ABSTRACT
BACKGROUND: Peripheral artery disease (PAD) is underdiagnosed due to poor patient and clinician awareness. Despite this, no widely accepted PAD screening is recommended. OBJECTIVES: The authors used machine learning to develop an automated risk stratification tool for identifying patients with a high likelihood of PAD. METHODS: Using data from the electronic health record (EHR), ankle-brachial indices (ABIs) were extracted for 3,298 patients. In addition to ABI, we extracted 60 other patient characteristics and used a random forest model to rank the features by association with ABI. The model identified several features independently correlated with PAD. We then built a logistic regression model to predict PAD status on a validation set of patients (n = 1,089), an external cohort of patients (n = 2,922), and a national database (n = 2,488). The model was compared to an age-based and random forest model. RESULTS: The model had an area under the curve (AUC) of 0.68 in the validation set. When evaluated on an external population using EHR data, it performed similarly with an AUC of 0.68. When evaluated on a national database, it had an AUC of 0.72. The model outperformed an age-based model (AUC: 0.62; P < 0.001). A random forest model with inclusion of all 60 features did not perform significantly better (AUC: 0.71; P = 0.31). CONCLUSIONS: Statistical techniques can be used to build models which identify individuals at high risk for PAD using information accessible from the EHR. Models such as this may allow large health care systems to efficiently identify patients that would benefit from aggressive preventive strategies or targeted-ABI screening.
ABSTRACT
BACKGROUND/AIMS: Extracellular volume (ECV) overload is a mortality risk factor in hemodialysis patients, but no standard approach exists to objectively assess this clinically. We aimed to quantify relationships between slopes of repeated intradialytic blood pressure (BP) measurements and ECV. METHODS: In a cross-sectional study of 71 hemodialysis patients, we calculated BP slopes from all intradialytic measurements using Gaussian regression. We measured extracellular and total body water (TBW) with bioimpedance spectroscopy. We analyzed unconditional and conditional associations between BP slope and volume metrics with mixed linear models and sensitivity analyses using non-linear intradialytic BP trajectory. RESULTS: Mean systolic intradialytic BP slope (IBPS) was -0.06 (0.1) mm Hg/min. Post-dialysis extracellular water (ECW)/weight was the volume metric mostly strongly associated with slope (r = 0.34, p = 0.007 for unconditional analysis; ß = 1.45, p = 0.001 for conditional analysis). Among subjects with post-dialysis systolic BP ≥130 mm Hg, the association strengthened (r = 0.40, p = 0.006; ß = 1.42, p = 0.003). ECV was more strongly associated with the BP slope than with pre-dialysis, post-dialysis, or delta systolic BP (r = -0.07, 0.19, 0.28; p = 0.6, 0.1, 0.03). In nonlinear models, BP trajectory also had the strongest association with post-dialysis ECW/body weight (p < 0.001). CONCLUSIONS: In hypertensive hemodialysis patients, measurements of ECV excess are more strongly associated with IBPSs than with pre-dialysis, post-dialysis, or change in systolic BP. Among varying volume metrics, post-dialysis ECW/weight has the strongest association with these slopes. Determining IBPS is a novel method to optimize clinical assessment of ECV in hemodialysis patients.
