Predicting AF after Cardiac Surgery - the PARADISE Score A Clinical Prediction Rule for Post-operative Atrial Fibrillation in Patients Undergoing Cardiac Surgery

Research question

To develop clinical prediction models for clinicians to reliably determine the pre- and immediately post-operative risk of a patient developing Atrial Fibrillation After Cardiac Surgery (AFACS). Background

Atrial fibrillation after cardiac surgery (AFACS) is frequent, occurring in 30-50% of cases. Over 90% of those who develop AFACS do so by day 7 post-surgery. We reviewed available prediction models whilst producing international guidance.

None include modern risk stratification information such as echocardiographic measurements or meet TRIPOD quality criteria. Prophylaxis decreases length of hospital stay and cost. However prophylaxis treatments are not risk-free.

A prospectively validated score allowing targeted prophylaxis is the first step to improving outcomes, informing patients and resource planning. Aims To develop validated clinical prediction models to determine the risk of a patient developing AFACS: • In the pre-operative assessment clinic or on admission for surgery

• On arrival in the post-operative care unit Databases

Two United States retrospective datasets (Partners and Brigham), two United Kingdom (UK) prospective datasets and data from UK prospective trials. Objectives 1. Develop a long list of candidate pre-operative and intra/early post-operative risk factors 2. Enhance the Partners development database to include additional variables from 1.

3. Develop and internally validate pre- and immediate post-operative models 4. Externally validate models retrospectively and prospectively 5. Compare performance to existing prediction models Methods 1. The long list of candidate risk factors will be developed from: • a systematic review of the literature

• a large UK research platform (CALIBER) • expert (including public) opinion (Delphi process) • machine learning of (combinations of) input variables in the CALIBER and Partners databases We will then:

2. Enhance the Partners research database to include the additional variables identified in 1 by extracting the required additional features from the clinical data warehouses.

3. Compare state of the art statistical methods and machine-learning (ML) approaches to develop pre-operative and immediately post-operative prediction models. We will learn from each method to maximise final model performance. We will internally validate the models.

4. Assess external validity of the final models retrospectively using the Brigham dataset, to allow model adjustment if performance declines prior to prospective validation. We will also prospectively validate the final models within a prospective “real world” cohort from two major UK centres and using data from ongoing UK trials.

5. Compare model performance to published models identified in 1. Delivery timelines 30 months - see project timeline Impact and dissemination We will: • collaborate with Arrhythmia and ICU patient groups to maximise patient/public awareness • publish our prediction model in relevant specialist medical journals and present it at conferences

• use social media, podcasts and infographics to further publicise this work

• use the networks we developed writing the latest international AFACS guidelines to publicise our findings, seeking endorsement by relevant professional societies

The impact of our work will be substantial, providing an evidence base for rational AFACS prophylaxis throughout the UK and beyond.