Improved detection of congestive heart failure via probabilistic symbolic pattern recognition and heart rate variability metrics

• Published in: International journal of medical informatics (Shannon, Ireland) Vol. 108; pp. 55 - 63
• Main Authors: Mahajan, Ruhi, Viangteeravat, Teeradache, Akbilgic, Oguz
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.12.2017
• Subjects: Case-Control Studies; Congestive heart failure; Databases, Factual; Ensemble bagged trees; Female; Heart Failure - diagnosis; Heart Rate; Heart rate variability; Humans; Male; Middle Aged; Pattern Recognition, Automated - methods; Probabilistic symbolic pattern recognition; RR intervals; Probabilistic symbolic pattern recognition; Ensemble bagged trees; Congestive heart failure; Heart rate variability; RR intervals
• ISSN: 1386-5056; 1872-8243
• DOI: 10.1016/j.ijmedinf.2017.09.006

•A probabilistic symbolic pattern recognition (PSPR) approach was implemented to extract features from R-R intervals.•Features extracted by PSPR and time-domain HRV statistics were used in an ensemble bagged trees classifier.•The final classifier yielded an overall accuracy, specificity, and sensitivity of 99.5%, 100%, and 98.6%, respectively.•Comparison with the state-of-the-art techniques, our approach is amenable to accurately identify subjects with CHF condition. A timely diagnosis of congestive heart failure (CHF) is crucial to evade a life-threatening event. This paper presents a novel probabilistic symbol pattern recognition (PSPR) approach to detect CHF in subjects from their cardiac interbeat (R-R) intervals. PSPR discretizes each continuous R-R interval time series by mapping them onto an eight-symbol alphabet and then models the pattern transition behavior in the symbolic representation of the series. The PSPR-based analysis of the discretized series from 107 subjects (69 normal and 38 CHF subjects) yielded discernible features to distinguish normal subjects and subjects with CHF. In addition to PSPR features, we also extracted features using the time-domain heart rate variability measures such as average and standard deviation of R-R intervals. An ensemble of bagged decision trees was used to classify two groups resulting in a five-fold cross-validation accuracy, specificity, and sensitivity of 98.1%, 100%, and 94.7%, respectively. However, a 20% holdout validation yielded an accuracy, specificity, and sensitivity of 99.5%, 100%, and 98.57%, respectively. Results from this study suggest that features obtained with the combination of PSPR and long-term heart rate variability measures can be used in developing automated CHF diagnosis tools.