Assessing Machine Learning for Diagnostic Classification of Hypertension Types Identified by Ambulatory Blood...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Inaccurate blood pressure (BP) classification leads to suboptimal treatment strategies.
• Ambulatory BP monitoring (ABPM) is a gold standard but can be resource-intensive.
• Exploring machine learning (ML) as an alternative for BP classification is crucial.

Purpose of the Study:

• To evaluate the efficacy of ML models in classifying BP status using routine clinical data.
• To compare ML-based BP classification with ABPM.
• To assess the long-term cardiovascular and mortality risks associated with ML-derived BP groups.

Main Methods:

• A multicentre study utilized 3 derivation cohorts and 1 independent evaluation cohort.
• ML models were trained using office BP, ABPM, and various clinical data.
• Seven ML algorithms were employed to classify patients into five distinct BP categories.
• Cox proportional hazards models assessed 10-year cardiovascular outcomes and 27-year all-cause mortality.

Main Results:

• Extreme gradient boosting (XGBoost) achieved the highest area under the receiver operating characteristic curve (0.85-0.88).
• However, ML model accuracy (0.57-0.72) and F1 scores (0.57-0.69) were generally low across derivation cohorts.
• The evaluation cohort showed increased cardiovascular event risk in Normal/Target-White-Coat and Hypertension-White-Coat groups.

Conclusions:

• Machine learning shows limited potential for accurate BP classification when ABPM is unavailable.
• Current ML models may not reliably replace ABPM for precise BP status determination.
• Larger, diverse studies are required to enhance ML model performance in varied clinical settings.