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

Highlights

• A novel integrated approach for adequate heart failure risk prediction is proposed.

• The proposed method’s performances were validated via standard performance metrics.

• The obtained results indicated that our method is better than existing methods.

• Our method may provide improved and realistic prediction in clinical applications.

Abstract

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.

Introduction

The emergence of Internet of Things (IoT), in which small-sized smart sensing devices are configured to capture and transmit data related to environment conditions, human actions, and biological signals, is arguably the next wave of technological move [1], [2], [3]. Considering the potential of IoT concept, it has been embraced by researchers in the healthcare community which led to the currently evolving Internet of Medical Things (IoMT) paradigm [4], [5]. Meanwhile, the successful implementation of such a remote healthcare system would normally provide, efficient diagnostic decision support platforms and secured mechanism for patient’s medical information transmission. Unlike the former, the latter had received substantial research attention in the recent years [3], [6], [7], [8]. Therefore, the implementation of efficient diagnostic decision support platform in the context of the emerging healthcare paradigm has rarely been considered, thus necessitating the need for further research.

The emerging IoMT paradigm fundamentally require the use of health information from multiple sources including miniaturized sensors, medical devices, and applications connected to electronic health records as shown in Fig. 1. Importantly, the diagnostic decision support platform incorporated in such a system is considered a core component since it is aimed at providing consistently accurate and timely diagnoses outcomes upon which effective therapies are administered, especially for chronic diseases. Amongst the existing chronic diseases, heart failure has received significant research attention due to its complicated diagnostic procedure [9], [10]. At the same time, cardiac related diseases would normally lead to different degrees of complications that may cause reduced quality of life and even deaths across developed countries [9], [10], [11]. Additionally, the morbidity and mortality associated with heart failure are reported to be more prominent in developing countries characterized by inadequate healthcare facilities [12]. Hence, a computationally adequate method that could guarantee accurate and early prediction of the heart failure risk (HFR) in patients is highly essential to mitigating all forms of associated risk in the emerging remote healthcare paradigm [13], [14].

In the recent years, medical decision support systems based on emerging intelligent computing paradigms have been proposed to diagnose HFR and its complications. These diagnostic systems usually adopt algorithms that are machine learning-driven including support vector machine, reinforcement learning, fuzzy logic, and artificially inspired neural-network (ANN) among others for the prediction of HFR in patients [15], [16], [17], [18], [19]. In fact, the ANN-based methods are arguably the most adopted machine learning algorithms for HFR prediction and in other fields because they could effectively model linear and non-linear problems with great learning characteristics [20].

For instance, in literature [21], a three-phase model driven by ANN for HFR diagnosis that could be seamlessly integrated into the existing hospital information systems is proposed with prediction accuracy of 88.89%. By utilizing statistically driven enterprise miner system of version 5.2, Resul and colleagues built an ANN based-ensemble method for HFR diagnosis that achieved an average prediction outcome of 89.01% with a sensitivity of 80.95% [22]. In another study conducted recently by Kim and Kang, the use of a feature correlation-based ANN model for coronary heart disease prediction was proposed. The experimental results of their proposed method using dataset acquired from potential HFR patients indicated that their model could achieve 0.749, which was found to be better than that of the standard Framingham risk score (0.393) based on the receiver operating characteristic evaluation metric [22]. While these studies have been significant towards enhancing the prediction of HFR in patients, a major limitation, is that the existing ANN-based methods assume that HFR factors otherwise known as heart failure diagnoses variables, contribute equally towards the prediction. Nonetheless, it has been established previously that the risk factors have non-uniform contributions towards the HFR prediction, and this has rarely been considered in the available HFR prediction methods [16], [23], [24], [25]. Therefore, it is rational to argue that the existing ANN-driven models may not exactly represent patients’ diagnosis status, thus leading to the administration of inadequate treatment measures that may cause reduced quality of life or death at worst case scenario [26], [27], [28], [29], [30]. Towards resolving this issue, a ranking technique that firstly computes a set of local weights representing individual heart failure attribute’s contribution upon which a global weight vector of the attributes’ alternatives is determined was implemented. Then in the second stage, an ANN model is constructed and trained for HFR prediction based on the initially obtained global weight vector [31]. However, it is worthy to note that in the study, the attributes’ local weights were determined based on report from medical practitioners that were rather subjective and sometimes inaccurate. It should be noted that this approach to determining the attributes’ local weights is limited by human errors which are inevitable, and may lead to imprecise and somewhat bias diagnosis outcome, that may cause inadequate therapy [32], [33]. Importantly, methods that intelligently learns the interactions amongst heart failure risk factors to construct a reduced set of highly informative features for accurate and reliable HFR prediction have not been considered to date. Therefore, an objective means of constructing the individual contributive weights of the risk factors/diagnoses attributes that would guarantee consistently accurate prediction of HFR in patients is still a challenge to date.

Towards developing an intelligently efficient clinical decision support platform that could be integrated into the emerging remote healthcare paradigm for HFR prediction, the following systematic procedures were ensured in this study. Firstly, a technique called hierarchical neighborhood component-based-learning (HNCL) was implemented to adequately construct the heart failure attributes’ local weights upon which the global weight vector that reflects the contribution of each attribute towards heart failure diagnosis was constructed. Secondly, the constructed global weight vector was applied to build an adaptive multi-layer network (AMLN) model that was subsequently utilized for the HFR prediction task. Thirdly, the proposed integrated diagnostic method’s (AMLN–HNCL)​ performances were extensively assessed using an online clinical database of potential heart failure patients and as well compared with the other commonly applied methods using different standard evaluation metrics. In summary, towards addressing the limitations of the existing heart failure risk diagnosis methods this study has contributed to the existing body of knowledge by developing:

(I) An objective and computationally efficient technique driven by hierarchical neighborhood component-based learning that constructs the individual contribution of heart failure risk factors.

(II) A multilayer network that adaptively tunes its connection weights based on the individual contribution of heart failure risk factors to minimize the prediction error associated with the task diagnosis.

(III) The performance of the proposed heart failure risk prediction method was extensively validated in comparison to notable existing methods, and its potential applicability in real-life was also verified.