Recurrent convolutional neural network based multimodal disease risk prediction

Highlights

• We propose a new MD-RCNN for disease risk prediction.

• We propose a feature fusion scheme based on DBN.

• We experiment with the medical big data of a Chinese grade-A hospital of second class.

Abstract

With the rapid growth of biomedical and healthcare data, machine learning methods are used in more and more work to predict disease risk. However, most works use single-mode data to predict disease risk and only few works use multimodal data to predict disease risk. Thus, a new multimodal data-based recurrent convolutional neural network (MD-RCNN) for disease risk prediction is proposed. This model not only can use patient’s structured data and text data, but also can extract structured and unstructured features in fine-grained. Furthermore, in order to obtain the highly non-linear relationships between structured data and unstructured data, we use deep belief network (DBN)to fuse the features. Finally, we experiment with the medical big data of a Chinese two grade hospital during 2013–2015. Experimental results show that the accuracy of MD-RCNN algorithm can reaches 96% and outperforms several state-of-the-art methods.

Introduction

According to McKinsey Report [1], half of Americans suffer from chronic diseases. Furthermore, America spends 27 thousand billion dollars on chronic diseases every year. That accounts for 18% of the annual GDP of America [2]. This situation is also common in other countries. For example, in china, the rate of death rises continuously, accounting for 86.6% of total number of dead people [3]. Therefore, it is crucial to predict the risk of chronic diseases [4].

With the development of deep learning [5] and wearable computing [6], [7], more and more researches predict disease risks from the perspective of big data analysis [8], [9]. Especially, electronic medical record (EHR) [10] becomes more and more convenient. The statistical information [11], detection result and medical history of patients are all recorded in EHR [12]. With some advanced information fusion mechanisms based on distributed systems like [13], data collected from different types of medical sensors and wearables can be used to enhance the information accuracy [14]. Therefore, the potential opportunity centering on data is offered for us to research medical treatment cases and decrease expenditure. For the risk prediction of chronic diseases, many researchers improve the accuracy of risk classification by automatically extracting features from a large amount of structured data [15], rather than the previously selected characteristics [16], [17]. Such as, Yang et al. [18] introduced a multi-label learning algorithm based on convolutional neural network (CNN). Under this algorithm, features can be extracted automatically from EHR for diagnosing diseases.

Furthermore, some research tried to use unstructured text data to predict the risk of chronic diseases [19], [20]. Weng et al. [21] used machine learning technology based on natural language processingto extract features for classification of clinical diagnosis. Jonnagaddala et al. [22] used unstructured electronic medical treatment record to operate text mining for predicting the risk of coronary artery disease. Framingham risk scoring was used to evaluate and calculate the risk of coronary heart disease in ten years, and imputation strategy was used to compensate the lost information. Khalifa et al. [23] suggested using clinical record to evaluate cardiovascular danger factors. However, most existing disease risk assessment schemes only considered structured data or unstructured data [24]. Few works considered multimodal disease risk assessment. For example, Chen et al. [25] propose multimodal disease risk prediction algorithm based on CNN (CNN-MDRP). The result indicated that the combination of structured data and unstructured data could improve the accuracy of model comparing with the only use of text data.

However, existing methods still has the following two problems: (i) As for extracting the features of medical text data, the existing most schemes do not consider context information of text. Therefore, the fine-grained feature of text cannot be extracted [26]. (ii) As for fusing the features of multimodal data, most existing fusion schemes do not consider the highly non-linear relationships among different modes of data [27]. Therefore, the degree of fusion is low.

In order to solve these challenges, we process the structured data and unstructured data of cerebral infarction patients to predict the patients’ disease risk. In details, we propose MD-RCNN for disease risk prediction. We first use RCNN to process unstructured text data to get the textual features of patients’ disease risk. At the same time, referring to the data of patients’ physical examination, the structured features of patients can be got, such as patients’ stature, weight, age, gender, disease and blood type. Then, the deep belief network (DBN) is used to combine structured features and unstructured features deeply. Finally, patients’ disease risk can be predicted. The main contributions of the paper are summarized as follows.

• We propose a MD-RCNN algorithm based on the structured and unstructured data. The RCNN is used to effectively extract fine-grained features of unstructured textual data.

• In order to deal with the highly nonlinear relationship between multimodal data and better fuse the features of structured data and unstructured text data, we propose a feature fusion scheme based on DBN.

• We experiment with the medical big data of a Chinese grade-A hospital of second class. Experimental results show that the proposed MD-RCNN algorithm is better than other prediction algorithms. The accuracy of MD-RCNN can reach 96%.

The rest of the paper is organized as follow. The framework of disease risk prediction is given in Section 2. The MD-RCNN for disease risk prediction model is describe in Section 3. Our result and discussions are given in Section 4. Finally, Section 5 concludes this paper.