2-058 How representative are heart failure clinical trials? A comparative study using natural language processing

Abstract

<h3>Introduction</h3> Natural language processing (NLP) is a powerful tool in the field of healthcare, offering the ability to efficiently analyse vast amounts of data. In randomized controlled trials (RCTs), the representativeness of the participants is crucial for the generalisability of the findings. However, there has been a lack of scalable and effective methods to assess how well these trials reflect the broader patient population. We applied NLP to evaluate the representativeness of heart failure (HF) RCTs cohort by comparing it to a large real-world HF cohort. We aim to uncover disparities and ensure that clinical trials are as inclusive and representative as possible, which is essential for the advancement of patient-centred care. <h3>Methods</h3> This is a retrospective cohort study of patients with HF at a large regional tertiary centre. Identifiers of patients who enrolled in 15 HF RCTs in the centre from 2012 to 2023 were manually collated using local enrolment records. The overall HF cohort was identified using NLP based on CogStack and MedCAT deployed in the centre. The NLP pipeline enables the integration and analysis of both structured and unstructured data, allowing comprehensive cohort characterisation. Demographics, Index of Multiple Deprivation (IMD), comorbidities, symptoms, smoking behaviour and hospitalisation of patients were analysed. Principal component analysis (PCA) was used to visualise the 2 cohorts. <h3>Results</h3> We identified 11,885 patients with a diagnosis of HF. Of these, 236 were recruited into HF RCTs. Characteristics of the patients are shown in table&nbsp;1. The age (figure&nbsp;1) of participants in HF RCTs was significantly younger (66 ± 12 years) compared to the whole HF cohort (70 ± 15 years) (p&lt;0.001). The proportion of females in HF RCTs was significantly lower (27%) than in the whole HF cohort (42%) (p&lt;0.001). No significant differences were found in terms of ethnicity and IMD. Regarding comorbidities, compared to the whole HF cohort, myocardial infarction was more common in the HF RCTs group (56% vs 38%, p&lt;0.001) while kidney disease was less common (48% vs 60%, p&lt;0.001). Both cohorts show similar rates of symptoms, smoking behaviour, and hospitalization. Distributions of the 2 cohorts using PCA are shown in figure&nbsp;2. <h3>Conclusions</h3> Our NLP-based analyses revealed that HF trials at a tertiary cardiology centre were more likely to recruit patients who were younger and male compared to the broader patient population. This study shows NLP methods can be used to assess the representativeness of RCTs. The same approach can be applied across different domains within cardiovascular medicine, and potentially, for any medical condition. For future work, we plan to develop metrics of representativeness that will further quantify the degree to which clinical trials reflect the broader patient population. This can also enhance patient recruitment by incorporating local cohort characteristics into randomisation strategies.

Ähnliche Arbeiten