How Will Machine Learning Inform the Clinical Care of Atrial Fibrillation?

Machine learning applications in cardiology have rapidly evolved in the past decade. With the availability of machine learning tools coupled with vast data sources, the management of atrial fibrillation (AF), a common chronic disease with significant associated morbidity and socioeconomic impact, is undergoing a knowledge and practice transformation in the increasingly complex healthcare environment. Among other advances, deep-learning machine learning methods, including convolutional neural networks, have enabled the development of AF screening pathways using the ubiquitous 12-lead ECG to detect asymptomatic paroxysmal AF in at-risk populations (such as those with cryptogenic stroke), the refinement of AF and stroke prediction schemes through comprehensive digital phenotyping using structured and unstructured data abstraction from the electronic health record or wearable monitoring technologies, and the optimization of treatment strategies, ranging from stroke prophylaxis to monitoring of antiarrhythmic drug (AAD) therapy. Although the clinical and population-wide impact of these tools continues to be elucidated, such transformative progress does not come without challenges, such as the concerns about adopting black box technologies, assessing input data quality for training such models, and the risk of perpetuating rather than alleviating health disparities. This review critically appraises the advances of machine learning related to the care of AF thus far, their potential future directions, and its potential limitations and challenges.

Evidence for models of diagnostic service provision in the community: literature mapping exercise and focused rapid reviews

Background

Current NHS policy favours the expansion of diagnostic testing services in community and primary care settings.

Objectives

Our objectives were to identify current models of community diagnostic services in the UK and internationally and to assess the evidence for quality, safety and clinical effectiveness of such services. We were also interested in whether or not there is any evidence to support a broader range of diagnostic tests being provided in the community.

Review methods

We performed an initial broad literature mapping exercise to assess the quantity and nature of the published research evidence. The results were used to inform selection of three areas for investigation in more detail. We chose to perform focused reviews on logistics of diagnostic modalities in primary care (because the relevant issues differ widely between different types of test); diagnostic ultrasound (a key diagnostic technology affected by developments in equipment); and a diagnostic pathway (assessment of breathlessness) typically delivered wholly or partly in primary care/community settings. Databases and other sources searched, and search dates, were decided individually for each review. Quantitative and qualitative systematic reviews and primary studies of any design were eligible for inclusion.

Results

We identified seven main models of service that are delivered in primary care/community settings and in most cases with the possible involvement of community/primary care staff. Not all of these models are relevant to all types of diagnostic test. Overall, the evidence base for community- and primary care-based diagnostic services was limited, with very few controlled studies comparing different models of service. We found evidence from different settings that these services can reduce referrals to secondary care and allow more patients to be managed in primary care, but the quality of the research was generally poor. Evidence on the quality (including diagnostic accuracy and appropriateness of test ordering) and safety of such services was mixed.

Conclusions

In the absence of clear evidence of superior clinical effectiveness and cost-effectiveness, the expansion of community-based services appears to be driven by other factors. These include policies to encourage moving services out of hospitals; the promise of reduced waiting times for diagnosis; the availability of a wider range of suitable tests and/or cheaper, more user-friendly equipment; and the ability of commercial providers to bid for NHS contracts. However, service development also faces a number of barriers, including issues related to staffing, training, governance and quality control.

Limitations

We have not attempted to cover all types of diagnostic technology in equal depth. Time and staff resources constrained our ability to carry out review processes in duplicate. Research in this field is limited by the difficulty of obtaining, from publicly available sources, up-to-date information about what models of service are commissioned, where and from which providers.

Future work

There is a need for research to compare the outcomes of different service models using robust study designs. Comparisons of ‘true’ community-based services with secondary care-based open-access services and rapid access clinics would be particularly valuable. There are specific needs for economic evaluations and for studies that incorporate effects on the wider health system. There appears to be no easy way of identifying what services are being commissioned from whom and keeping up with local evaluations of new services, suggesting a need to improve the availability of information in this area.

Funding

The National Institute for Health Research Health Services and Delivery Research programme.

Automated detection of atrial fibrillation using R-R intervals and multivariate-based classification

Automated detection of AF from the electrocardiogram (ECG) still remains a challenge. In this study, we investigated two multivariate-based classification techniques, Random Forests (RF) and k-nearest neighbor (k-nn), for improved automated detection of AF from the ECG. We have compiled a new database from ECG data taken from existing sources. R-R intervals were then analyzed using four previously described R-R irregularity measurements: (1) the coefficient of sample entropy (CoSEn), (2) the coefficient of variance (CV), (3) root mean square of the successive differences (RMSSD), and (4) median absolute deviation (MAD). Using outputs from all four R-R irregularity measurements, RF and k-nn models were trained. RF classification improved AF detection over CoSEn with overall specificity of 80.1% vs. 98.3% and positive predictive value of 51.8% vs. 92.1% with a reduction in sensitivity, 97.6% vs. 92.8%. k-nn also improved specificity and PPV over CoSEn; however, the sensitivity of this approach was considerably reduced (68.0%).

Screening for atrial fibrillation in general practice: a national, cross-sectional study of an innovative technology

BACKGROUND

To test the use of three lead monitoring as a screening tool for atrial fibrillation (AF) in general practice. AF is responsible for up to a quarter of all strokes and is often asymptomatic until a stroke occurs.

METHODS

26 randomly selected general practices identified 80 randomly selected patients aged 70 or older from their database and excluded those known to have AF, those with clinical issues or who had not attended for three years. Up to 40 eligible patients/practice were invited to attend for screening. A 2min three-lead ECG was recorded and collected centrally for expert cardiology assessment. Risk factor data was gathered.

OUTCOMES

(i) point prevalence of AF, (ii) proportion of ECG tracings which were adequate for interpretation, (iii) uptake rate by patients and (iv) acceptability of the screening process to patients and staff (reported separately).

RESULTS

Of 1447 current patients, 1003 were eligible for inclusion, 639 (64%) agreed to take part in screening and 566 (56%) completed screening. The point prevalence rate for AF was 10.3%-2.1% new cases (12 of 566 who were screened) and 9.5% existing cases (137 of 1447 eligible patients). Only four of 570 (0.7%) screening visits did not record a usable ECG and 11 (2.6%) three lead ECGs required a clarifying 12 lead ECG.

CONCLUSIONS

Three lead screening for AF is feasible, effective and offers an alternative to pulse taking or 12 lead ECGs. The availability of this technology may facilitate more effective screening, leading to reduced stroke incidence.

Noninvasive cardiac event monitoring to detect atrial fibrillation after ischemic stroke: a randomized, controlled trial

BACKGROUND AND PURPOSE

Atrial fibrillation (AF) elevates risk of recurrent stroke but is incompletely identified by standard investigation after stroke, though detection rates correlate with monitoring duration. We hypothesized that 7 days of noninvasive cardiac-event monitoring early after stroke would accelerate detection of AF and thus uptake of effective therapy.

METHODS

We performed a pragmatic randomized trial with objective outcome assessment among patients presenting in sinus rhythm with no AF history, within 7 days of ischemic stroke symptom onset. Patients were randomized to standard practice investigations (SP) to detect AF, or SP plus additional monitoring (SP-AM). AM comprised 7 days of noninvasive cardiac-event monitoring reported by an accredited cardiac electrocardiology laboratory. Primary outcome was detection of AF at 14 days.

RESULTS

One-hundred patients were enrolled from 2 centers. Within 14 days of stroke, sustained paroxysms of AF were detected in 18% of patients undergoing SP-AM versus 2% undergoing SP (P<0.05). Paroxysms of any-duration were detected in 44% of patients undergoing SP-AM versus 4% undergoing SP (P<0.001). These differences persisted at 90 days. Anticoagulant therapy was commenced within 14 days in 16% of SP-AM patients versus none randomized to SP (P<0.01). This difference persisted to 90 days (22% versus 6%; P<0.05).

CONCLUSIONS

Routine noninvasive cardiac-event monitoring after acute stroke enhances detection of paroxysmal AF and early anticoagulation. Extended monitoring should be offered to all eligible patients soon after acute stroke. Guidelines on investigation for AF in stroke patients could be strengthened.

CLINICAL TRIAL REGISTRATION URL

http://www.controlled-trials.com/isrctn/. Unique identifier: ISRCTN97412358.

Effectiveness of systematic screening for the detection of atrial fibrillation

BACKGROUND

Atrial fibrillation (AF) is the most common arrhythmia in clinical practice and is a leading cause of morbidity and mortality. Screening for AF in asymptomatic patients has been proposed as a way of reducing the burden of the disease by detecting people who would benefit from prophylactic anticoagulation therapy prior to the onset of symptoms. However, for screening to be an effective intervention it must improve the detection of AF and provide benefit for those who are detected earlier as a result of screening.

OBJECTIVES

The primary objective of this review was to examine whether screening programmes increase the detection of new cases of AF compared to routine practice. The secondary objectives were to identify which combination of screening strategy and patient population is most effective, as well as assessing any safety issues associated with screening, its acceptability within the target population and the costs involved.

SEARCH METHODS

We searched the Cochrane Central Register of Controlled Trials (CENTRAL) on The Cochrane Library, MEDLINE (Ovid) and EMBASE (Ovid) up to March 2012. Other relevant research databases, trials registries and websites were searched up to June 2012. Reference lists of identified studies were also searched for potentially relevant studies and we contacted corresponding authors for information about additional published or unpublished studies that may be relevant. No language restrictions were applied.

SELECTION CRITERIA

Randomised controlled trials, controlled before and after studies and interrupted time series studies comparing screening for AF with routine practice in people aged 40 years and over were eligible. Two authors (PM, CT or MF) independently selected the trials for inclusion.

DATA COLLECTION AND ANALYSIS

Assessment of risk of bias and data extraction were performed independently by two authors (PM, CT). Odds ratios (OR) and 95% confidence intervals (CI) were used to present the results for the primary outcome, which is a dichotomous variable. Since only one included study was identified, no meta-analysis was performed.

MAIN RESULTS

One cluster randomised controlled trial met the inclusion criteria for this review. This study compared systematic screening (by invitation to have an electrocardiogram (ECG)) and opportunistic screening (pulse palpation during a general practitioner (GP) consultation for any reason followed by an ECG if pulse was irregular) to routine practice (normal case finding on the basis of clinical presentation) in people aged 65 years or older. The risk of bias in the included study was judged to be low.Both systematic and opportunistic screening of people over the age of 65 years are more effective than routine practice (OR 1.57, 95% CI 1.08 to 2.26 and OR 1.58, 95% CI 1.10 to 2.29, respectively). The number needed to screen in order to detect one additional case compared to routine practice was 172 (95% CI 94 to 927) for systematic screening and 167 (95% CI 92 to 806) for opportunistic screening. Both systematic and opportunistic screening were more effective in men (OR 2.68, 95% CI 1.51 to 4.76 and OR 2.33, 95% CI 1.29 to 4.19, respectively) than in women (OR 0.98, 95% CI 0.59 to 1.62 and OR 1.2, 95% CI 0.74 to 1.93, respectively). No data on the effectiveness of screening in different ethnic or socioeconomic groups were available. There were insufficient data to compare the effectiveness of screening programmes in different healthcare settings.Systematic screening was associated with a better overall uptake rate than opportunistic screening (53% versus 46%) except in the ≥ 75 years age group where uptake rates were similar (43% versus 42%). In both screening programmes men were more likely to participate than women (57% versus 50% in systematic screening, 49% versus 41% in opportunistic screening) and younger people (65 to 74 years) were more likely to participate than people aged 75 years and over (61% versus 43% systematic, 49% versus 42% opportunistic). No adverse events associated with screening were reported.The incremental cost per additional case detected by opportunistic screening was GBP 337, compared to GBP 1514 for systematic screening. All cost estimates were based on data from the single included trial, which was conducted in the UK between 2001 and 2003.

AUTHORS' CONCLUSIONS

Systematic and opportunistic screening for AF increase the rate of detection of new cases compared with routine practice. While both approaches have a comparable effect on the overall AF diagnosis rate, the cost of systematic screening is significantly more than that of opportunistic screening from the perspective of the health service provider. The lack of studies investigating the effect of screening in other health systems and younger age groups means that caution needs to be exercised in relation to the transferability of these results beyond the setting and population in which the included study was conducted.Additional research is needed to examine the effectiveness of alternative screening strategies and to investigate the effect of the intervention on the risk of stroke for screened versus non-screened populations.