Real-world detection of atrial fibrillation after ischaemic stroke using the FIND-AFDAS meta-machine learning algorithm: an unmet need

Background

Ischaemic stroke patients have a high risk of stroke recurrence, which can be reduced by diagnosing and treating underlying atrial fibrillation (AF) (1). Prolonged ECG monitoring is often required to diagnose AF, but it is inequitably provided. In a recent multinational analysis, we have shown the FIND-AFDAS meta-machine learning algorithm can accurately identify ischaemic stroke patients at higher risk of AF diagnosis, with a positive predictive value of 83% at 12 months with implantable loop recorder (ILR) (2).

Purpose

To establish current use of prolonged ECG monitoring after ischaemic stroke for AF detection, establish the feasibility of implementing the FIND-AFDAS algorithm in routine care records, and identify the diagnostics gap for patients at high FIND-AFDAS risk.

Methods

We performed a retrospective analysis of adult patients (≥18 years) without prior atrial fibrillation admitted to a single regional stroke centre between 1st March 2023 and 31st May 2023 with ischaemic stroke. Individuals with AF on admission ECG or telemetry were excluded. Using the FIND-AFDAS algorithm, patients were stratified into low and high risk of AF detection (2). The primary outcome was the composite of all-cause mortality, recurrent stroke or AF at one year. Secondary outcomes were the components of the primary outcome. The Fisher’s Exact test, Chi-Square test, Mann-Whitney U test and Student’s t test were used to compare groups. Mantel-Haenszel tests were used for survival curve hazard ratio comparisons with the low-risk group being the reference group.

Results

There were 222 consecutive patients identified. Seven had known AF and nine had AF detected on the admission ECG/telemetry. Of the remaining 206 patients, the FIND-AFDAS algorithm could be calculated in all patients. There were 154 (74.8%) low-risk patients and 52 (25.2%) high-risk patients (Table 1). Prolonged ECG monitoring was performed in 153 (74.3%) patients, with 152 receiving Holter ECG monitoring (median duration: 72 hours) and two receiving an ILR. Monitoring rates were similar between low and high-risk patients (76.0% vs 69.2% respectively; p = 0.336). Of 152 Holter monitors fitted, AF was detected in nine (5.9%) patients. AF was diagnosed in an additional four (2.0%) patients within one year from stroke.

The primary outcome occurred in 26 (16.9%) patients in the low-risk group and 14 (26.9%) in the high-risk group (HR 1.762, 95% CI: 0.856 to 3.626). All-cause death occurred in 14 (9.1%) low-risk patients and 10 (19.2%) high-risk patients (HR 2.517; 95% CI: 0.993 to 6.377) (Figure 1). At 12 months, 39 (75%) of the high-risk patients were alive, had not had AF detected and had not received prolonged ECG monitoring longer than a 72-hour Holter.

Conclusion

The FIND-AFDAS algorithm can be implemented in routine care. In routine practice, there is a low rate of prolonged ECG monitoring longer than a 72-hour Holter. This is a diagnostics gap for high-risk patients.Table 1Figure 1