INVESTIGATING THE QUALITY OF MACHINE LEARNING RESEARCH AND REPORTING IN HYPERTENSION

Abstract

Objective: Artificial intelligence and machine learning (AI/ML) are increasingly being applied to big clinical data to tackle research questions that cannot be answered with traditional statistical methods. The field is still in its nascent stages and there is a paucity of guidelines for conducting and reporting AI/ML research in hypertension. The objective was to apply the HUMANE checklist to survey the present landscape of AI/ML in hypertension to inform the development of hypertension-specific guidelines and recommendations. Design and method: The HUMANE checklist was developed by global clinical and AI/ML experts through the Delphi method. It assesses the quality of medical AI/ML articles based on whether they cover subjects expected in any peer-reviewed, clinical or AI/ML research publication. A cooping review was carried out to identify articles presenting original research in AI/ML and hypertension published in 2019–2021. Two independent reviewers applied the checklist to each article and in the case of discordance, the response was adjudicated by an AI/ML expert. Results were analysed to assess compliance with the survey (% of papers satisfying checklist requirements). Results: A total of 63 manuscripts was reviewed. A summary of results is shown in Figure 1. Highest compliance was seen for items relating to general article presentation, with compliance ranging from 68% to 98% (description of statistical analysis methods and background context, respectively). Lowest compliance was seen with checklist items relating to clinical research and AI/ML methods. 44% of reviewed articles described the demographics of their dataset and 48% stated their inclusion/exclusion criteria. Nonetheless, datasets were deemed appropriate for investigative aims in 93% of articles. 30% of manuscripts reported their calibration measures, while 73% stated their performance metrics. Internal validation was carried out in 75% of studies, but external validity was assessed in only 14% of cases. Algorithmic bias was addressed in 11% of papers. Conclusions: Application of AI/ML methods in hypertension research is growing, but the majority of current work has major shortfalls in reporting quality, model validation and algorithmic bias. Our study identifies areas of improvement to enable the full realisation of the potential of AI/ML in hypertension.