Impact of Atrial Fibrillation Prediction Model on Prescribing Oral Anti-Coagulation in Patients With Cryptogenic Stroke

Patients presenting to outpatient or inpatient stroke services or to the inpatient cardiac procedure unit for implantation of implantable cardiac monitor at the hospitals of UPMC and who have suffered an ischemic stroke without overt clinical etiology will be considered for participation in this clinical trial. The research coordinator for the trial will screen the inpatient services and outpatient clinics and will verify that any prospective subject meets all the inclusion and none of the exclusion criteria. Once this information is ascertained, the investigator, research coordinator, or both will approach the patient to explain the rationale of the study and its requirements after a clinician known to the participant has obtained their permission to be approached. After signing written informed consent, patients will be enrolled into the study whereby the normal sinus rhythm EKG will be analyzed using the novel machine learning method to assess the patient's risk for developing AF in the following year. Participants will undergo standard of care monitoring on an outpatient basis which includes initially a wearable monitor and then an implantable monitor if the wearable monitor does not reveal AF. If any monitoring reveals AF, then the patients will be on anticoagulation, as indicated in any patient with stroke and AF. This information will be shared with the treating neurologist.

Once a patient is enrolled, their baseline normal EKG will be analyzed using the machine learning algorithm and their risk category for developing AF will be determined. The results which are classified into low, intermediate, and high risk will be presented to the stroke team. The stroke specialists will then share these risk prediction data with their patient for shared decision-making regarding anticoagulation strategy -- oral anticoagulation (OAC) vs. Aspirin. When the information from the novel learning method is shared with the treating neurologist or stroke specialist, the research team will remind them that this is a research tool and has not been fully validated.

The machine learning algorithm, referred to for research purposes as the UPMC AF Prediction Model, was developed under STUDY20030038. The algorithm performance was tested and validated on more than 26,000 patients, 23% of whom developed AF over the course of one year after their original normal EKG. The algorithm performance shows an area under the Receiver Operating Characteristic curve of 0.800 with an F1 score of 0.42. Accordingly, the risk groups in the study were defined as follows, after consultation with stroke specialists:

A. Low risk group, based on a negative predictive value of 97.5% (i.e., only 2.5% of patients classified as low risk will develop AF over 1-year of follow-up) B. High risk group, based on a positive predictive value of 30% (i.e., 30% of patients classified as high risk will develop AF over 1-year of follow-up) C. Intermediate risk group, which will include patients who do not belong in either of the two previous risk groups

Patients will be followed up for 90 days for clinical events. Follow-up at 30 and 90 days will consist of phone calls and review of electronic health records. The rate of use of OAC will be assessed at 30 and 90 days in each AF risk category.