A new technique for the prediction of heart failure risk driven by hierarchical neighborhood component-based learning and adaptive multi-layer networks

The recently evolving remote healthcare technology could potentially aid the realization of cost-effective and lasting solutions to life-threatening diseases such as heart failure. Such a remote healthcare system should integrate an effectual heart failure risk monitoring and prediction platform. However, developing a heart failure risk (HFR) prediction method that objectively incorporate individual contributive characteristics of HFR risk factors, that are required for adequate prediction remains a challenge. Towards addressing this research gap, a new approach driven by hierarchical neighborhood component-based-learning (HNCL) and adaptive multi-layer networks (AMLN) is proposed. In the proposed method, the HNCL module firstly learns the interrelations among the HFR attributes/ risk factors to construct a set of informative features, regarded as the global weight vector that reflects individual contribution of each risk factor. Subsequently, the constructed global weight vector is applied in building an AMLN model for the prediction of HFR. Moreover, the proposed method's performances were extensively validated with a benchmark clinical database of potential heart failure patients and compared with previous studies using prediction accuracies, performance plots, receiving operating characteristic analysis, error-histogram analysis, specificity, and sensitivity metrics. From the experimental results, we found that the proposed method (AMLN–HNCL)​ achieved significantly higher and stable predictions with an improvement of approximately 11.10% over the commonly applied method. Additionally, the proposed method recorded 9.09% and 12.48% improvements for specificity and sensitivity, respectively compared to the commonly applied method. The superiority in performances achieved by the proposed method should be because the interrelations amongst the risk factors were adequately learnt and their individual contribution was objectively accounted for in the prediction task. Thus, we believe that the proposed method could potentially facilitate the practical implementation of accurately robust HFR prediction module in the context of the currently emerging remote healthcare system, especially in Internet of Medical Things (IoMT) systems. Also, the method may be applied in wearable mobile health-care gadgets capable of monitoring the heart failure status in individuals.