bit, which ensured a feasible model capable of fast on-device inference.

Abstract Background Disease populations, clinical practice, and healthcare systems are constantly evolving. This can result in clinical prediction models quickly becoming outdated and less accurate over time. A potential solution is to develop ‘dynamic’ prediction models capable of retaining accuracy by evolving over time in response to observed changes. Our aim was to review the literature in this area to understand the current state-of-the-art in dynamic prediction modelling and identify unresolved methodological challenges. Methods MEDLINE, Embase and Web of Science were searched for papers which used or developed dynamic clinical prediction models. Information was extracted on methods for model updating, choice of update windows and decay factors and validation of models. We also extracted reported limitations of methods and recommendations for future research. Results We identified eleven papers that discussed seven dynamic clinical prediction modelling methods which split into three categories. The first category uses frequentist methods to update models in discrete steps, the second uses Bayesian methods for continuous updating and the third, based on varying coefficients, explicitly describes the relationship between predictors and outcome variable as a function of calendar time. These methods have been applied to a limited number of healthcare problems, and few empirical comparisons between them have been made. Conclusion Dynamic prediction models are not well established but they overcome one of the major issues with static clinical prediction models, calibration drift. However, there are challenges in choosing decay factors and in dealing with sudden changes. The validation of dynamic prediction models is still largely unexplored terrain.

Objective Most current cardiovascular disease (CVD) risk stratification tools are for people without CVD, but very few are for prevalent CVD. In this study, we developed and validated a CVD severity score in people with coronary heart disease (CHD) and evaluated the association between severity and adverse outcomes.Methods Primary and secondary care data for 213 088 people with CHD in 398 practices in England between 2007 and 2017 were used. The cohort was randomly divided into training and validation datasets (80%/20%) for the severity model. Using 20 clinical severity indicators (each assigned a weight=1), baseline and longitudinal CVD severity scores were calculated as the sum of indicators. Adjusted Cox and competing-risk regression models were used to estimate risks for all-cause and cause-specific hospitalisation and mortality.Results Mean age was 64.5±12.7 years, 46% women, 16% from deprived areas, baseline severity score 1.5±1.2, with higher scores indicating a higher burden of disease. In the training dataset, 138 510 (81%) patients were hospitalised at least once, and 39 944 (23%) patients died. Each 1-unit increase in baseline severity was associated with 41% (95% CI 37% to 45%, area under the receiver operating characteristics (AUROC) curve=0.79) risk for 1 year for all-cause mortality; 59% (95% CI 52% to 67%, AUROC=0.80) for cardiovascular (CV)/diabetes mortality; 27% (95% CI 26% to 28%) for any-cause hospitalisation and 37% (95% CI 36% to 38%) for CV/diabetes hospitalisation. Findings were consistent in the validation dataset.Conclusions Higher CVD severity score is associated with higher risks for any-cause and cause-specific hospital admissions and mortality in people with CHD. Our reproducible score based on routinely collected data can help practitioners better prioritise management of people with CHD in primary care.

Introduction Multimorbidity is widely recognised as the presence of two or more concurrent long-term conditions, yet remains a poorly understood global issue despite increasing in prevalence.We have created the Wales Multimorbidity e-Cohort (WMC) to provide an accessible research ready data asset to further the understanding of multimorbidity. Our objectives are to create a platform to support research which would help to understand prevalence, trajectories and determinants in multimorbidity, characterise clusters that lead to highest burden on individuals and healthcare services, and evaluate and provide new multimorbidity phenotypes and algorithms to the National Health Service and research communities to support prevention, healthcare planning and the management of individuals with multimorbidity.Methods and analysis The WMC has been created and derived from multisourced demographic, administrative and electronic health record data relating to the Welsh population in the Secure Anonymised Information Linkage (SAIL) Databank. The WMC consists of 2.9 million people alive and living in Wales on the 1 January 2000 with follow-up until 31 December 2019, Welsh residency break or death. Published comorbidity indices and phenotype code lists will be used to measure and conceptualise multimorbidity.Study outcomes will include: (1) a description of multimorbidity using published data phenotype algorithms/ontologies, (2) investigation of the associations between baseline demographic factors and multimorbidity, (3) identification of temporal trajectories of clusters of conditions and multimorbidity and (4) investigation of multimorbidity clusters with poor outcomes such as mortality and high healthcare service utilisation.Ethics and dissemination The SAIL Databank independent Information Governance Review Panel has approved this study (SAIL Project: 0911). Study findings will be presented to policy groups, public meetings, national and international conferences, and published in peer-reviewed journals.

BackgroundMissing values are a key issue in the statistical analysis of proteomic data. Defining the strategy to address missing values is a complex task in each study, potentially affecting the quality of statistical analyses.ResultsWe have developed OptiMissP, a dashboard to visually and qualitatively evaluate missingness and guide decision making in the handling of missing values in proteomics studies that use data-independent acquisition mass spectrometry. It provides a set of visual tools to retrieve information about missingness through protein densities and topology-based approaches, and facilitates exploration of different imputation methods and missingness thresholds.ConclusionsOptiMissP provides support for researchers' and clinicians' qualitative assessment of missingness in proteomic datasets in order to define study-specific strategies for the handling of missing values. OptiMissP considers biases in protein distributions related to the choice of imputation method and helps analysts to balance the information loss caused by low missingness thresholds and the noise introduced by selecting high missingness thresholds. This is complemented by topological data analysis which provides additional insight to the structure of the data and their missingness. We use an example in Chronic Kidney Disease to illustrate the main functionalities of OptiMissP.

Abstract Background Significant evidence suggests that the cholesterol‐lowering statins can affect cognitive function and reduce the risk for Alzheimer's disease (AD) and dementia. These potential effects may be constrained by specific combinations of an individual's sex and apolipoprotein E (APOE) genotype. Methods Here we examine data from 252,327 UK Biobank participants, aged 55 or over, and compare the effects of statin use in males and females. We assessed difference in statin treatments taking a matched cohort approach, and identified key stratifiers using regression models and conditional inference trees. Using statistical modeling, we further evaluated the effect of statins on survival, cognitive decline over time, and on AD prevalence. Results We identified that in the selected population, males were older, had a higher level of education, better cognitive scores, higher incidence of cardiovascular and metabolic diseases, and a higher rate of statin use. We observed that males and those participants with an APOE ε4–positive genotype had higher probabilities of being treated with statins; while participants with an AD diagnosis had slightly lower probabilities. We found that use of statins was not significantly associated with overall higher rates of survival. However, when considering the interaction of statin use with sex, the results suggest higher survival rates in males treated with statins. Finally, examination of cognitive function indicates a potential beneficial effect of statins that is selective for APOE ε4–positive genotypes. Discussion Our evaluation of the aging population in a large cohort from the UK Biobank confirms sex and APOE genotype as fundamental risk stratifiers for AD and cognitive function, furthermore it extends them to the specific area of statin use, clarifying their specific interactions with treatments.

BackgroundWe evaluated the impact of the pharmacist-led Safety Medication dASHboard (SMASH) intervention on medication safety in primary care.Methods and findingsSMASH comprised (1) training of clinical pharmacists to deliver the intervention; (2) a web-based dashboard providing actionable, patient-level feedback; and (3) pharmacists reviewing individual at-risk patients, and initiating remedial actions or advising general practitioners on doing so. It was implemented in 43 general practices covering a population of 235,595 people in Salford (Greater Manchester), UK. All practices started receiving the intervention between 18 April 2016 and 26 September 2017. We used an interrupted time series analysis of rates (prevalence) of potentially hazardous prescribing and inadequate blood-test monitoring, comparing observed rates post-intervention to extrapolations from a 24-month pre-intervention trend. The number of people registered to participating practices and having 1 or more risk factors for being exposed to hazardous prescribing or inadequate blood-test monitoring at the start of the intervention was 47,413 (males: 23,073 [48.7%]; mean age: 60 years [standard deviation: 21]). At baseline, 95% of practices had rates of potentially hazardous prescribing (composite of 10 indicators) between 0.88% and 6.19%. The prevalence of potentially hazardous prescribing reduced by 27.9% (95% CI 20.3% to 36.8%, p < 0.001) at 24 weeks and by 40.7% (95% CI 29.1% to 54.2%, p < 0.001) at 12 months after introduction of SMASH. The rate of inadequate blood-test monitoring (composite of 2 indicators) reduced by 22.0% (95% CI 0.2% to 50.7%, p = 0.046) at 24 weeks; the change at 12 months (23.5%) was no longer significant (95% CI -4.5% to 61.6%, p = 0.127). After 12 months, 95% of practices had rates of potentially hazardous prescribing between 0.74% and 3.02%. Study limitations include the fact that practices were not randomised, and therefore unmeasured confounding may have influenced our findings.ConclusionsThe SMASH intervention was associated with reduced rates of potentially hazardous prescribing and inadequate blood-test monitoring in general practices. This reduction was sustained over 12 months after the start of the intervention for prescribing but not for monitoring of medication. There was a marked reduction in the variation in rates of hazardous prescribing between practices.

Background and purposeMachine learning (ML) has attracted much attention with the hope that it could make use of large, routinely collected datasets and deliver accurate personalised prognosis. The aim of this systematic review is to identify and critically appraise the reporting and developing of ML models for predicting outcomes after stroke.MethodsWe searched PubMed and Web of Science from 1990 to March 2019, using previously published search filters for stroke, ML, and prediction models. We focused on structured clinical data, excluding image and text analysis. This review was registered with PROSPERO (CRD42019127154).ResultsEighteen studies were eligible for inclusion. Most studies reported less than half of the terms in the reporting quality checklist. The most frequently predicted stroke outcomes were mortality (7 studies) and functional outcome (5 studies). The most commonly used ML methods were random forests (9 studies), support vector machines (8 studies), decision trees (6 studies), and neural networks (6 studies). The median sample size was 475 (range 70-3184), with a median of 22 predictors (range 4-152) considered. All studies evaluated discrimination with thirteen using area under the ROC curve whilst calibration was assessed in three. Two studies performed external validation. None described the final model sufficiently well to reproduce it.ConclusionsThe use of ML for predicting stroke outcomes is increasing. However, few met basic reporting standards for clinical prediction tools and none made their models available in a way which could be used or evaluated. Major improvements in ML study conduct and reporting are needed before it can meaningfully be considered for practice.

Background: Patient safety is vital to well-functioning health systems. A key component is safe prescribing, particularly in primary care where most medications are prescribed. Previous research demonstrated that the number of patients exposed to potentially hazardous prescribing can be reduced by interrogating the electronic health record (EHR) database of general practices and providing feedback to general practitioners in a pharmacist-led intervention. We aimed to develop and roll out an online dashboard application that delivers this audit and feedback intervention in a continuous fashion. Method: Based on initial system requirements we designed the dashboard’s user interface over 3 iterations with 6 general practitioners (GPs), 7 pharmacists and a member of the public. Prescribing safety indicators from previous work were implemented in the dashboard. Pharmacists were trained to use the intervention and deliver it to general practices. Results: A web-based electronic dashboard was developed and linked to shared care records in Salford, UK. The completed dashboard was deployed in all but one (n=43) general practices in the region. By November 2017, 36 pharmacists had been trained in delivering the intervention to practices. There were 135 registered users of the dashboard, with an average of 91 user sessions a week. Conclusion: We have developed and successfully rolled out of a complex, pharmacist-led dashboard intervention in Salford, UK. System usage statistics indicate broad and sustained uptake of the intervention. The use of systems that provide regularly updated audit information may be an important contributor towards medication safety in primary care

Abstract Background Audit and feedback aims to guide health professionals in improving aspects of their practice that need it most. Evidence suggests that feedback fails to increase accuracy of professional perceptions about clinical performance, which likely reduces audit and feedback effectiveness. This study investigates health professionals’ perceptions about their clinical performance and the influence of feedback on their intentions to change practice. Methods We conducted an online laboratory experiment guided by Control Theory with 72 intensive care professionals from 21 units. For each of four new pain management indicators, we collected professionals’ perceptions about their clinical performance; peer performance; targets; and improvement intentions before and after receiving first-time feedback. An electronic audit and feedback dashboard provided ICU’s own performance, median and top 10% peer performance, and improvement recommendations. The experiment took place approximately 1 month before units enrolled into a cluster-randomised trial assessing the impact of adding a toolbox with suggested actions and materials to improve intensive care pain management. During the experiment, the toolbox was inaccessible; all participants accessed the same version of the dashboard. Results We analysed 288 observations. In 53.8%, intensive care professionals overestimated their clinical performance; but in only 13.5%, they underestimated it. On average, performance was overestimated by 22.9% (on a 0–100% scale). Professionals similarly overestimated peer performance, and set targets 20.3% higher than the top performance benchmarks. In 68.4% of cases, intentions to improve practice were consistent with actual gaps in performance, even before professionals had received feedback; which increased to 79.9% after receiving feedback (odds ratio, 2.41; 95% CI, 1.53 to 3.78). However, in 56.3% of cases, professionals still wanted to improve care aspects at which they were already top performers. Alternatively, in 8.3% of cases, they lacked improvement intentions because they did not consider indicators important; did not trust the data; or deemed benchmarks unrealistic. Conclusions Audit and feedback helps health professionals to work on aspects for which improvement is recommended. Given the abundance of professionals’ prior good improvement intentions, the limited effects typically found by audit and feedback studies are likely predominantly caused by barriers to translation of intentions into actual change in clinical practice. Trial registration ClinicalTrials.gov NCT02922101. Registered 26 September 2016.

Abstract Background Patient portals are considered valuable instruments for self-management of long term conditions, however, there are concerns over how patients might interpret and act on the clinical information they access. We hypothesized that visual cues improve patients’ abilities to correctly interpret laboratory test results presented through patient portals. We also assessed, by applying eye-tracking methods, the relationship between risk interpretation and visual search behaviour. Methods We conducted a controlled study with 20 kidney transplant patients. Participants viewed three different graphical presentations in each of low, medium, and high risk clinical scenarios composed of results for 28 laboratory tests. After viewing each clinical scenario, patients were asked how they would have acted in real life if the results were their own, as a proxy of their risk interpretation. They could choose between: 1) Calling their doctor immediately (high interpreted risk); 2) Trying to arrange an appointment within the next 4 weeks (medium interpreted risk); 3) Waiting for the next appointment in 3 months (low interpreted risk). For each presentation, we assessed accuracy of patients’ risk interpretation, and employed eye tracking to assess and compare visual search behaviour. Results Misinterpretation of risk was common, with 65% of participants underestimating the need for action across all presentations at least once. Participants found it particularly difficult to interpret medium risk clinical scenarios. Participants who consistently understood when action was needed showed a higher visual search efficiency, suggesting a better strategy to cope with information overload that helped them to focus on the laboratory tests most relevant to their condition. Conclusions This study confirms patients’ difficulties in interpreting laboratories test results, with many patients underestimating the need for action, even when abnormal values were highlighted or grouped together. Our findings raise patient safety concerns and may limit the potential of patient portals to actively involve patients in their own healthcare.

Abstract Background Childhood pneumonia is the leading infectious cause of mortality in children younger than 5 years old. Recent updates to World Health Organization pneumonia guidelines recommend outpatient care for a population of children previously classified as high risk. This revision has been challenged by policymakers in Africa, where mortality related to pneumonia is higher than in other regions and often complicated by comorbidities. This study aimed to identify factors that best discriminate inpatient mortality risk in non-severe pneumonia and explore whether these factors offer any added benefit over the current criteria used to identify children with pneumonia requiring inpatient care. Methods We undertook a retrospective cohort study of children aged 2–59 months admitted with a clinical diagnosis of pneumonia at 14 public hospitals in Kenya between February 2014 and February 2016. Using machine learning techniques, we analysed whether clinical characteristics and common comorbidities increased the risk of inpatient mortality for non-severe pneumonia. The topmost risk factors were subjected to decision curve analysis to explore if using them as admission criteria had any net benefit above the current criteria. Results Out of 16,162 children admitted with pneumonia during the study period, 10,687 were eligible for subsequent analysis. Inpatient mortality within this non-severe group was 252/10,687 (2.36%). Models demonstrated moderately good performance; the partial least squares discriminant analysis model had higher sensitivity for predicting mortality in comparison to logistic regression. Elevated respiratory rate (≥70 bpm), age 2–11 months and weight-for-age Z-score (WAZ)

This article presents presentation and poster abstracts that were mistakenly omitted from the original publication.

Abstract Background Audit and feedback is a common intervention for supporting clinical behaviour change. Increasingly, health data are available in electronic format. Yet, little is known regarding if and how electronic audit and feedback (e-A&F) improves quality of care in practice. Objective The study aimed to assess the effectiveness of e-A&F interventions in a primary care and hospital context and to identify theoretical mechanisms of behaviour change underlying these interventions. Methods In August 2016, we searched five electronic databases, including MEDLINE and EMBASE via Ovid, and the Cochrane Central Register of Controlled Trials for published randomised controlled trials. We included studies that evaluated e-A&F interventions, defined as a summary of clinical performance delivered through an interactive computer interface to healthcare providers. Data on feedback characteristics, underlying theoretical domains, effect size and risk of bias were extracted by two independent review authors, who determined the domains within the Theoretical Domains Framework (TDF). We performed a meta-analysis of e-A&F effectiveness, and a narrative analysis of the nature and patterns of TDF domains and potential links with the intervention effect. Results We included seven studies comprising of 81,700 patients being cared for by 329 healthcare professionals/primary care facilities. Given the extremely high heterogeneity of the e-A&F interventions and five studies having a medium or high risk of bias, the average effect was deemed unreliable. Only two studies explicitly used theory to guide intervention design. The most frequent theoretical domains targeted by the e-A&F interventions included ‘knowledge’, ‘social influences’, ‘goals’ and ‘behaviour regulation‘, with each intervention targeting a combination of at least three. None of the interventions addressed the domains ‘social/professional role and identity’ or ‘emotion’. Analyses identified the number of different domains coded in control arm to have the biggest role in heterogeneity in e-A&F effect size. Conclusions Given the high heterogeneity of identified studies, the effects of e-A&F were found to be highly variable. Additionally, e-A&F interventions tend to implicitly target only a fraction of known theoretical domains, even after omitting domains presumed not to be linked to e-A&F. Also, little evaluation of comparative effectiveness across trial arms was conducted. Future research should seek to further unpack the theoretical domains essential for effective e-A&F in order to better support strategic individual and team goals.

Abstract Background Audit and feedback is often used as a strategy to improve quality of care, however, its effects are variable and often marginal. In order to learn how to design and deliver effective feedback, we need to understand their mechanisms of action. This theory-informed study will investigate how electronic audit and feedback affects improvement intentions (i.e. information–intention gap), and whether an action implementation toolbox with suggested actions and materials helps translating those intentions into action (i.e. intention–behaviour gap). The study will be executed in Dutch intensive care units (ICUs) and will be focused on pain management. Methods and design We will conduct a laboratory experiment with individual ICU professionals to assess the impact of feedback on their intentions to improve practice. Next, we will conduct a cluster randomised controlled trial with ICUs allocated to feedback without or feedback with action implementation toolbox group. Participants will not be told explicitly what aspect of the intervention is randomised; they will only be aware that there are two variations of providing feedback. ICUs are eligible for participation if they submit indicator data to the Dutch National Intensive Care Evaluation (NICE) quality registry and agree to allocate a quality improvement team that spends 4 h per month on the intervention. All participating ICUs will receive access to an online quality dashboard that provides two functionalities: gaining insight into clinical performance on pain management indicators and developing action plans. ICUs with access to the toolbox can develop their action plans guided by a list of potential barriers in the care process, associated suggested actions, and supporting materials to facilitate implementation of the actions. The primary outcome measure for the laboratory experiment is the proportion of improvement intentions set by participants that are consistent with recommendations based on peer comparisons; for the randomised trial it is the proportion of patient shifts during which pain has been adequately managed. We will also conduct a process evaluation to understand how the intervention is implemented and used in clinical practice, and how implementation and use affect the intervention’s impact. Discussion The results of this study will inform care providers and managers in ICU and other clinical settings how to use indicator-based performance feedback in conjunction with an action implementation toolbox to improve quality of care. Within the ICU context, this study will produce concrete and directly applicable knowledge with respect to what is or is not effective for improving pain management, and under which circumstances. The results will further guide future research that aims to understand the mechanisms behind audit and feedback and contribute to identifying the active ingredients of successful interventions. Trial registration ClinicalTrials.gov NCT02922101 . Registered 26 September 2016.

Introduction: The Informatics for Health congress, 24-26 April 2017, in Manchester, UK, brought together the Medical Informatics Europe (MIE) conference and the Farr Institute International Conference. This special issue of the Journal of Innovation in Health Informatics contains 113 presentation abstracts and 149 poster abstracts from the congress. Discussion: The twin programmes of “Big Data” and “Digital Health” are not always joined up by coherent policy and investment priorities. Substantial global investment in health IT and data science has led to sound progress but highly variable outcomes. Society needs an approach that brings together the science and the practice of health informatics. The goal is multi-level Learning Health Systems that consume and intelligently act upon both patient data and organizational intervention outcomes. Conclusions: Informatics for Health demonstrated the art of the possible, seen in the breadth and depth of our contributions. We call upon policy makers, research funders and programme leaders to learn from this joined-up approach.

Abstract Background Clinical prediction models (CPMs) are increasingly deployed to support healthcare decisions but they are derived inconsistently, in part due to limited data. An emerging alternative is to aggregate existing CPMs developed for similar settings and outcomes. This simulation study aimed to investigate the impact of between-population-heterogeneity and sample size on aggregating existing CPMs in a defined population, compared with developing a model de novo. Methods Simulations were designed to mimic a scenario in which multiple CPMs for a binary outcome had been derived in distinct, heterogeneous populations, with potentially different predictors available in each. We then generated a new ‘local’ population and compared the performance of CPMs developed for this population by aggregation, using stacked regression, principal component analysis or partial least squares, with redevelopment from scratch using backwards selection and penalised regression. Results While redevelopment approaches resulted in models that were miscalibrated for local datasets of less than 500 observations, model aggregation methods were well calibrated across all simulation scenarios. When the size of local data was less than 1000 observations and between-population-heterogeneity was small, aggregating existing CPMs gave better discrimination and had the lowest mean square error in the predicted risks compared with deriving a new model. Conversely, given greater than 1000 observations and significant between-population-heterogeneity, then redevelopment outperformed the aggregation approaches. In all other scenarios, both aggregation and de novo derivation resulted in similar predictive performance. Conclusion This study demonstrates a pragmatic approach to contextualising CPMs to defined populations. When aiming to develop models in defined populations, modellers should consider existing CPMs, with aggregation approaches being a suitable modelling strategy particularly with sparse data on the local population.

This paper examines the role that lived experience plays in the human capacity to reason about uncertainty. Previous research shows that people are more likely to provide accurate responses in Bayesian tasks when the data are presented in natural frequencies, the problem in question describes a familiar event, and the values of the data are in line with beliefs. Precisely why these factors are important remains open to debate. We elucidate the issue in two ways. Firstly, we hypothesize that in a task that requires people to reason about conditional probabilities, they are more likely to respond accurately when the values of the problem reflect their own lived experience, than when they reflect the experience of the average participant. Secondly, to gain further understanding of the underlying reasoning process, we employ a novel interaction analysis method that tracks mouse movements in an interactive web application and applies transition analysis to model how the approach to reasoning differs depending on whether data are presented using percentages or natural frequencies. We find (1) that the closer the values of the data in the problem are to people's self-reported lived experience, the more likely they are to provide a correct answer, and (2) that the reasoning process employed when data are presented using natural frequencies is qualitatively different to that employed when data are presented using percentages. The results indicate that the benefits of natural frequency presentation are due to a clearer representation of the relationship between sets and that the prior humans acquire through experience has an overwhelming influence on their ability to reason about uncertainty.

ObjectiveClinical code sets are vital to research using routinely-collected electronic healthcare data. Existing code set engineering methods pose significant limitations when considering reproducible research. To improve the transparency and reusability of research, these code sets must abide by FAIR principles; this is not currently happening. We propose 'term sets', an equivalent alternative to code sets that are findable, accessible, interoperable and reusable.Materials and methodsWe describe a new code set representation, consisting of natural language inclusion and exclusion terms (term sets), and explain its relationship to code sets. We formally prove that any code set has a corresponding term set. We demonstrate utility by searching for recently published code sets, representing them as term sets, and reporting on the number of inclusion and exclusion terms compared with the size of the code set.ResultsThirty-one code sets from 20 papers covering diverse disease domains were converted into term sets. The term sets were on average 74% the size of their equivalent original code set. Four term sets were larger due to deficiencies in the original code sets.DiscussionTerm sets can concisely represent any code set. This may reduce barriers for examining and reusing code sets, which may accelerate research using healthcare databases. We have developed open-source software that supports researchers using term sets.ConclusionTerm sets are independent of clinical code terminologies and therefore: enable reproducible research; are resistant to terminology changes; and are less error-prone as they are shorter than the equivalent code set.

The original article [1] contains an omitted grant acknowledgement and affiliation as relates to the contribution of co-author, Rafael Perera-Salazar. As such, the following two amendments should apply to the original article:

PURPOSE:To assess and improve the effectiveness of ICU care, in-hospital mortality rates are often used as principal quality indicator for benchmarking purposes. Two other often used, easily quantifiable, quality indicators to assess the efficiency of ICU care are based on readmission to the ICU and ICU length of stay. Our aim was to examine whether there is an association between case-mix adjusted outcome-based quality indicators in the general ICU population as well as within specific subgroups. MATERIALS AND METHODS:We included patients admitted in 2015 of all Dutch ICUs. We derived the standardized in-hospital mortality ratio (SMR); the standardized readmission ratio (SRR); and the standardized length of stay ratio (SLOSR). We expressed association through Pearson's correlation coefficients. RESULTS:The SMR ranged from 0.6 to 1.5; the SRR ranged from 0.7 to 2.1; and the SLOSR ranged from 0.7 to 1.3. For the total ICU population we found no significant associations. We found a positive, non-significant, association between SMR and SLOSR for admissions with low-mortality risk, (r = 0.25; p = 0.024), and a negative association between these indicators for admissions with high-mortality risk (r = -0.49; p

INTRODUCTION:Developments in information technology offer opportunities to enhance medication safety in primary care. We evaluated the implementation and adoption of a complex pharmacist-led intervention involving the use of an electronic audit and feedback surveillance dashboard to identify patients potentially at risk of hazardous prescribing or monitoring of medicines in general practices. The intervention aimed to create a rapid learning health system for medication safety in primary care. This study aimed to explore how the intervention was implemented, adopted and embedded into practice using a qualitative process evaluation. METHODS:Twenty two participants were purposively recruited from eighteen out of forty-three general practices receiving the intervention as well as clinical commissioning group staff across Salford UK, which reflected the range of contexts in which the intervention was implemented. Interviews explored how pharmacists and GP staff implemented the intervention and how this affected care practice. Data analysis was thematic with emerging themes developed into coding frameworks based on Normalisation Process Theory (NPT). RESULTS:Engagement with the dashboard involved a process of sense-making in which pharmacists considered it added value to their work. The intervention helped to build respect, improve trust and develop relationships between pharmacists and GPs. Collaboration and communication between pharmacists and clinicians was primarily initiated by pharmacists and was important for establishing the intervention. The intervention operated as a rapid learning health system as it allowed for the evidence in the dashboard to be translated into changes in work practices and into transformations in care. CONCLUSIONS:Our study highlighted the importance of the combined use of information technology and the role of pharmacists working in general practice settings. Medicine optimisation activities in primary care may be enhanced by the implementation of a pharmacist-led electronic audit and feedback system. This intervention established a rapid learning health system that swiftly translated data from electronic health records into changes in practice to improve patient care. Using NPT provided valuable insights into the ways in which developing relationships, collaborations and communication between health professionals could lead to the implementation, adoption and sustainability of the intervention.

ABSTRACT Objectives The secondary use of health data for public and private good raises complex questions of privacy and governance. Such questions are ill-suited to opinion polling where citizens must choose quickly with little information between multiple choice answers. This project aims to extend our knowledge about what control informed citizens would seek over health records after participating in a deliberative process. Approach Two citizens’ juries, of 17 citizens each, were convened; each jury was chosen to reflect national demographics, from a pool of almost 400 volunteers. Juries met separately over three days, to address the charge “To what extent should patients control access to patient records for secondary use?” Citizens heard from and questioned five expert witnesses (chosen either to inform the jury, or to air arguments for and against the secondary use of data), interspersed with opportunities to deliberate amongst themselves, using a range of tools including facilitated discussion and guided role-play. Jurors voted on a series of questions associated with the jury charge, giving their rationale. Individual views were polled at the beginning and end of the process against the jury charge. Results Thirty-three out of 34 jurors voted in support of the secondary use of data, with 24 wanting individuals to be able to opt-out and 6 favouring opt-in. Many jurors changed their opinion about who should get access to these records, with more people supporting wider information sharing by the end of day 3. When considering who should get access to data, the two juries had very similar rationales. Both thought that public benefit was a key justification for access. Jury 1 was more strongly supportive of sharing patient records for public benefit, whilst jury 2 was more cautious and sought to give patients more control. The questionnaire results suggest that over the course of three days a few people moved towards more patient control over patient records, but overall more people moved towards enabling greater sharing of health data for public benefit. Despite our endeavours to ensure otherwise, bias was reported by a few jurors, particularly regarding the impartiality of information from expert witnesses. Conclusion These findings illustrate that citizens often change their minds about complex policy questions when they become more informed about a public policy problem. It may also suggest that many, but not all, people become less sceptical about health data sharing as they become better informed of its benefits and risks.