Improving stratification and management of patients with atrial fibrillation and heart failure using novel technology

Gill, Simrat (2024). Improving stratification and management of patients with atrial fibrillation and heart failure using novel technology. University of Birmingham. Ph.D.

Preview

Gill2024PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (3MB) | Preview

Abstract

Atrial fibrillation (AF) and heart failure (HF) commonly co-exist, and exhibit higher rates of morbidity and mortality. Treatment remain unclear due to the lack of robust data regarding treatment efficacy in this specific population, and current ones have unpredictable responses. The use of big data and artificial intelligence (AI) based analysis has allowed us to find better ways to identify and sub-categorise this cohort of patients based on treatment response, that can improve outcomes and quality of life.
This PhD looked at the use of big data and machine learning to improve phenotyping and prognoses in patients with AF and HF. A systematic review and meta-analysis of novel screening techniques for AF detection found smartphone photoplethysmography had a high sensitivity and specificity for AF detection when compared to an ECG (n=28 studies). A meta-analysis of 20 comparisons (n=17 studies; n=6,891; 2,299 with AF) found a pooled sensitivity of 94% (95% CI 92-95%) and specificity of 97% (CI 96-98%), with significant heterogeneity (p<0.01). Studies were found to be small, of poor quality and biased.
Following this, a consumer wearable Fitbit device measuring PPG heart rate and physical activity (step count), was embedded within an on-going randomised clinical trial (RATE-AF) to compare the effectiveness of beta blockers and low-dose digoxin in AF rate control. Over 143 million heart rate recordings and 23 million corresponding physical activity intervals were collected over a mean of 20 weeks. No significant difference was seen in heart rate control with beta-blockers or low-dose digoxin, the regression coefficient was 1.22 (95% CI -2.82 to 5.27; p=0.55). There remained no difference in heart rate after adjusting for clinical variables (p=0.75), and individual physical activity (p=0.74). Conventional trial assessments and a wearable neural network trained on sensor data, were used to predict future trial outcomes. The wearable neural network showed a similar performance for predicting future New York Heart Association class: F1 score 0.55 (95% CI 0.40 to 0.70), versus 0.59 for electrocardiogram and 6-minute walk test (95% CI 0.44 to 0.74; p=0.72 for comparison).
The limitations of conventional statistical analysis in the presence of multi-morbidity were highlighted in the meta-analysis, assessing beta-blocker efficacy in heart failure with reduced ejection fraction (HFrEF) and renal dysfunction. This was only able to detect a weak interaction between beta blocker efficacy and renal function, with low power (p=0.021). An AI based clustering technique was evaluated to overcome this. The clustering technique identified clusters of HFrEF patients with different beta-blocker efficacy in sinus rhythm and AF. A cluster of older patients in sinus rhythm were found to have no benefit from beta-blockers, and a cluster of younger patients with AF were found to have a significant benefit.
These studies have highlighted the potential wearable devices and AI-based analysis can offer in terms of improving patient care.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):