Decoding 2.3 million ECGs: interpretable deep learning for advancing cardiovascular diagnosis and mortality risk stratification

Aims

Electrocardiogram (ECG) is widely considered the primary test for evaluating cardiovascular diseases. However, the use of artificial intelligence (AI) to advance these medical practices and learn new clinical insights from ECGs remains largely unexplored. We hypothesize that AI models with a specific design can provide fine-grained interpretation of ECGs to advance cardiovascular diagnosis, stratify mortality risks, and identify new clinically useful information.

Methods and results

Utilizing a data set of 2 322 513 ECGs collected from 1 558 772 patients with 7 years follow-up, we developed a deep-learning model with state-of-the-art granularity for the interpretable diagnosis of cardiac abnormalities, gender identification, and hypertension screening solely from ECGs, which are then used to stratify the risk of mortality. The model achieved the area under the receiver operating characteristic curve (AUC) scores of 0.998 (95% confidence interval (CI), 0.995-0.999), 0.964 (95% CI, 0.963-0.965), and 0.839 (95% CI, 0.837-0.841) for the three diagnostic tasks separately. Using ECG-predicted results, we find high risks of mortality for subjects with sinus tachycardia (adjusted hazard ratio (HR) of 2.24, 1.96-2.57), and atrial fibrillation (adjusted HR of 2.22, 1.99-2.48). We further use salient morphologies produced by the deep-learning model to identify key ECG leads that achieved similar performance for the three diagnoses, and we find that the V1 ECG lead is important for hypertension screening and mortality risk stratification of hypertensive cohorts, with an AUC of 0.816 (0.814-0.818) and a univariate HR of 1.70 (1.61-1.79) for the two tasks separately.

Conclusion

Using ECGs alone, our developed model showed cardiologist-level accuracy in interpretable cardiac diagnosis and the advancement in mortality risk stratification. In addition, it demonstrated the potential to facilitate clinical knowledge discovery for gender and hypertension detection which are not readily available.

Assessing the value of incorporating a polygenic risk score with non-genetic factors for predicting breast cancer diagnosis in the UK Biobank

BACKGROUND

Previous studies have demonstrated that incorporating a polygenic risk score (PRS) to existing risk prediction models for breast cancer improves model fit, but to determine its clinical utility the impact on risk categorisation needs to be established. We add a PRS to two well-established models and quantify the difference in classification using the net reclassification improvement (NRI).

METHODS

We analysed data from 126,490 post-menopausal women of "White British" ancestry, aged 40-69 years at baseline from the UK Biobank prospective cohort. The breast cancer outcome was derived from linked registry data and hospital records. We combined a PRS for breast cancer with 10-year risk scores from the Tyrer-Cuzick and Gail models, and compared these to the risk scores from the models using phenotypic variables alone. We report metrics of discrimination and classification, and consider the importance of the risk threshold selected.

RESULTS

The Harrell's C statistic of the 10-year risk from the Tyrer-Cuzick and Gail models was 0.57 and 0.54, respectively, increasing to 0.67 when the PRS was included. Inclusion of the PRS gave a positive NRI for cases in both models (0.080 (95% confidence interval: 0.053, 0.104) and 0.051 (95% CI: 0.030, 0.073), respectively), with negligible impact on controls.

CONCLUSIONS

The addition of a PRS for breast cancer to the well-established Tyrer-Cuzick and Gail models provides a substantial improvement in the prediction accuracy and risk stratification.

IMPACT

These findings could have important implications for the ongoing discussion about the value of PRS in risk prediction models and screening.

Incorporating polygenic risk into the Leicester Risk Assessment score for 10-year risk prediction of type 2 diabetes

AIMS

We evaluated whether incorporating information on ethnic background and polygenic risk enhanced the Leicester Risk Assessment (LRA) score for predicting 10-year risk of type 2 diabetes.

METHODS

The sample included 202,529 UK Biobank participants aged 40-69 years. We computed the LRA score, and developed two new risk scores using training data (80% sample): LRArev, which incorporated additional information on ethnic background, and LRAprs, which incorporated polygenic risk for type 2 diabetes. We assessed discriminative and reclassification performance in a test set (20% sample). Type 2 diabetes was ascertained using primary care, hospital inpatient and death registry records.

RESULTS

Over 10 years, 7,476 participants developed type 2 diabetes. The Harrell's C indexes were 0.796 (95% Confidence Interval [CI] 0.785, 0.806), 0.802 (95% CI 0.792, 0.813), and 0.829 (95% CI 0.820, 0.839) for the LRA, LRArev and LRAprs scores, respectively. The LRAprs score significantly improved the overall reclassification compared to the LRA (net reclassification index [NRI] = 0.033, 95% CI 0.015, 0.049) and LRArev (NRI = 0.040, 95% CI 0.024, 0.055) scores.

CONCLUSIONS

Polygenic risk moderately improved the performance of the existing LRA score for 10-year risk prediction of type 2 diabetes.

AutoNet-Generated Deep Layer-Wise Convex Networks for ECG Classification

The design of neural networks typically involves trial-and-error, a time-consuming process for obtaining an optimal architecture, even for experienced researchers. Additionally, it is widely accepted that loss functions of deep neural networks are generally non-convex with respect to the parameters to be optimised. We propose the Layer-wise Convex Theorem to ensure that the loss is convex with respect to the parameters of a given layer, achieved by constraining each layer to be an overdetermined system of non-linear equations. Based on this theorem, we developed an end-to-end algorithm (the AutoNet) to automatically generate layer-wise convex networks (LCNs) for any given training set. We then demonstrate the performance of the AutoNet-generated LCNs (AutoNet-LCNs) compared to state-of-the-art models on three electrocardiogram (ECG) classification benchmark datasets, with further validation on two non-ECG benchmark datasets for more general tasks. The AutoNet-LCN was able to find networks customised for each dataset without manual fine-tuning under 2 GPU-hours, and the resulting networks outperformed the state-of-the-art models with fewer than 5% parameters on all the above five benchmark datasets. The efficiency and robustness of the AutoNet-LCN markedly reduce model discovery costs and enable efficient training of deep learning models in resource-constrained settings.

Assessing the importance of primary care diagnoses in the UK Biobank

The UK Biobank has made general practitioner (GP) data (censoring date 2016-2017) available for approximately 45% of the cohort, whilst hospital inpatient and death registry (referred to as "HES/Death") data are available cohort-wide through 2018-2022 depending on whether the data comes from England, Wales or Scotland. We assessed the importance of case ascertainment via different data sources in UKB for three diseases that are usually first diagnosed in primary care: Parkinson's disease (PD), type 2 diabetes (T2D), and all-cause dementia. Including GP data at least doubled the number of incident cases in the subset of the cohort with primary care data (e.g. from 619 to 1390 for dementia). Among the 786 dementia cases that were only captured in the GP data before the GP censoring date, only 421 (54%) were subsequently recorded in HES. Therefore, estimates of the absolute incidence or risk-stratified incidence are misleadingly low when based only on the HES/Death data. For incident cases present in both HES/Death and GP data during the full follow-up period (i.e. until the HES censoring date), the median time difference between an incident diagnosis of dementia being recorded in GP and HES/Death was 2.25 years (i.e. recorded 2.25 years earlier in the GP records). Similar lag periods were also observed for PD (median 2.31 years earlier) and T2D (median 2.82 years earlier). For participants with an incident GP diagnosis, only 65.6% of dementia cases, 69.0% of PD cases, and 58.5% of T2D cases had their diagnosis recorded in HES/Death within 7 years since GP diagnosis. The effect estimates (hazard ratios, HR) of established risk factors for the three health outcomes mostly remain in the same direction and with a similar strength of association when cases are ascertained either using HES only or further adding GP data. The confidence intervals of the HR became narrower when adding GP data, due to the increased statistical power from the additional cases. In conclusion, it is desirable to extend both the coverage and follow-up period of GP data to allow researchers to maximise case ascertainment of chronic health conditions in the UK.

A medical multimodal large language model for future pandemics

Deep neural networks have been integrated into the whole clinical decision procedure which can improve the efficiency of diagnosis and alleviate the heavy workload of physicians. Since most neural networks are supervised, their performance heavily depends on the volume and quality of available labels. However, few such labels exist for rare diseases (e.g., new pandemics). Here we report a medical multimodal large language model (Med-MLLM) for radiograph representation learning, which can learn broad medical knowledge (e.g., image understanding, text semantics, and clinical phenotypes) from unlabelled data. As a result, when encountering a rare disease, our Med-MLLM can be rapidly deployed and easily adapted to them with limited labels. Furthermore, our model supports medical data across visual modality (e.g., chest X-ray and CT) and textual modality (e.g., medical report and free-text clinical note); therefore, it can be used for clinical tasks that involve both visual and textual data. We demonstrate the effectiveness of our Med-MLLM by showing how it would perform using the COVID-19 pandemic "in replay". In the retrospective setting, we test the model on the early COVID-19 datasets; and in the prospective setting, we test the model on the new variant COVID-19-Omicron. The experiments are conducted on 1) three kinds of input data; 2) three kinds of downstream tasks, including disease reporting, diagnosis, and prognosis; 3) five COVID-19 datasets; and 4) three different languages, including English, Chinese, and Spanish. All experiments show that our model can make accurate and robust COVID-19 decision-support with little labelled data.

Independent external validation of the QRISK3 cardiovascular disease risk prediction model using UK Biobank

OBJECTIVE

To externally evaluate the performance of QRISK3 for predicting 10 year risk of cardiovascular disease (CVD) in the UK Biobank cohort.

METHODS

We used data from the UK Biobank, a large-scale prospective cohort study of 403 370 participants aged 40-69 years recruited between 2006 and 2010 in the UK. We included participants with no previous history of CVD or statin treatment and defined the outcome to be the first occurrence of coronary heart disease, ischaemic stroke or transient ischaemic attack, derived from linked hospital inpatient records and death registrations.

RESULTS

Our study population included 233 233 women and 170 137 men, with 9295 and 13 028 incident CVD events, respectively. Overall, QRISK3 had moderate discrimination for UK Biobank participants (Harrell's C-statistic 0.722 in women and 0.697 in men) and discrimination declined by age (<0.62 in all participants aged 65 years or older). QRISK3 systematically overpredicted CVD risk in UK Biobank, particularly in older participants, by as much as 20%.

CONCLUSIONS

QRISK3 had moderate overall discrimination in UK Biobank, which was best in younger participants. The observed CVD risk for UK Biobank participants was lower than that predicted by QRISK3, particularly for older participants. It may be necessary to recalibrate QRISK3 or use an alternate model in studies that require accurate CVD risk prediction in UK Biobank.

Kidney function, albuminuria, and their modification by genetic factors and risk of incident dementia in UK Biobank

BACKGROUND

Associations between kidney function and dementia risk are inconclusive. Chronic kidney disease (CKD) severity is determined by levels of both estimated glomerular filtration rate (eGFR) and the urine albumin to creatinine ratio (ACR). However, whether there is a graded increase in dementia risk for worse eGFR in each ACR category is unclear. Also, whether genetic risk for dementia impacts the associations is unknown. The current study aims to investigate the associations between eGFR and albuminuria with dementia risk both individually and jointly, whether the associations vary by different follow-up periods, and whether genetic factors modified the associations.

METHODS

In 202,702 participants aged ≥ 60 years from the UK Biobank, Cox proportional-hazards models were used to examine the associations between eGFR and urine albumin creatinine ratio (ACR) with risk of incident dementia. GFR was estimated based on serum creatinine, cystatin C, or both. The models were restricted to different follow-up periods (< 5 years, 5-10 years, and ≥ 10 years) to investigate potential reverse causation.

RESULTS

Over 15 years of follow-up, 6,042 participants developed dementia. Decreased kidney function (eGFR < 60 ml/min/1.73m2) was associated with an increased risk of dementia (Hazard Ratio [HR] = 1.42, 95% Confidence Interval [CI] 1.28-1.58), compared to normal kidney function (≥ 90 ml/min/1.73m2). The strength of the association remained consistent when the models were restricted to different periods of follow-up. The HRs for incident dementia were 1.16 (95% CI 1.07-1.26) and 2.24 (95% CI 1.79-2.80) for moderate (3-30 mg/mmol) and severely increased ACR (≥ 30 mg/mmol) compared to normal ACR (< 3 mg/mmol). Dose-response associations were observed when combining eGFR and ACR, with those in the severest eGFR and ACR group having the greatest risk of dementia (HR = 4.70, 95% CI 2.34-9.43). APOE status significantly modified the association (p = 0.04), with stronger associations observed among participants with a lower genetic risk of dementia. There was no evidence of an interaction between kidney function and non-APOE polygenic risk of dementia with dementia risk (p = 0.42).

CONCLUSIONS

Kidney dysfunction and albuminuria were individually and jointly associated with higher dementia risk. The associations were greater amongst participants with a lower genetic risk of dementia based on APOE, but not non-APOE polygenic risk.

Combining machine learning with Cox models to identify predictors for incident post-menopausal breast cancer in the UK Biobank

We aimed to identify potential novel predictors for breast cancer among post-menopausal women, with pre-specified interest in the role of polygenic risk scores (PRS) for risk prediction. We utilised an analysis pipeline where machine learning was used for feature selection, prior to risk prediction by classical statistical models. An "extreme gradient boosting" (XGBoost) machine with Shapley feature-importance measures were used for feature selection among [Formula: see text] 1.7 k features in 104,313 post-menopausal women from the UK Biobank. We constructed and compared the "augmented" Cox model (incorporating the two PRS, known and novel predictors) with a "baseline" Cox model (incorporating the two PRS and known predictors) for risk prediction. Both of the two PRS were significant in the augmented Cox model ([Formula: see text]). XGBoost identified 10 novel features, among which five showed significant associations with post-menopausal breast cancer: plasma urea (HR = 0.95, 95% CI 0.92-0.98, [Formula: see text]), plasma phosphate (HR = 0.68, 95% CI 0.53-0.88, [Formula: see text]), basal metabolic rate (HR = 1.17, 95% CI 1.11-1.24, [Formula: see text]), red blood cell count (HR = 1.21, 95% CI 1.08-1.35, [Formula: see text]), and creatinine in urine (HR = 1.05, 95% CI 1.01-1.09, [Formula: see text]). Risk discrimination was maintained in the augmented Cox model, yielding C-index 0.673 vs 0.667 (baseline Cox model) with the training data and 0.665 vs 0.664 with the test data. We identified blood/urine biomarkers as potential novel predictors for post-menopausal breast cancer. Our findings provide new insights to breast cancer risk. Future research should validate novel predictors, investigate using multiple PRS and more precise anthropometry measures for better breast cancer risk prediction.

Polygenic Risk of Prediabetes, Undiagnosed Diabetes, and Incident Type 2 Diabetes Stratified by Diabetes Risk Factors

Context

Early diagnosis of type 2 diabetes is crucial to reduce severe comorbidities and complications. Current screening recommendations for type 2 diabetes include traditional risk factors, primarily body mass index (BMI) and family history, however genetics also plays a key role in type 2 diabetes risk. It is important to understand whether genetic predisposition to type 2 diabetes modifies the effect of these traditional factors on type 2 diabetes risk.

Objective

This work aimed to investigate whether genetic risk of type 2 diabetes modifies associations between BMI and first-degree family history of diabetes with 1) prevalent prediabetes or undiagnosed diabetes; and 2) incident confirmed type 2 diabetes.

Methods

We included 431 658 individuals aged 40 to 69 years at baseline of multiethnic ancestry from the UK Biobank. We used a multiethnic polygenic risk score for type 2 diabetes (PRS_T2D) developed by Genomics PLC. Prediabetes or undiagnosed diabetes was defined as baseline glycated hemoglobin greater than or equal to 42 mmol/mol (6.0%), and incident type 2 diabetes was derived from medical records.

Results

At baseline, 43 472 participants had prediabetes or undiagnosed diabetes, and 17 259 developed type 2 diabetes over 15 years follow-up. Dose-response associations were observed for PRS_T2D with each outcome in each category of BMI or first-degree family history of diabetes. Those in the highest quintile of PRS_T2D with a normal BMI were at a similar risk as those in the middle quintile who were overweight. Participants who were in the highest quintile of PRS_T2D and did not have a first-degree family history of diabetes were at a similar risk as those with a family history who were in the middle category of PRS_T2D.

Conclusion

Genetic risk of type 2 diabetes remains strongly associated with risk of prediabetes, undiagnosed diabetes, and future type 2 diabetes within categories of nongenetic risk factors. This could have important implications for identifying individuals at risk of type 2 diabetes for prevention and early diagnosis programs.

Hypertension, a dementia polygenic risk score, APOE genotype, and incident dementia

INTRODUCTION

There is inconsistent evidence on whether genetic risk for dementia modifies the association between hypertension and dementia.

METHODS

In 198,965 dementia-free participants aged ≥60 years, Cox proportional-hazards models were used to investigate the association between hypertension and incident dementia. A polygenic risk score (PRS) based on 38 non-apolipoprotein E (APOE) single nucleotide polymorphisms and APOE ε4 status were used to determine genetic risk for dementia.

RESULTS

Over 15 years follow-up, 6270 participants developed dementia. Hypertension was associated with a 19% increased risk of dementia (hazard ratio = 1.19, 95% confidence interval 1.11-1.27). The associations remained similar when stratifying by genetic risk, with no evidence for multiplicative interaction by dementia PRS (P = 0.20) or APOE ε4 status (P = 0.16). However, the risk difference between those with and without hypertension was larger among those at higher genetic risk.

DISCUSSION

Hypertension was associated with an increased risk of dementia regardless of genetic risk for dementia.

Incidence of fractures in people with intellectual disabilities over the life course: a retrospective matched cohort study

BACKGROUND

Current osteoporosis guidelines do not identify individuals with intellectual disabilities (ID) as at risk of fracture, potentially missing opportunities for prevention. We aimed to assess the incidence of fractures in people with ID over the life course.

METHODS

Descriptive analysis of open cohort study using anonymised electronic health records from the UK Clinical Practice Research Datalink, linked to the Hospital Episode Statistics database (Jan 1, 1998-Dec 31, 2017). All individuals with ID were matched on age and sex to five individuals without ID.  We calculated the incidence rate (95% CI) per 10000 person-years (py) and incidence rate ratio (IRR, 95% CI) to compare fractures between individuals with and without ID (age 1-17 and ≥18 years) for any fracture, and in those aged 18-49 and ≥ 50 years for major osteoporotic fracture (vertebra, shoulder, wrist, hip), and for hip fracture.

FINDINGS

43176 individuals with ID (15470 children aged 1-17 years; 27706 adults aged ≥ 18 years) were identified and included (40.4% females) along with 215733 matched control individuals.  The median age at study entry was 24 (10th-90th centiles 3-54) years. Over a median (10th-90th centile) follow-up of 7.1 (0.9-17.6) and 6.5 (0.8-17.6) years, there were 5941 and 24363 incident fractures in the ID and non ID groups respectively. Incidence of any fracture was 143.5 (131.8-156.3) vs 120.7 (115.4-126.4)/10000 py (children), 174.2 (166.4-182.4)/10000 py vs 118.2 (115.3-121.2)/10000 py (adults) in females. In males it was 192.5 (182.4-203.2) vs 228.5 (223.0-234.1)/10000 py (children), 155.6 (149.3-162.1)/10000 py vs 128.4 (125.9-131.0)/10000 py (adults). IRR for major osteoporotic fracture was 1.81 (1.50-2.18) age 18-49 years, 1.69 (1.53-1.87) age ≥ 50 years in women. In men it was 1.56 (1.36-1.79) age 18-49 years, 2.45 (2.13-2.81) age ≥ 50 years. IRR for hip fracture was 7.79 (4.14-14.65) age 18-49 years, 2.28 (1.91-2.71) age ≥ 50 years in women. In men it was 6.04 (4.18-8.73) age 18-49 years, 3.91 (3.17-4.82) age ≥ 50 years. Comparable rates of major osteoporotic fracture and of hip fracture occurred approximately 15 and 20 years earlier respectively in women and 20 and 30 years earlier respectively in men with ID than without ID. Fracture distribution differed profoundly, hip fracture 9.9% vs 5.0% of any fracture in adults with ID vs without ID.

INTERPRETATION

The incidence, type, and distribution of fractures in people with intellectual disabilities suggest early onset osteoporosis. Prevention and management strategies are urgently required, particularly to reduce the incidence of hip fracture.

FUNDING

National Institute for Health and Care Research.

Assessing agreement between different polygenic risk scores in the UK Biobank

Polygenic risk scores (PRS) are proposed for use in clinical and research settings for risk stratification. However, there are limited investigations on how different PRS diverge from each other in risk prediction of individuals. We compared two recently published PRS for each of three conditions, breast cancer, hypertension and dementia, to assess the stability of using these algorithms for risk prediction in a single large population. We used imputed genotyping data from the UK Biobank prospective cohort, limited to the White British subset. We found that: (1) 20% or more of SNPs in the first PRS were not represented in the more recent PRS for all three diseases, by the same SNP or a surrogate with R² > 0.8 by linkage disequilibrium (LD). (2) Although the difference in the area under the receiver operating characteristic curve (AUC) obtained using the two PRS is hardly appreciable for all three diseases, there were large differences in individual risk prediction between the two PRS. For instance, for each disease, of those classified in the top 5% of risk by the first PRS, over 60% were not so classified by the second PRS. We found substantial discordance between different PRS for the same disease, indicating that individuals could receive different medical advice depending on which PRS is used to assess their genetic susceptibility. It is desirable to resolve this uncertainty before using PRS for risk stratification in clinical settings.

Are polygenic risk scores for systolic blood pressure and LDL-cholesterol associated with treatment effectiveness, and clinical outcomes among those on treatment?

AIMS

Many studies have investigated associations between polygenic risk scores (PRS) and the incidence of cardiovascular disease (CVD); few have examined whether risk factor-related PRS predict CVD outcomes among adults treated with risk-modifying therapies. We assessed whether PRS for systolic blood pressure (PRSSBP) and for low-density lipoprotein cholesterol (PRSLDL-C) were associated with achieving SBP and LDL-C-related targets, and with major adverse cardiovascular events (MACE: non-fatal stroke or myocardial infarction, CVD death, and revascularization procedures).

METHODS AND RESULTS

Using observational data from the UK Biobank (UKB), we calculated PRSSBP and PRSLDL-C and constructed two sub-cohorts of unrelated adults of White British ancestry aged 40-69 years and with no history of CVD, who reported taking medications used in the treatment of hypertension or hypercholesterolaemia. Treatment effectiveness in achieving adequate risk factor control was ascertained using on-treatment blood pressure (BP) or LDL-C levels measured at enrolment (uncontrolled hypertension: BP ≥ 140/90 mmHg; uncontrolled hypercholesterolaemia: LDL-C ≥ 3 mmol/L). We conducted multivariable logistic and Cox regression modelling for incident events, adjusting for socioeconomic characteristics, and CVD risk factors. There were 55 439 participants using BP lowering therapies (51.0% male, mean age 61.0 years, median follow-up 11.5 years) and 33 787 using LDL-C lowering therapies (58.5% male, mean age 61.7 years, median follow-up 11.4 years). PRSSBP was associated with uncontrolled hypertension (odds ratio 1.70; 95% confidence interval: 1.60-1.80) top vs. bottom quintile, equivalent to a 5.4 mmHg difference in SBP, and with MACE [hazard ratio (HR) 1.13; 1.04-1.23]. PRSLDL-C was associated with uncontrolled hypercholesterolaemia (HR 2.78; 2.58-3.00) but was not associated with subsequent MACE.

CONCLUSION

We extend previous findings in the UKB cohort to examine PRSSBP and PRSLDL-C with treatment effectiveness. Our results indicate that both PRSSBP and PRSLDL-C can help identify individuals who, despite being on treatment, have inadequately controlled SBP and LDL-C, and for SBP are at higher risk for CVD events. This extends the potential role of PRS in clinical practice from identifying patients who may need these interventions to identifying patients who may need more intensive intervention.

Speech-in-noise hearing impairment is associated with an increased risk of incident dementia in 82,039 UK Biobank participants

INTRODUCTION

Little is known about the association between speech-in-noise (SiN) hearing impairment and dementia.

METHODS

In 82,039 dementia-free participants aged ≥60 years were selected from the UK Biobank. Cox proportional-hazards models were used to investigate whether SiN hearing impairment is associated with an increased risk of incident dementia.

RESULTS

Over 11 years of follow-up (median = 10.1), 1285 participants developed dementia. Insufficient and poor SiN hearing were associated with a 61% (hazard ratio [HR] = 1.61, 95% confidence [CI] 1.41-1.84) and 91% (HR = 1.91, 95% CI 1.55-2.36) increased risk of developing dementia, respectively, compared to normal SiN hearing. The association remained similar when restricting to follow-up intervals of ≤3, >3 to  <6, >6 to <9, and >9 years. There was limited evidence for mediation through depressive symptoms and social isolation.

DISCUSSION

SiN hearing impairment is independently associated with incident dementia, providing further evidence for hearing impairment as a potential modifiable dementia risk factor.

Calculating Polygenic Risk Scores (PRS) in UK Biobank: A Practical Guide for Epidemiologists

A polygenic risk score estimates the genetic risk of an individual for some disease or trait, calculated by aggregating the effect of many common variants associated with the condition. With the increasing availability of genetic data in large cohort studies such as the UK Biobank, inclusion of this genetic risk as a covariate in statistical analyses is becoming more widespread. Previously this required specialist knowledge, but as tooling and data availability have improved it has become more feasible for statisticians and epidemiologists to calculate existing scores themselves for use in analyses. While tutorial resources exist for conducting genome-wide association studies and generating of new polygenic risk scores, fewer guides exist for the simple calculation and application of existing genetic scores. This guide outlines the key steps of this process: selection of suitable polygenic risk scores from the literature, extraction of relevant genetic variants and verification of their quality, calculation of the risk score and key considerations of its inclusion in statistical models, using the UK Biobank imputed data as a model data set. Many of the techniques in this guide will generalize to other datasets, however we also focus on some of the specific techniques required for using data in the formats UK Biobank have selected. This includes some of the challenges faced when working with large numbers of variants, where the computation time required by some tools is impractical. While we have focused on only a couple of tools, which may not be the best ones for every given aspect of the process, one barrier to working with genetic data is the sheer volume of tools available, and the difficulty for a novice to assess their viability. By discussing in depth a couple of tools that are adequate for the calculation even at large scale, we hope to make polygenic risk scores more accessible to a wider range of researchers.

Prevalence and determinants of hypertension control among almost 100 000 treated adults in the UK

OBJECTIVE

To identify factors associated with hypertension control among treated middle-aged UK adults.

METHODS

A cross-sectional population-based study including 99 468 previously diagnosed, treated hypertensives enrolled in the UK Biobank. Hypertension control was defined as systolic blood pressure <140 mm Hg and diastolic blood pressure <90 mm Hg.

RESULTS

Median age was 62.3 years (IQR 57.3 to 66.0), 45.9% female, 92.0% white, 40.1% obese, 9.3% current smokers and 19.4% had prior cardiovascular disease. 38.1% (95% CI 37.8% to 38.4%) were controlled. In multivariable logistic regression, associations with lack of hypertension control included: older age (OR 0.61, 95% CI 0.58 to 0.64 for 60-69 years compared with age 40-50 years), higher alcohol use (OR 0.61, 95% CI 0.58 to 0.64, for consuming >30 units per week compared with none), black ethnicity (OR 0.73, 95% CI 0.65 to 0.82 compared with white), obesity (OR 0.73, 95% CI 0.71 to 0.76 compared with normal body mass index). The strongest positive association with control was having ≥3 comorbidities (OR 2.09, 95% CI 1.95 to 2.23). Comorbidities associated with control included cardiovascular disease (OR 2.11, 95% CI 2.04 to 2.19), migraines (OR 1.68, 95% CI 1.56 to 1.81), diabetes (OR 1.32, 95% CI 1.27 to 1.36) and depression (OR 1.27, 95% CI 1.20 to 1.34).

CONCLUSIONS

In one of the largest population-based analyses of middle-aged adults with measured blood pressure, the majority of treated hypertensives were uncontrolled. Risk factors for hypertension were associated with a lower probability of control. Having a comorbidity was associated with higher probability of control, possibly due to more frequent interaction with the healthcare system and/or appropriate management of those at greater cardiovascular risk.

Early warning score adjusted for age to predict the composite outcome of mortality, cardiac arrest or unplanned intensive care unit admission using observational vital-sign data: a multicentre development and validation

OBJECTIVES

Early warning scores (EWS) alerting for in-hospital deterioration are commonly developed using routinely collected vital-sign data from the whole in-hospital population. As these in-hospital populations are dominated by those over the age of 45 years, resultant scores may perform less well in younger age groups. We developed and validated an age-specific early warning score (ASEWS) derived from statistical distributions of vital signs.

DESIGN

Observational cohort study.

SETTING

Oxford University Hospitals (OUH) July 2013 to March 2018 and Portsmouth Hospitals (PH) NHS Trust January 2010 to March 2017 within the Hospital Alerting Via Electronic Noticeboard database.

PARTICIPANTS

Hospitalised patients with electronically documented vital-sign observations OUTCOME: Composite outcome of unplanned intensive care unit admission, mortality and cardiac arrest.

METHODS AND RESULTS

Statistical distributions of vital signs were used to develop an ASEWS to predict the composite outcome within 24 hours. The OUH development set consisted of 2 538 099 vital-sign observation sets from 142 806 admissions (mean age (SD): 59.8 (20.3)). We compared the performance of ASEWS to the National Early Warning Score (NEWS) and our previous EWS (MCEWS) on an OUH validation set consisting of 581 571 observation sets from 25 407 emergency admissions (mean age (SD): 63.0 (21.4)) and a PH validation set consisting of 5 865 997 observation sets from 233 632 emergency admissions (mean age (SD): 64.3 (21.1)). ASEWS performed better in the 16-45 years age group in the OUH validation set (AUROC 0.820 (95% CI 0.815 to 0.824)) and PH validation set (AUROC 0.840 (95% CI 0.839 to 0.841)) than NEWS (AUROC 0.763 (95% CI 0.758 to 0.768) and AUROC 0.836 (95% CI 0.835 to 0.838) respectively) and MCEWS (AUROC 0.808 (95% CI 0.803 to 0.812) and AUROC 0.833 (95% CI 0.831 to 0.834) respectively). Differences in performance were not consistent in the elder age group.

CONCLUSIONS

Accounting for age-related vital sign changes can more accurately detect deterioration in younger patients.

How to maintain the maximal level of blinding in randomisation for a placebo-controlled drug trial

We illustrate the approach of randomising treatments and compare it with the traditional approach of randomising patients, using a case study drawn from the authors’ experience in clinical trials. The setting is a double-blind parallel two-arm randomised controlled trial (RCT), but the method in this paper can be extended to single-blind, cross-over, or multi-arm RCTs.

We propose the concept of two different levels of blinding: full blinding and partial blinding. We subsequently show how to maintain the maximal level of blinding. Using an example, we show that a pharmacist can be fully blinded if the investigational medical products (IMPs) that they prescribe (instead of patients) are randomised, and they can be partially blinded if they need to dispense replacement (i.e., surplus) IMPs. A small number of surplus IMPs is commonly required in a clinical trial to replace lost or damaged IMPs. We note that the concept of full blinding and partial blinding is different from double-blind trial, and the level of blinding is relevant in both single-blind and double-blind trials.

A trial statistician needs to work closely with all parties in the design of the randomisation, including the pharmacist, the trial manager, and the manufacturer. We detail what should and should not be shown in the various documents that the trial statistician need to provide to the pharmacist and to the manufacturer. We provide template tables for these documents.

Comparing different ways of calculating sample size for two independent means: A worked example

We discuss different methods of sample size calculation for two independent means, aiming to provide insight into the calculation of sample size at the design stage of a parallel two-arm randomised controlled trial (RCT). We compare different methods for sample size calculation, using published results from a previous RCT. We use variances and correlation coefficients to compare sample sizes using different methods, including

1. The choice of the primary outcome measure: post-intervention score vs. change from baseline score.

2. The choice of statistical methods: t-test without using correlation coefficients vs. analysis of covariance (ANCOVA).

We show that the required sample size will depend on whether the outcome measure is the post-intervention score, or the change from baseline score, with or without baseline score included as a covariate. We show that certain assumptions have to be met when using simplified sample size equations, and discuss their implications in sample size calculation when planning an RCT. We strongly recommend publishing the crucial result “mean change (SE, standard error)” in a study paper, because it allows (i) the calculation of the variance of the change score in each arm, and (ii) to pool the variances from both arms. It also enables us to calculate the correlation coefficient in each arm. This subsequently allows us to calculate sample size using change score as the outcome measure. We use simulation to demonstrate how sample sizes by different methods are influenced by the strength of the correlation.

Comparing the Efficacy of a Mobile Phone-Based Blood Glucose Management System With Standard Clinic Care in Women With Gestational Diabetes: Randomized Controlled Trial

Background

Treatment of hyperglycemia in women with gestational diabetes mellitus (GDM) is associated with improved maternal and neonatal outcomes and requires intensive clinical input. This is currently achieved by hospital clinic attendance every 2 to 4 weeks with limited opportunity for intervention between these visits.

Objective

We conducted a randomized controlled trial to determine whether the use of a mobile phone-based real-time blood glucose management system to manage women with GDM remotely was as effective in controlling blood glucose as standard care through clinic attendance.

Methods

Women with an abnormal oral glucose tolerance test before 34 completed weeks of gestation were individually randomized to a mobile phone-based blood glucose management solution (GDm-health, the intervention) or routine clinic care. The primary outcome was change in mean blood glucose in each group from recruitment to delivery, calculated with adjustments made for number of blood glucose measurements, proportion of preprandial and postprandial readings, baseline characteristics, and length of time in the study.

Results

A total of 203 women were randomized. Blood glucose data were available for 98 intervention and 85 control women. There was no significant difference in rate of change of blood glucose (–0.16 mmol/L in the intervention and –0.14 mmol/L in the control group per 28 days, P=.78). Women using the intervention had higher satisfaction with care (P=.049). Preterm birth was less common in the intervention group (5/101, 5.0% vs 13/102, 12.7%; OR 0.36, 95% CI 0.12-1.01). There were fewer cesarean deliveries compared with vaginal deliveries in the intervention group (27/101, 26.7% vs 47/102, 46.1%, P=.005). Other glycemic, maternal, and neonatal outcomes were similar in both groups. The median time from recruitment to delivery was similar (intervention: 54 days; control: 49 days; P=.23). However, there were significantly more blood glucose readings in the intervention group (mean 3.80 [SD 1.80] and mean 2.63 [SD 1.71] readings per day in the intervention and control groups, respectively; P<.001). There was no significant difference in direct health care costs between the two groups, with a mean cost difference of the intervention group compared to control of –£1044 (95% CI –£2186 to £99). There were no unexpected adverse outcomes.

Conclusions

Remote blood glucocse monitoring in women with GDM is safe. We demonstrated superior data capture using GDm-health. Although glycemic control and maternal and neonatal outcomes were similar, women preferred this model of care. Further studies are required to explore whether digital health solutions can promote desired self-management lifestyle behaviors and dietetic adherence, and influence maternal and neonatal outcomes. Digital blood glucose monitoring may provide a scalable, practical method to address the growing burden of GDM around the world.

Trial Registration

ClinicalTrials.gov NCT01916694; https://clinicaltrials.gov/ct2/show/NCT01916694 (Archived by WebCite at http://www.webcitation.org/6y3lh2BOQ)

Feasibility of a staff training and support programme to improve pain assessment and management in people with dementia living in care homes

OBJECTIVES

The objective of this study was to establish the feasibility and initial effectiveness of training and support intervention for care staff to improve pain management in people with dementia living in care homes (PAIN-Dem).

METHODS

PAIN-Dem training was delivered to care staff from three care homes in South London, followed by intervention support and resources to encourage improved pain management by staff over 4 weeks. Feasibility was assessed through fidelity to intervention materials and qualitative approaches. Focus group discussions with staff explored the use of the PAIN-Dem intervention, and interviews were held with six residents and family carers. Pain was assessed in all residents at baseline, 3 and 4 weeks, and goal attainment scaling was assessed at 4 weeks.

RESULTS

Delivery of training was a key driver for success and feasibility of the PAIN-Dem intervention. Improvements in pain management behaviour and staff confidence were seen in homes where training was delivered in a care home setting across the care team with good manager buy-in. Family involvement in pain management was highlighted as an area for improvement. Goal attainment in residents was significantly improved across the cohort, although no significant change in pain was seen.

CONCLUSIONS

This study shows good initial feasibility of the PAIN-Dem intervention and provides valuable insight into training and support paradigms that deliver successful learning and behaviour change. There is a need for a larger trial of PAIN-Dem to establish its impact on resident pain and quantifiable staff behaviour measures. Copyright © 2017 John Wiley & Sons, Ltd.

Efficacy of sonographic and biological pleurodesis indicators of malignant pleural effusion (SIMPLE): protocol of a randomised controlled trial

Introduction

Malignant pleural effusion (MPE) is common and currently in UK there are an estimated 50 000 new cases of MPE per year. Talc pleurodesis remains one of the most popular methods for fluid control. The value of thoracic ultrasound (TUS) imaging, before and after pleurodesis, in improving the quality and efficacy of care for patients with MPE remains unknown. Additionally, biomarkers of successful pleurodesis including measurement of pleural fluid proteins have not been validated in prospective studies.The SIMPLE trial is an appropriately powered, multicentre, randomised controlled trial designed to assess 'by the patient bedside' use of TUS imaging and pleural fluid analysis in improving management of MPE.

Methods and analysis

262 participants with a confirmed MPE requiring intervention will be recruited from hospitals in UK and The Netherlands. Participants will be randomised (1:1) to undergo either chest drain insertion followed by instillation of sterile talc, or medical thoracoscopy and simultaneous poudrage. The allocated procedure will be done while the patient is hospitalised, and within 3 days of randomisation. Following hospital discharge, participants will be followed up at 1, 3 and 12 months. The primary outcome measure is the length of hospital stay during initial hospitalisation.

Ethics and dissemination

The trial has received ethical approval from the South Central-Oxford C Research Ethics Committee (Reference number 15/SC/0600). The Trial Steering Committee includes an independent chair and members, and a patient representative. The trial results will be published in a peer-reviewed journal and presented at international conferences.

Trial registration number

ISRCTN: 16441661.

Likelihood-based artefact detection in continuously-acquired patient vital signs

Robust continuous monitoring of patient vital signs (VS) is limited by artefactual data yielding measurements that are not representative of the patient's physiology. These artefacts are typified by several distinct "archetypes". We present several of these archetypal artefacts for heart rate (HR) monitoring, and propose a light weight, real-time algorithm to remove the majority of these artefacts. Most artefacts are not identifiable by their values in absolute terms, but instead by their values relative to other measurements nearby in time. We model temporally-proximate measurements as independent and identically distributed (i.i.d.) samples from a Gamma distribution. Measurements with low likelihood with respect to the distribution are candidates for artefact removal. This lightweight algorithm is important for real-time deployment on wearable sensors, which are becoming increasingly common in hospital and home care. The clinical applicability of artefact-removal is demonstrated in its ability to enhance patient deterioration detection. A Kalman filter-based patient monitoring algorithm is shown to improve early warning of deterioration when the proposed artefact-removal algorithm is used. We demonstrate this real-time system with patient data from a clinical trial that we have undertaken.

The landscape of pain management in people with dementia living in care homes: a mixed methods study

OBJECTIVES

The aim of this study is to explore the current landscape of pain management in people with dementia living in care home settings. Pain is extremely common in this patient group, yet there is very limited guidance for healthcare professionals.

METHODS

Triangulation of stakeholder consultation and quality review of pain management guidance were performed. A review of existing pain management guidance was conducted using published quality criteria adapted for the field. Three focus group discussions were held with care home staff and two focus group discussions and an online survey with family carers. Data were subjected to thematic analysis to identify themes and sub-themes. Outcomes were reviewed by an expert panel, which gave recommendations.

RESULTS

Fifteen existing guidelines were identified, of which three were designed for use in dementia and none were tailored for care home settings. Thematic analysis revealed six major themes in current pain management in dementia: importance of person-centredness, current lack of pain awareness in staff, communication as a core element, disparities in staff responsibility and confidence, the need for consistency of care and current lack of staff training. In addition to the needs for practice, the expert panel identified promising pharmacological treatment candidates, which warrant clinical evaluation.

CONCLUSIONS

The findings of this study clearly articulate a need for an evidence-based pain management programme for care homes, which is informed by stakeholder input and based within a conceptual framework for this setting. There are novel opportunities for clinical trials of alternative analgesics for use in this patient group. Copyright © 2016 John Wiley & Sons, Ltd.

Impact of a web-based tool (WebCONSORT) to improve the reporting of randomised trials: results of a randomised controlled trial

BACKGROUND

The CONSORT Statement is an evidence-informed guideline for reporting randomised controlled trials. A number of extensions have been developed that specify additional information to report for more complex trials. The aim of this study was to evaluate the impact of using a simple web-based tool (WebCONSORT, which incorporates a number of different CONSORT extensions) on the completeness of reporting of randomised trials published in biomedical publications.

METHODS

We conducted a parallel group randomised trial. Journals which endorsed the CONSORT Statement (i.e. referred to it in the Instruction to Authors) but do not actively implement it (i.e. require authors to submit a completed CONSORT checklist) were invited to participate. Authors of randomised trials were requested by the editor to use the web-based tool at the manuscript revision stage. Authors registering to use the tool were randomised (centralised computer generated) to WebCONSORT or control. In the WebCONSORT group, they had access to a tool allowing them to combine the different CONSORT extensions relevant to their trial and generate a customised checklist and flow diagram that they must submit to the editor. In the control group, authors had only access to a CONSORT flow diagram generator. Authors, journal editors, and outcome assessors were blinded to the allocation. The primary outcome was the proportion of CONSORT items (main and extensions) reported in each article post revision.

RESULTS

A total of 46 journals actively recruited authors into the trial (25 March 2013 to 22 September 2015); 324 author manuscripts were randomised (WebCONSORT n = 166; control n = 158), of which 197 were reports of randomised trials (n = 94; n = 103). Over a third (39%; n = 127) of registered manuscripts were excluded from the analysis, mainly because the reported study was not a randomised trial. Of those included in the analysis, the most common CONSORT extensions selected were non-pharmacologic (n = 43; n = 50), pragmatic (n = 20; n = 16) and cluster (n = 10; n = 9). In a quarter of manuscripts, authors either wrongly selected an extension or failed to select the right extension when registering their manuscript on the WebCONSORT study site. Overall, there was no important difference in the overall mean score between WebCONSORT (mean score 0.51) and control (0.47) in the proportion of CONSORT and CONSORT extension items reported pertaining to a given study (mean difference, 0.04; 95% CI -0.02 to 0.10).

CONCLUSIONS

This study failed to show a beneficial effect of a customised web-based CONSORT checklist to help authors prepare more complete trial reports. However, the exclusion of a large number of inappropriately registered manuscripts meant we had less precision than anticipated to detect a difference. Better education is needed, earlier in the publication process, for both authors and journal editorial staff on when and how to implement CONSORT and, in particular, CONSORT-related extensions.

TRIAL REGISTRATION

ClinicalTrials.gov: NCT01891448 [registered 24 May 2013].

Efficacy of cognitive behavioural therapy for sleep improvement in patients with persistent delusions and hallucinations (BEST): a prospective, assessor-blind, randomised controlled pilot trial

BACKGROUND

Sleep disturbance occurs in most patients with delusions or hallucinations and should be treated as a clinical problem in its own right. However, cognitive behavioural therapy (CBT)-the best evidence-based treatment for insomnia-has not been tested in this patient population. We aimed to pilot procedures for a randomised trial testing CBT for sleep problems in patients with current psychotic experiences, and to provide a preliminary assessment of potential benefit.

METHODS

We did this prospective, assessor-blind, randomised controlled pilot trial (Better Sleep Trial [BEST]) at two mental health centres in the UK. Patients (aged 18-65 years) with persistent distressing delusions or hallucinations in the context of insomnia and a schizophrenia spectrum diagnosis were randomly assigned (1:1), via a web-based randomisation system with minimisation to balance for sex, insomnia severity, and psychotic experiences, to receive either eight sessions of CBT plus standard care (medication and contact with the local clinical team) or standard care alone. Research assessors were masked to group allocation. Assessment of outcome was done at weeks 0, 12 (post-treatment), and 24 (follow-up). The primary efficacy outcomes were insomnia assessed by the Insomnia Severity Index (ISI) and delusions and hallucinations assessed by the Psychotic Symptoms Rating Scale (PSYRATS) at week 12. We did analysis by intention to treat, with an aim to provide confidence interval estimation of treatment effects. This study is registered with ISRCTN, number 33695128.

FINDINGS

Between Dec 14, 2012, and May 22, 2013, and Nov 7, 2013, and Aug 26, 2014, we randomly assigned 50 patients to receive CBT plus standard care (n=24) or standard care alone (n=26). The last assessments were completed on Feb 10, 2015. 48 (96%) patients provided follow-up data. 23 (96%) patients offered CBT took up the intervention. Compared with standard care, CBT led to reductions in insomnia in the large effect size range at week 12 (adjusted mean difference 6.1, 95% CI 3.0-9.2, effect size d=1.9). By week 12, nine (41%) of 22 patients receiving CBT and one (4%) of 25 patients receiving standard care alone no longer had insomnia, with ISI scores lower than the cutoff for insomnia. The treatment effect estimation for CBT covered a range from reducing but also increasing delusions (adjusted mean difference 0.3, 95% CI -2.0 to 2.6) and hallucinations (-1.9, -6.5 to 2.7). Three patients, all in the CBT group, had five adverse events, although none were regarded as related to study treatment.

INTERPRETATION

Our findings show that CBT for insomnia might be highly effective for improving sleep in patients with persistent delusions or hallucinations. A larger, suitably powered phase 3 study is now needed to provide a precise estimate of the effects of CBT for sleep problems, both on sleep and psychotic experiences.

FUNDING

Research for Patient Benefit Programme, National Institute for Health Research.

Multitask Gaussian processes for multivariate physiological time-series analysis

Gaussian process (GP) models are a flexible means of performing nonparametric Bayesian regression. However, GP models in healthcare are often only used to model a single univariate output time series, denoted as single-task GPs (STGP). Due to an increasing prevalence of sensors in healthcare settings, there is an urgent need for robust multivariate time-series tools. Here, we propose a method using multitask GPs (MTGPs) which can model multiple correlated multivariate physiological time series simultaneously. The flexible MTGP framework can learn the correlation between multiple signals even though they might be sampled at different frequencies and have training sets available for different intervals. Furthermore, prior knowledge of any relationship between the time series such as delays and temporal behavior can be easily integrated. A novel normalization is proposed to allow interpretation of the various hyperparameters used in the MTGP. We investigate MTGPs for physiological monitoring with synthetic data sets and two real-world problems from the field of patient monitoring and radiotherapy. The results are compared with standard Gaussian processes and other existing methods in the respective biomedical application areas. In both cases, we show that our framework learned the correlation between physiological time series efficiently, outperforming the existing state of the art.

'Errors' and omissions in paper-based early warning scores: the association with changes in vital signs--a database analysis

OBJECTIVES

To understand factors associated with errors using an established paper-based early warning score (EWS) system. We investigated the types of error, where they are most likely to occur, and whether 'errors' can predict subsequent changes in patient vital signs.

METHODS

Retrospective analysis of prospectively collected early warning system database from a single large UK teaching hospital.

RESULTS

16,795 observation sets, from 200 postsurgical patients, were collected. Incomplete observation sets were more likely to contain observations which should have led to an alert than complete observation sets (15.1% vs 7.6%, p<0.001), but less likely to have an alerting score correctly calculated (38.8% vs 30.0%, p<0.001). Mis-scoring was much more common when leaving a sequence of three or more consecutive observation sets with aggregate scores of 0 (55.3%) than within the sequence (3.0%, p<0.001). Observation sets that 'incorrectly' alerted were more frequently followed by a correctly alerting observation set than error-free non-alerting observation sets (14.7% vs 4.2%, p<0.001). Observation sets that 'incorrectly' did not alert were more frequently followed by an observation set that did not alert than error-free alerting observation sets (73.2% vs 45.8%, p<0.001).

CONCLUSIONS

Missed alerts are particularly common in incomplete observation sets and when a patient first becomes unstable. Observation sets that 'incorrectly' alert or 'incorrectly' do not alert are highly predictive of the next observation set, suggesting that clinical staff detect both deterioration and improvement in advance of the EWS system by using information not currently encoded within it. Work is urgently needed to understand how best to capture this information.

Probabilistic estimation of respiratory rate using Gaussian processes

The presence of respiratory information within the electrocardiogram (ECG) signal is a well-documented phenomenon. We present a Gaussian process framework for the estimation of respiratory rate from the different sources of modulation in a single-lead ECG. We propose a periodic covariance function to model the frequency- and amplitude-modulation time series derived from the ECG, where the hyperparameters of the process are used to derive the respiratory rate. The approach is evaluated using data taken from 40 healthy subjects each with 2 hours of monitoring, containing ECG and respiration waveforms. Results indicate that the accuracy of our proposed method is comparable with that of existing methods, but with the advantages of a principled probabilistic approach, including the direct quantification of the uncertainty in the estimation.

Predictive monitoring of mobile patients by combining clinical observations with data from wearable sensors

The majority of patients in the hospital are ambulatory and would benefit significantly from predictive and personalized monitoring systems. Such patients are well suited to having their physiological condition monitored using low-power, minimally intrusive wearable sensors. Despite data-collection systems now being manufactured commercially, allowing physiological data to be acquired from mobile patients, little work has been undertaken on the use of the resultant data in a principled manner for robust patient care, including predictive monitoring. Most current devices generate so many false-positive alerts that devices cannot be used for routine clinical practice. This paper explores principled machine learning approaches to interpreting large quantities of continuously acquired, multivariate physiological data, using wearable patient monitors, where the goal is to provide early warning of serious physiological determination, such that a degree of predictive care may be provided. We adopt a one-class support vector machine formulation, proposing a formulation for determining the free parameters of the model using partial area under the ROC curve, a method arising from the unique requirements of performing online analysis with data from patient-worn sensors. There are few clinical evaluations of machine learning techniques in the literature, so we present results from a study at the Oxford University Hospitals NHS Trust devised to investigate the large-scale clinical use of patient-worn sensors for predictive monitoring in a ward with a high incidence of patient mortality. We show that our system can combine routine manual observations made by clinical staff with the continuous data acquired from wearable sensors. Practical considerations and recommendations based on our experiences of this clinical study are discussed, in the context of a framework for personalized monitoring.

A large-scale clinical validation of an integrated monitoring system in the emergency department

We consider an integrated patient monitoring system, combining electronic patient records with high-rate acquisition of patient physiological data. There remain many challenges in increasing the robustness of "e-health" applications to a level at which they are clinically useful, particularly in the use of automated algorithms used to detect and cope with artifact in data contained within the electronic patient record, and in analyzing and communicating the resultant data for reporting to clinicians. There is a consequential "plague of pilots," in which engineering prototype systems do not enter into clinical use. This paper describes an approach in which, for the first time, the Emergency Department (ED) of a major research hospital has adopted such systems for use during a large clinical trial. We describe the disadvantages of existing evaluation metrics when applied to such large trials, and propose a solution suitable for large-scale validation. We demonstrate that machine learning technologies embedded within healthcare information systems can provide clinical benefit, with the potential to improve patient outcomes in the busy environment of a major ED and other high-dependence areas of patient care.

Gaussian processes for personalized e-health monitoring with wearable sensors

Advances in wearable sensing and communications infrastructure have allowed the widespread development of prototype medical devices for patient monitoring. However, such devices have not penetrated into clinical practice, primarily due to a lack of research into "intelligent" analysis methods that are sufficiently robust to support large-scale deployment. Existing systems are typically plagued by large false-alarm rates, and an inability to cope with sensor artifact in a principled manner. This paper has two aims: 1) proposal of a novel, patient-personalized system for analysis and inference in the presence of data uncertainty, typically caused by sensor artifact and data incompleteness; 2) demonstration of the method using a large-scale clinical study in which 200 patients have been monitored using the proposed system. This latter provides much-needed evidence that personalized e-health monitoring is feasible within an actual clinical environment, at scale, and that the method is capable of improving patient outcomes via personalized healthcare.

Extending the Generalised Pareto Distribution for Novelty Detection in High-Dimensional Spaces

Novelty detection involves the construction of a "model of normality", and then classifies test data as being either "normal" or "abnormal" with respect to that model. For this reason, it is often termed one-class classification. The approach is suitable for cases in which examples of "normal" behaviour are commonly available, but in which cases of "abnormal" data are comparatively rare. When performing novelty detection, we are typically most interested in the tails of the normal model, because it is in these tails that a decision boundary between "normal" and "abnormal" areas of data space usually lies. Extreme value statistics provides an appropriate theoretical framework for modelling the tails of univariate (or low-dimensional) distributions, using the generalised Pareto distribution (GPD), which can be demonstrated to be the limiting distribution for data occurring within the tails of most practically-encountered probability distributions. This paper provides an extension of the GPD, allowing the modelling of probability distributions of arbitrarily high dimension, such as occurs when using complex, multimodel, multivariate distributions for performing novelty detection in most real-life cases. We demonstrate our extension to the GPD using examples from patient physiological monitoring, in which we have acquired data from hospital patients in large clinical studies of high-acuity wards, and in which we wish to determine "abnormal" patient data, such that early warning of patient physiological deterioration may be provided.

Gaussian process regression in vital-sign early warning systems

The current standard of clinical practice for patient monitoring in most developed nations is connection of patients to vital-sign monitors, combined with frequent manual observation. In some nations, such as the UK, manual early warning score (EWS) systems have been mandated for use, in which scores are assigned to the manual observations, and care escalated if the scores exceed some pre-defined threshold. We argue that this manual system is far from ideal, and can be improved using machine learning techniques. We propose a system based on Gaussian process regression for improving the efficacy of existing EWS systems, and then demonstrate the method using manual observation of vital signs from a large-scale clinical study.

A Non-Invasive Method for Estimating Cardiopulmonary Variables Using Breath-by-Breath Injection of Two Tracer Gases

Conventional methods for estimating cardiopulmonary variables usually require complex gas analyzers and the active co-operation of the patient. Therefore, they are not compatible with the crowded environment of the intensive care unit (ICU) or operating theatre, where patient co-operation is typically impossible. However, it is these patients that would benefit the most from accurate estimation of cardiopulmonary variables, because of their critical condition. This paper describes the results of a collaborative development between an anesthesiologists and biomedical engineers to create a compact and non-invasive system for the measurement of cardiopulmonary variables such as lung volume, airway dead space volume, and pulmonary blood flow. In contrast with conventional methods, the compact apparatus and non-invasive nature of the proposed method allow it to be used in the ICU, as well as in general clinical settings. We propose the use of a non-invasive method, in which tracer gases are injected into the patient's inspired breath, and the concentration of the tracer gases is subsequently measured. A novel breath-by-breath tidal ventilation model is then used to estimate the value of a patient's cardiopulmonary variables. Experimental results from an artificial lung demonstrate minimal error in the estimation of known parameters using the proposed method. Results from analysis of a cohort of 20 healthy volunteers (within the Oxford University Hospitals NHS Trust) show that the values of estimated cardiopulmonary variables from these subjects lies within the expected ranges. Advantages of this method are that it is non-invasive, compact, portable, and can perform analysis in real time with less than 1 min of acquired respiratory data.

Assessment of lung function using a non-invasive oscillating gas-forcing technique

•

We propose a compact and non-invasive system for the measurement and monitoring of lung function.

•

We develop a novel tidal ventilation model using a non-invasive oscillating gas-forcing technique.

•

We compare a conventional continuous ventilation model with the proposed tidal ventilation model.

•

The proposed technique has several advantages over conventional methods, and can be used to assess patient lung function.

Conventional methods for monitoring lung function can require complex, or special, gas analysers, and may therefore not be practical in clinical areas such as the intensive care unit (ICU) or operating theatre. The system proposed in this article is a compact and non-invasive system for the measurement and monitoring of lung variables, such as alveolar volume, airway dead space, and pulmonary blood flow. In contrast with conventional methods, the compact apparatus and non-invasive nature of the proposed method could eventually allow it to be used in the ICU, as well as in general clinical settings. We also propose a novel tidal ventilation model using a non-invasive oscillating gas-forcing technique, where both nitrous oxide and oxygen are used as indicator gases. Experimental results are obtained from healthy volunteers, and are compared with those obtained using a conventional continuous ventilation model. Our findings show that the proposed technique can be used to assess lung function, and has several advantages over conventional methods such as compact and portable apparatus, easy usage, and quick estimation of cardiopulmonary variables.

A two-class approach to the detection of physiological deterioration in patient vital signs, with clinical label refinement

Hospital patient outcomes can be improved by the early identification of physiological deterioration. Automatic methods of detecting patient deterioration in vital-sign data typically attempt to identify deviations from assumed normal physiological conditions, which is a one-class approach to classification. This paper investigates the use of a two-class approach, in which abnormal physiology is modelled explicitly. The success of such a method relies on the accuracy of data labels provided by clinical experts, which may be incomplete (due to large dataset size) or imprecise (due to clinical labels covering intervals, rather than each data point within those intervals). We propose a novel method of refining clinical labels such that the two-class classification approach may be adopted for identifying patient deterioration. We demonstrate the effectiveness of the proposed methods using a large dataset acquired in a 24-bed hospital step-down unit.

Engineering preformed cobalt-doped platinum nanocatalysts for ultraselective hydrogenation

Bimetallic heterostructures are used as industrial catalysts for many important transformations. However, conventional catalysts are primarily prepared in cost-effective manners without much appreciation in metal size control and metal-metal interaction. By employing recent nanotechnology, Pt nanocrystals with tailored sizes can be decorated with Co atoms in a controlled manner in colloid solution as preformed nanocatalysts before they are applied on support materials. Thus, we show that the terminal CO hydrogenation can be achieved in high activity, while the undesirable hydrogenation of the CC group can be totally suppressed in the selective hydrogenation of alpha,beta-unsaturated aldehydes to unsaturated alcohols, when Co decorated Pt nanocrystals within a critical size range are used. This is achieved through blockage of unselective low coordination sites and the optimization in electronic influence of the Pt nanoparticle of appropriate size by the Co decoration. This work clearly demonstrates the advantage in engineering preformed nanoparticles via a bottom-up construction and illustrates that this route of catalyst design may lead to improved catalytic processes.

Treatment of neuroendocrine tumours with infusional 5-fluorouracil, folinic acid and streptozocin

A standardised, effective systemic therapy for metastatic neuroendocrine tumours (NETs) has not been established to date. We reviewed the management of 15 patients with inoperable, metastatic NET treated systematically with a combination chemotherapy regimen of infusional 5-fluorouracil, folinic acid and streptozocin. Overall objective response rate was 53% and tolerability was excellent.

The efficacy of caries detection using three intraoral films under different processing conditions

This study compares the diagnostic accuracy of caries detection using Ultra-speed (Eastman Kodak), Ektaspeed (Eastman Kodak) and Ektaspeed Plus (Eastman Kodak) films after they were developed in both new and used processing solutions. Ektaspeed Plus film provided significantly better diagnostic accuracy for small proximal-surface caries limited to the outer third of the dentin than did Ektaspeed film. Ektaspeed Plus film did not differ significantly from Ultra-speed film in aiding in the diagnosis of small carious lesions, and it maintained diagnostic accuracy in used processing solutions. Dentists can offer the X-ray dose-reducing technology of Ektaspeed Plus film to their patients and maintain consistently high diagnostic quality.

A canine model for determination of the therapeutic index of cytokine inhibitors

Using tumor necrosis factor (TNF) inhibition in dog blood as a measure of efficacy, and canine emesis as a measure of toxicity, we were able to assign a therapeutic index to rolipram, a prototypic anti-inflammatory compound. Because both assays were performed in the same species, the ambiguities associated with comparing the physiologic effects of drugs on various species was avoided. Rolipram, a standard phosphodiesterase type IV inhibitor, was a prototypic test compound characterized by a number of cardiovascular and central nervous system side effects, as well as its in vitro and in vivo inhibition of TNF. Initial experiments with canine whole blood incubated with lipopolysaccharide resulted in nanogram-per-milliliter concentrations of TNF that could be significantly reduced by in vitro addition of a 0.03 microM concentration of rolipram. Because rolipram inhibited canine TNF production in vitro, a protocol was devised in which TNF inhibitory activity was measured in a series of blood samples from dogs infused with increasingly high doses of rolipram. This yielded the efficacy half of the therapeutic index, whereas the emetogenic dose represented the side effect portion of the index. Rolipram was infused stepwise into conscious dogs at gradually increasing doses. The infusion was stopped when vomiting occurred, and the cumulative dose was reported as the emetic dose. Rolipram caused emesis in dogs at a cumulative dose of 0.1 mg/kg. At each dose of rolipram, blood was collected. The whole blood was incubated in vitro with lipopolysaccharide to induce TNF production, which in turn was quantified by the L929 bio-assay. Theoretically, if the rolipram infusion raised blood values high enough, the rolipram in whole blood would inhibit TNF production and be reflected by a lack of TNF activity in the L929 assay. In this assay system, rolipram's 50% effective dose in the TNF assay was always at least 33-fold lower than its emetic dose of 0.1 mg/kg. This gave rolipram a therapeutic index of at least 33:1 (0.003 versus 0.1 mg/kg) on the basis of its activity in a canine efficacy model (TNF inhibition) and a toxicity model (emesis induction). Experimental compounds were tested for their emetic dose as well as TNF 50% effective dose, with the goal of obtaining a therapeutic index better than that of rolipram. Thus the coupling of cytokine activity with overt toxicity was used to arrive at the therapeutic index of a compound. The therapeutic index was used to rank compounds as to their efficacy/toxicity profile. This ranking was used to eliminate several anti-inflammatory compounds that had a therapeutic index less than that of rolipram.