Cholesterol Paradox: Multi-point Heart Failure Prediction from Lipid Panels in Electronic Health Records
- 1North Carolina Agricultural and Technical State University
Abstract
In cardiovascular medicine, low-density lipoprotein cholesterol (LDL-C) is the main lipid marker used in routine screening to assess cardiovascular risk and guide preventive treatment and dietary changes. An increasing number of studies have recently documented the cholesterol paradox: patients with or at risk of heart failure (HF) often have unexpectedly low levels of LDL-C despite leading to cardiovascular diseases. This paper examines how well retrospective lipid panel measurements, supplemented with apolipoprotein B (ApoB), can predict an initial heart failure (HF) event in several pre-diagnosis time windows, using age-matched patient data from Electronic Health Records (EHRs). This retrospective study uses lipid profiles from multicenter nationwide EHRs from the All of Us Research Program. We apply machine learning and explainable AI (SHAP) to lipid-based prediction and to interpret cholesterol contributions to the diagnosis of first heart failure (HF). One experiment predicts the diagnosis of HF in 36 months; the other distinguishes HF patients from people who have never been diagnosed with cardiovascular disease. A baseline lipid panel that includes ApoB predicts the first HF event up to 36 months in advance, achieving an AUC of 0.624 in men and 0.627 in women. Compared to a group of the same age without previous cardiovascular disease, a lipid profile including ApoB predicts a first HF event with a substantially higher AUC: 0.937 for men and 0.925 for women. Shapley value analyses suggest that higher triglycerides and lower HDL and LDL levels are linked to subsequent HF events, with LDL appearing as the strongest yet counterintuitive predictor. ApoB ranks second in importance; in males, higher ApoB is associated with a greater risk of HF, unlike LDL. In general, the study highlights an unexpected association with LDL and emphasizes the prognostic utility of sex-specific metabolic indicators (TG, HDL) for HF using routinely collected clinical data.
Related
- Dynamic Prediction to Predict Virologic Failure in Antiretroviral Treated People Living with HIV: Cohort Analysis of Routine Electronic Health Records Data in the Western Cape, South Africa
- Dynamic near-term risk prediction of clinically significant immune-related adverse events using longitudinal electronic health records
- Cohort Versus Nested Case-Control Design for Prediction Models of Rare Outcomes Using Electronic Health Records
- Prediction of incident inflammatory bowel disease from routine electronic health records: development and external validation of a longitudinal machine learning model
- Scalable Causal-Interpretable Machine Learning for Cancer Prescreening Using Electronic Health Records
BibTeX
@misc{hou2026cholesterol,
title = {{Cholesterol Paradox: Multi-point Heart Failure Prediction from Lipid Panels in Electronic Health Records}},
author = {Y. Hou and M. K. Bhuiyan and M. Samad},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.21.26366145},
url = {https://www.medrxiv.org/content/10.64898/2026.10.21.26366145v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Cholesterol Paradox: Multi-point Heart Failure Prediction from Lipid Panels in Electronic Health Records
AU - Y. Hou
AU - M. K. Bhuiyan
AU - M. Samad
PY - 2026
DA - 2026/10/21
PB - medRxiv
DO - 10.64898/2026.10.21.26366145
UR - https://www.medrxiv.org/content/10.64898/2026.10.21.26366145v1
AB - In cardiovascular medicine, low-density lipoprotein cholesterol (LDL-C) is the main lipid marker used in routine screening to assess cardiovascular risk and guide preventive treatment and dietary changes. An increasing number of studies have recently documented the cholesterol paradox: patients with or at risk of heart failure (HF) often have unexpectedly low levels of LDL-C despite leading to cardiovascular diseases. This paper examines how well retrospective lipid panel measurements, supplemented with apolipoprotein B (ApoB), can predict an initial heart failure (HF) event in several pre-diagnosis time windows, using age-matched patient data from Electronic Health Records (EHRs). This retrospective study uses lipid profiles from multicenter nationwide EHRs from the All of Us Research Program. We apply machine learning and explainable AI (SHAP) to lipid-based prediction and to interpret cholesterol contributions to the diagnosis of first heart failure (HF). One experiment predicts the diagnosis of HF in 36 months; the other distinguishes HF patients from people who have never been diagnosed with cardiovascular disease. A baseline lipid panel that includes ApoB predicts the first HF event up to 36 months in advance, achieving an AUC of 0.624 in men and 0.627 in women. Compared to a group of the same age without previous cardiovascular disease, a lipid profile including ApoB predicts a first HF event with a substantially higher AUC: 0.937 for men and 0.925 for women. Shapley value analyses suggest that higher triglycerides and lower HDL and LDL levels are linked to subsequent HF events, with LDL appearing as the strongest yet counterintuitive predictor. ApoB ranks second in importance; in males, higher ApoB is associated with a greater risk of HF, unlike LDL. In general, the study highlights an unexpected association with LDL and emphasizes the prognostic utility of sex-specific metabolic indicators (TG, HDL) for HF using routinely collected clinical data.
ER -