Budget impact analysis of a machine learning algorithm to predict high risk of atrial fibrillation among primary care patients.

AIMS: We investigated whether the use of an atrial fibrillation (AF) risk prediction algorithm could improve AF detection compared with opportunistic screening in primary care and assessed the associated budget impact. METHODS AND RESULTS: Eligible patients were registered with a general practice in UK, aged 65 years or older in 2018/19, and had complete data for weight, height, body mass index, and systolic and diastolic blood pressure recorded within 1 year. Three screening scenarios were assessed: (i) opportunistic screening and diagnosis (standard care); (ii) standard care replaced by the use of the algorithm; and (iii) combined use of standard care and the algorithm. The analysis considered a 3-year time horizon, and the budget impact for the National Health Service (NHS) costs alone or with personal social services (PSS) costs. Scenario 1 would identify 79 410 new AF cases (detection gap reduced by 22%). Scenario 2 would identify 70 916 (gap reduced by 19%) and Scenario 3 would identify 99 267 new cases (gap reduction 27%). These rates translate into 2639 strokes being prevented in Scenario 1, 2357 in Scenario 2, and 3299 in Scenario 3. The 3-year NHS budget impact of Scenario 1 would be £45.3 million, £3.6 million (difference ‒92.0%) with Scenario 2, and £46.3 million (difference 2.2%) in Scenario 3, but for NHS plus PSS would be ‒£48.8 million, ‒£80.4 million (64.8%), and ‒£71.3 million (46.1%), respectively. CONCLUSION: Implementation of an AF risk prediction algorithm alongside standard opportunistic screening could close the AF detection gap and prevent strokes while substantially reducing NHS and PSS combined care costs.

Detecting undiagnosed atrial fibrillation in UK primary care: Validation of a machine learning prediction algorithm in a retrospective cohort study.

AIMS: To evaluate the ability of a machine learning algorithm to identify patients at high risk of atrial fibrillation in primary care. METHODS: A retrospective cohort study was undertaken using the DISCOVER registry to validate an algorithm developed using a Clinical Practice Research Datalink (CPRD) dataset. The validation dataset included primary care patients in London, England aged ≥30 years from 1 January 2006 to 31 December 2013, without a diagnosis of atrial fibrillation in the prior 5 years. Algorithm performance metrics were sensitivity, specificity, positive predictive value, negative predictive value (NPV) and number needed to screen (NNS). Subgroup analysis of patients aged ≥65 years was also performed. RESULTS: Of 2,542,732 patients in DISCOVER, the algorithm identified 604,135 patients suitable for risk assessment. Of these, 3.0% (17,880 patients) had a diagnosis of atrial fibrillation recorded before study end. The area under the curve of the receiver operating characteristic was 0.87, compared with 0.83 in algorithm development. The NNS was nine patients, matching the CPRD cohort. In patients aged ≥30 years, the algorithm correctly identified 99.1% of patients who did not have atrial fibrillation (NPV) and 75.0% of true atrial fibrillation cases (sensitivity). Among patients aged ≥65 years (n = 117,965), the NPV was 96.7% with 91.8% sensitivity. CONCLUSIONS: This atrial fibrillation risk prediction algorithm, based on machine learning methods, identified patients at highest risk of atrial fibrillation. It performed comparably in a large, real-world population-based cohort and the developmental registry cohort. If implemented in primary care, the algorithm could be an effective tool for narrowing the population who would benefit from atrial fibrillation screening in the United Kingdom.

Hospital admissions for stroke and bleeding in Hounslow following a quality improvement initiative.

OBJECTIVE: Atrial fibrillation (AF) is the most common arrhythmia. Undiagnosed and poorly managed AF increases risk of stroke. The Hounslow AF quality improvement (QI) initiative was associated with improved quality of care for patients with AF through increased detection of AF and appropriate anticoagulation. This study aimed to evaluate whether there has been a change in stroke and bleeding rates in the Hounslow population following the QI initiative. METHODS: Using hospital admissions data from January 2011 to August 2018, interrupted time series analysis was performed to investigate the changes in standardised rates of admission with stroke and bleeding, following the start of the QI initiative in October 2014. RESULTS: There was a 17% decrease in the rate of admission with stroke as primary diagnosis (incidence rate ratio (IRR) 0.83; 95% CI 0.712 to 0.963; p<0.014). There was an even larger yet not statistically significant decrease in admission with stroke as primary diagnosis and AF as secondary diagnosis (IRR 0.75; 95% CI 0.550 to 1.025; p<0.071). No significant changes were observed in bleeding admissions. For each outcome, an additional regression model including both the level change and an interaction term for slope change was created. In all cases, the slope change was small and not statistically significant. CONCLUSION: Reduction in stroke admissions may be associated with the AF QI initiative. However, the immediate level change and non-significant slope change suggests a lack of effect of the intervention over time and that the decrease observed may be attributable to other events.

"Bow-tie" optimal pathway discovery analysis of sepsis hospital admissions using the Hospital Episode Statistics database in England.

OBJECTIVE: The "Bow-tie" optimal pathway discovery analysis uses large clinical event datasets to map clinical pathways and to visualize risks (improvement opportunities) before, and outcomes after, a specific clinical event. This proof-of-concept study assesses the use of NHS Hospital Episode Statistics (HES) in England as a potential clinical event dataset for this pathway discovery analysis approach. MATERIALS AND METHODS: A metaheuristic optimization algorithm was used to perform the "bow-tie" analysis on HES event log data for sepsis (ICD-10 A40/A41) in 2016. Analysis of hospital episodes across inpatient and outpatient departments was performed for the period 730 days before and 365 days after the index sepsis hospitalization event. RESULTS: HES data captured a sepsis event for 76 523 individuals (>13 years), relating to 580 000 coded events (across 220 sepsis and non-sepsis event classes). The "bow-tie" analysis identified several diagnoses that most frequently preceded hospitalization for sepsis, in line with the expectation that sepsis most frequently occurs in vulnerable populations. A diagnosis of pneumonia (5 290 patients) and urinary tract infections (UTIs; 2 057 patients) most often preceded the sepsis event, with recurrent UTIs acting as a potential indicative risk factor for sepsis. DISCUSSION: This proof-of-concept study demonstrates that a "bow-tie" pathway discovery analysis of the HES database can be undertaken and provides clinical insights that, with further study, could help improve the identification and management of sepsis. The algorithm can now be more widely applied to HES data to undertake targeted clinical pathway analysis across multiple healthcare conditions.

Application of a machine learning algorithm for detection of atrial fibrillation in secondary care.

Atrial fibrillation (AF) is the most common sustained heart arrhythmia and significantly increases risk of stroke. Opportunistic AF testing in high-risk patients typically requires frequent electrocardiogram tests to capture the arrhythmia. Risk-prediction algorithms may help to more accurately identify people with undiagnosed AF and machine learning (ML) may aid in the diagnosis of AF. Here, we applied an AF-risk prediction algorithm to secondary care data linked to primary care data in the DISCOVER database in order to evaluate changes in model performance, and identify patients not previously detected in primary care. We identified an additional 5,444 patients who had an AF diagnosis only in secondary care during the data extraction period. 2,696 (49.5%) were accepted by the algorithm and the algorithm correctly assigned 2,637 (97.8%) patients to the AF cohort. Using a risk threshold of 7.4% in patients aged ≥ 30 years, algorithm sensitivity and specificity was 38% and 95%, respectively. Approximately 15% of AF patients assigned to the AF cohort by the algorithm had a secondary care diagnosis with no record of AF in primary care. These additional patients did not substantially alter algorithm performance. The additional detection of previously undiagnosed AF patients in secondary care highlights unexpected potential utility of this ML algorithm.