Mapping the landscape of heart failure – lessons learnt using machine learning

Background Heart Failure (HF) is a complex condition, which is on the increase in UK due to the most rapidly increasing aged population. As a result, HF services are likely ill-equipped to deal with the increasing demand. HF requires timely expert input for diagnosis and rapid medication optimisation to reduce morbidity and mortality, but it is unclear how HF care operates in the Outpatient (OP) settings.

Aims An analysis of the landscape of HF, and confounding factors impacting on prevalence and mortality, through application of various machine learning (ML) techniques to HF datasets.

Methods A systematic literature review on ML application to HF. Exploratory data analysis of open source epidemiological and government data. Supervised and unsupervised ML used for clustering and mortality prediction. Survival analysis of clinical data curated exclusively for the purpose of this research from Southern HSC Trust (SHSCT) electronic health records (figure 1).

Results (1) Literature review showed a lack of clinical experts’ involvement in the development of HF predictive models. The usefulness of ML application in clinical practice was overclaimed. A practical checklist for data scientists was co-designed by clinical experts and published.

(2) The analysis of open-source data from NI shows from that there is a decreasing trend in coronary artery disease (CAD) prevalence, alongside a gradual increase in HF prevalence since 2013. However, the prevalence of HF is at 0.5 – 1%. Based on an analysis of GP catchment areas, HF was more prevalent in rural areas. The HF prevalence was higher in areas located more than 60 minutes by car from 2 primary PCI centres regardless of the urban/rural status.

(3) Mortality analysis of 5121 patients (36% females, median age 75 (SD=13)) attending HF specialist outpatient (OP) clinic in SHSCT between May 2007 and August 2019 (median follow up 8 years (SD 3.8)) showed that 48% (2444) patients died during the follow up period. The 1-year survival was 81% (CI 0.79 – 0.81), 5-year: 53% (CI 0.51 – 0.54), 10-year: 35% (CI 0.32 – 0.37). There was a statistically significant difference in the overall survival of patients who were referred to a specialist HF OP clinic by non-cardiology teams vs. those referred by cardiology teams, p<.005, HR 0.49 (95% CI 0.45 – 0.54) indicating a lower survival rate of patients referred by non-cardiology teams, figure 2. The prescription of evidence-based pharmacotherapy varied significantly between the cohorts referred by cardiology vs. non-cardiology teams (Chi Squared test: 44 p<.005), including angiotensin-converting enzyme inhibitor (ACEi) (78% vs. 67.5%), beta blocker (BB) (85.5% vs. 78.5%), mineralocorticoid receptor antagonist (MRA) (63% vs. 49.5%) and aldosterone receptor/neprilysin inhibitor (ARNi) (16.5% vs. 9.5%), and combination of ACEi/BB/MRA (45% vs. 31.5%) respectively. Patients referred by non-cardiology teams were more frequently prescribed symptomatic treatment such as diuretics: furosemide 74.5% vs. 61.5%, metolazone 10% vs. 5.5%.

Conclusions (1) New generations of healthcare professionals should learn how to curate, analyse and interpret data with the use of ML to efficiently drive change in HF landscape.

(2) HF shows slow increasing trend in NI, with higher prevalence of HF in rural areas.

(3) Patients in NI were less likely to be on optimal medical therapy if referred from general medical and primary care teams to HF services that resulted in lower survival. Findings from this research project will be used to advocate for policy change to argue for an in-patient and in-reach HF service and to improve patient care interface between primary and secondary care. Changes to HF care must be driven by data to justify funding streaming to most appropriate place of need.