Budget impact analysis of a machine learning algorithm to predict high-risk of AF among primary care patients

Discover-NOW, UCL Partners, The Health Economics Unit and Bristol
Myers Squibb Pharmaceuticals Ltd investigated whether the use of an atrial fibrillation (AF) risk prediction algorithm could improve AF detection compared with opportunistic screening in primary care.

AF is the most common sustained heart arrhythmia. In 2019 in UK, nearly 1.5 million (2.5%) people were living with AF, but 20% of cases were thought to be undiagnosed AF. AF increases the risk of ischaemic stroke around five-fold and AF-associated strokes are more severe than those in patients without AF. Around 40–50% of costs associated with all strokes in the UK are direct NHS care costs, with the remainder being indirect personal social services (PSS) costs provided in the community (e.g. physiotherapy, occupational therapy, and psychological support).

However, systematic population screening for AF is not currently endorsed by the UK National Screening Committee, although they suggest combining routine clinical diagnosis and opportunistic screening in general practices.

Read the full publication


The collaborative 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.


Study design and population
The target population is represented by adults aged 65 years or older without a diagnosis of AF. The size of the population was calculated by applying data about the algorithm performance to national data for patients registered at a general practitioner in UK in 2018/19. As the data are publicly available aggregated data, neither written informed consent from patients nor ethics approval was required.

AF risk-prediction algorithm
The AF risk-prediction model generates a risk score for AF from baseline risk factors (e.g. age, previous cardiovascular disease, antihypertensive medication usage) and time-varying predictors [e.g. proximity of cardiovascular events, body mass index (BMI) (level and change), pulse pressure, and the frequency of blood pressure measurements].

The algorithm was developed using Clinical Practice Research Datalink (CPRD) data and validated using anonymized data for a retrospective cohort of approximately 2.5 million patients in the North West London Whole Systems Integrated Care data warehouse (WSIC).19 Data were obtained via the DISCOVER secure environment, in which only anonymized data are stored. In the CPRD development dataset, a threshold of 7.4% for the current high risk of AF was associated with 50% sensitivity. The risk threshold is predicated on a number of aforementioned intrinsic risk factors, such as age, BMI, and systolic hypertension.19 In the WSIC cohort, we found that among patients aged 65 years or older, sensitivity with this threshold was 51%. A confusion matrix of the AF risk-prediction algorithm in the validation cohort was created to confirm the performance of this risk threshold, and it was set as the threshold for this study.

As there is potential for the AF risk prediction algorithm to be practically implemented in multiple ways, the effects of three different implementation scenarios were assessed: Scenario 1 (base case), in which standard care of routine clinical diagnosis and opportunistic screening would continue; Scenario 2, in which the algorithm would replace standard care and run for all patients in primary care, with all flagged high-risk being invited to attend the practice for assessment; and Scenario 3, which would be a combination of Scenarios 1 and 2. Scenarios 2 and 3 were compared with Scenario 1 in all instances.

Budget impact analysis
A budget impact model was developed to assess the economic impact of introducing the AF risk prediction algorithm into the NHS from the perspective of NHS and PSS at the national level in UK. Values were derived from the WSIC validation cohort for patients aged 65 years and older who had complete data for weight, height, BMI, and systolic and diastolic blood pressure recorded within the 12-month period before their latest visit in 2018/19 (index date). Individuals with no valid index date or who had a code for AF recorded prior to their index date were excluded. The outputs were used to extrapolate national values for UK since the WSIC cohort is broadly representative of the UK population. Budget impact estimates were calculated for a time horizon of three consecutive years.


AF detection gap
The prevalence of AF in the WSIC cohort was 3.0% (17,800 of 604,135). We calculated that the national total AF prevalence in 2019 was 1,480,221 (2.5%), but the number of patients registered as having an AF diagnosis in 2019 was 1 174 959 (2.0%). Thus, we estimated an AF detection gap of 305 262, which equates to 20.6%

Clinical outcomes
In the year 2018/19, 9 495 830 patients (16.2%) of 58 735 241 registered with general practices in UK were aged 65 years or older and were eligible for screening by the AF risk prediction algorithm.

In Scenario 1, we estimated that 2 772 782 of 9 495 830 patients (29.2%) would receive opportunistic screening in primary care. In Scenario 2, only 376 984 (4.0%) patients at risk of AF would receive screening, reducing the rate by 84.4% compared with Scenario 1. In Scenario 3, the addition of the algorithm to routine screening and diagnosis would increase the number of patients who received screening 2 878 338 (30.3%).

Over the 3-year study time horizon, it was estimated that 79 410 new cases of AF would be identified in Scenario 1, closing the detection gap by 22% and preventing 2639 strokes. The numbers would be lower in Scenario 2 (70 916 AF cases detected, closing the detection gap by 19%, and 2357 strokes prevented). In contrast, Scenario 3 notably improves AF detection (99 267 new cases detected), closing the detection gap by 27% and preventing 3299 strokes (Figure 2). This pattern was similar for additional major bleeds (Figure 2).

Read the full publication