Estimating Heart Failure Health Status From Multimodal Wearables: Implementation of a Feature-Engineering Framework for KCCQ-12 Prediction.
7 affiliations
- 1Department of Cardiovascular Diseases, Mayo Clinic, Jacksonville, FL.
- 2OutcomesAI, Inc, formerly of Biofourmis, Inc, Singapore.
- 3Section of General Internal Medicine, Yale School of Medicine, New Haven, CT.
- 4Independent Researcher, Singapore, formerly of Biofourmis, Inc, Singapore.
- 5Singapore Management University, Singapore, formerly of Biofourmis, Inc., Singapore.
- 6Kern Center for the Science of Health Care Delivery, Mayo Clinic, Rochester, MN.
- 7Biofourmis, Inc, Boston, MA.
Abstract
Routine monitoring of symptom burden and quality of life in heart failure (HF) patients is limited by intermittent sampling of validated instruments, such as the Kansas City Cardiomyopathy Questionnaire (KCCQ-12). As more wearable devices track physiological measures in real-world settings than ever before, we propose an exploratory implementation of a feature-engineering framework to bridge this sampling gap, approximating KCCQ-12 scores and patient health status from wearable features. In this narrative, we processed data from the 8-week multicenter observational study of patients with HF (New York Heart Association, class I-III) wearing Everion multisensor smartwatches: "Evaluating Mobile Health Tool Use for Capturing Patient-centered Outcomes Measures in Patients with HF" (NCT04191356). Then, Least Absolute Shrinkage and Selection Operator logistic regression was applied for feature selection, followed by binary classification and linear mixed-effects modeling to predict KCCQ-12 scores. Among patients generating sufficient data (n=59), there were 173 unique KCCQ-12 observations. From 7235 candidate features, Least Absolute Shrinkage and Selection Operator retained 88 predictors (Supplemental Table 1, available online at https://www.mcpdigitalhealth.org/), including the top predictors: passive functional capacity metrics (positive), daily steps/walking time (positive), nocturnal respiratory variability (negative), and patient-reported dyspnea (negative). Binary classification models achieved a test of area under the curve=0.77 (sensitivity, 91%; specificity, 63%), and linear mixed-effects models produced a root mean squared error of 14.7 (R²=0.58) for predicting patient health status and continuous KCCQ-12 scores, respectively. A patient-level split analysis revealed decreased model performance (root mean squared error=18.9; R²=0.33), suggesting limited generalizability to new individuals. This exploratory implementation with a small sample size outlines wearable features that warrant future investigation for KCCQ-12 score monitoring and offers early-stage lessons for assessing such features as proxies of KCCQ-12 scores in real-world settings.
Related
- Estimating presenteeism from repeated smartphone-based multimodal behavioral responses
- Nationwide Implementation of Multimodal Prehabilitation and Complications After Colorectal Cancer Surgery
- Artificial Intelligence-Driven Nanobody Engineering for Pathogen Recognition: From Functional Epitope Prioritization to Multimodal Foundation Models
- From revascularization to multimodal integrated therapy for peripheral arterial disease: immunomodulation, metabolic correction, and tissue engineering
- Echocardiography-based multimodal models for thromboembolic risk prediction in atrial fibrillation: a narrative review
BibTeX
@article{mcleod2026estimating,
title = {{Estimating Heart Failure Health Status From Multimodal Wearables: Implementation of a Feature-Engineering Framework for KCCQ-12 Prediction.}},
author = {Christopher J McLeod and Charles Bruce and Kuldeep Singh Rajput and Joseph S Ross and Sy Jackie Lou Velarde and Felis Dwiyasa and Chen Hao and Yan He and Pyae Phyo Tun and Stephanie Zawada and Maulik Majmudar},
journal = {Mayo Clinic proceedings. Digital health},
year = {2026},
month = sep,
doi = {10.1016/j.mcpdig.2026.100403},
url = {https://europepmc.org/article/MED/42908891},
note = {CC BY},
}
RIS
TY - JOUR
TI - Estimating Heart Failure Health Status From Multimodal Wearables: Implementation of a Feature-Engineering Framework for KCCQ-12 Prediction.
AU - Christopher J McLeod
AU - Charles Bruce
AU - Kuldeep Singh Rajput
AU - Joseph S Ross
AU - Sy Jackie Lou Velarde
AU - Felis Dwiyasa
AU - Chen Hao
AU - Yan He
AU - Pyae Phyo Tun
AU - Stephanie Zawada
AU - Maulik Majmudar
PY - 2026
DA - 2026/09/17
JO - Mayo Clinic proceedings. Digital health
DO - 10.1016/j.mcpdig.2026.100403
UR - https://europepmc.org/article/MED/42908891
AB - Routine monitoring of symptom burden and quality of life in heart failure (HF) patients is limited by intermittent sampling of validated instruments, such as the Kansas City Cardiomyopathy Questionnaire (KCCQ-12). As more wearable devices track physiological measures in real-world settings than ever before, we propose an exploratory implementation of a feature-engineering framework to bridge this sampling gap, approximating KCCQ-12 scores and patient health status from wearable features. In this narrative, we processed data from the 8-week multicenter observational study of patients with HF (New York Heart Association, class I-III) wearing Everion multisensor smartwatches: "Evaluating Mobile Health Tool Use for Capturing Patient-centered Outcomes Measures in Patients with HF" (NCT04191356). Then, Least Absolute Shrinkage and Selection Operator logistic regression was applied for feature selection, followed by binary classification and linear mixed-effects modeling to predict KCCQ-12 scores. Among patients generating sufficient data (n=59), there were 173 unique KCCQ-12 observations. From 7235 candidate features, Least Absolute Shrinkage and Selection Operator retained 88 predictors (Supplemental Table 1, available online at https://www.mcpdigitalhealth.org/), including the top predictors: passive functional capacity metrics (positive), daily steps/walking time (positive), nocturnal respiratory variability (negative), and patient-reported dyspnea (negative). Binary classification models achieved a test of area under the curve=0.77 (sensitivity, 91%; specificity, 63%), and linear mixed-effects models produced a root mean squared error of 14.7 (R²=0.58) for predicting patient health status and continuous KCCQ-12 scores, respectively. A patient-level split analysis revealed decreased model performance (root mean squared error=18.9; R²=0.33), suggesting limited generalizability to new individuals. This exploratory implementation with a small sample size outlines wearable features that warrant future investigation for KCCQ-12 score monitoring and offers early-stage lessons for assessing such features as proxies of KCCQ-12 scores in real-world settings.
SN - 2949-7612
ER -