Can the development of a patient’s condition be predicted through intelligent inquiry under the e-health business mode? Sequential feature map-based disease risk prediction upon features selected from cognitive diagnosis big data

Highlights

• Online intelligent medical inquiry-based physician’s cognitive diagnosis big data was fused with offline EMR to obtain VEMR.

• A sequential feature map-based model for disease risk prediction was presented to obtain online users’ medical conditions.

• The optimized TRApriori method is used to mine a frequent feature map on the basis of the temporal graph.

• The frequent feature graph to obtain the online user’s reconstruction coefficient and to realize disease risk prediction.

• A neighborhood-based collaborative prediction model was presented for prediction of an online user’s possible diseases.

Abstract

The data-driven mode has promoted the researches of preventive medicine. In prediction of disease risks, physicians’ clinical cognitive diagnosis data can be used for early prevention of diseases and, therefore, to reduce medical cost, to improve accessibility of medical services and to lower medical risk. However, researches involved no physicians’ cognition of patients’ conditions in intelligent inquiry under e-health business mode, offered no diagnosis big data, neglected the values of the fused text information generated by joint activities of online and offline medical data, and failed to thoroughly analyze the phenomenon of redundancy-complementarity dispersion caused by high-order information shortage from the online inquiry data-driven perspective. Besides, the risk prediction simply based on offline clinical cognitive diagnosis data undoubtedly reduces prediction precision. Importantly, relevant researches rarely considered temporal relationships of different medical events, did not conduct detailed analysis on practical problems of pattern explosion, did not offer a thought of intelligent portrayal map, and did not conduct relevant risk prediction based on the sub-maps obtained from the map. In consequence, the paper presents a disease risk prediction method with the model for redundancy-complementarity dispersion-based feature selection from physicians’ online cognitive diagnosis big data to realize features selection from the cognitive diagnosis big data of online intelligent inquiry; the obtained features were ranked intelligently for subsequent high-dimensional information shortage compensation; the compensated key feature information of the cognitive diagnosis big data was fused with offline electronic medical record (EMR) to form the virtual electronic medical record (VEMR). The formed VEMR was combined with the method of the sequential feature map for modelling, and a sequential feature map-based model for disease risk prediction was presented to obtain online users’ medical conditions. A neighborhood-based collaborative prediction model was presented for prediction of an online intelligent medical inquiry user’s possible diseases in the future and to intelligently rank the risk probabilities of the diseases. In the experiments, the online intelligent medical inquiry users’ VEMRs were used as the foundation of the simulation experiments to predict disease risks in chronic obstructive pulmonary disease (OCPD) population and rheumatic heart disease (RHD) population. The experiments demonstrated that the presented method showed relatively good metric performances in the VEMR and improved disease risk prediction.

Introduction

With the growth in the living standard, people are paying increasingly attention to health. The rapid development of the internet industry has facilitated lots of online medical communities and the business mode of information platform-based intelligent inquiry to provide users with multiple accesses to medical information. These e-health business platforms mostly focus on health knowledge, disease information, and medical news and so on, and also serve users with functions of online medical consultation. Well-known commercial e-health websites in China include Sina Health, QIUYI.CN, haoyisheng.com, ask.39.net, etc., while famous foreign websites include patientslikeme, Daily Strength, well Sphere and MDJunction. It is suggested by investigation that only XYWY.com contains diseases-concerned Q&A data for above a decade from November 24, 2004, and has thousands of newly raised questions every day. As time goes on, the diseases-concerned information will accumulate to form big data. Such big data is a product of people’s extensive involvement, includes substantial true cases, and has high medical values. However, the big data presents a trend of high complexity and high dimensions on the commercial medical platforms. A high amount of feature dimension would increase the complexity of learning algorithm complexity for online medical data, reduce classification performance, and results in failure concerning the problem of feature sparsity in intelligent online medical inquiry-based physicians’ cognitive diagnosis big data; consequently, offline EMR cannot be fused for analysis based on sequential feature map, and online intelligent medical inquiry users’diseases risks will not be obtained.

Feature selection is a powerful tool to overcome the “dimension curse” faced with a learning algorithm. (Pascucci, 2002) Namely, feature selection can be used to select specific inquiry data on an online medical platform, and the selected feature subsets are then adopted for construction and training of a learning model. Therefore, under the background of e-health commercial platforms and big data-driven science, feature selection has become an important research orientation for network economic decision-making and smart business decision-making. On that account, feature selection methods fitting for physicians’ cognitive diagnosis big data of specific intelligent inquiry on e-health commercial platforms were studied thoroughly in this paper. Feature selection in physicians’ cognitive diagnosis big data of intelligent online inquiry is generally based on feature evaluation, which has been a hot and focus problem in feature selection researches and has led to plentiful measures, standards and methods for feature evaluation. Unfortunately, most of the methods carry a presupposed assumption or a given constant; more impertinently, weights of relevance and redundancy in physicians’ cognitive diagnosis big data of intelligent online inquiry cannot be identified with these measures, standards and methods. Therefore, it is very necessary to find a more effective non-prior parameter-based feature evaluation method for cognitive diagnosis big data. Based on in-depth exploration on redundancy and complementarity dispersions caused by high-order information loss in physicians’ cognitive diagnosis big data of intelligent online inquiry, possible conditions of high-order information was judged from low-order inter-feature relationships from the perspective of high-order mutual information “projection” in a low-order domain, and parameters were hereby determined according to the relationships of low-order item features (Bolón-Canedo, Sánchez-Maroño, & Alonso-Betanzos, 2016). Then, a feature selection method based on big data concerning redundancy- complementarity dispersion of intelligent online inquiry-based physicians’ cognitive diagnosis big data was presented.

Based on abovementioned feature selection in physicians’ cognitive diagnosis big data, the key feature information was fused with offline EMR for prediction of online intelligent medical inquiry users’ diseases risks, suggesting a promising future for change of current medical risks. With offline EMR as the principal part, VEMR is digital medical record centered on online intelligent medical inquiry-based physicians’ cognitive diagnosis big data, allows collection of one user-concerned physicians’ cognitive diagnosis information at different timings, integrates offline EMR in the form of online text, and acts as a main carrier guiding big data-driven medical researches (Nambiar, Bhardwaj, Sethi, & Vargheese, 2013; Wu, Roy, & Stewart, 2010). Firstly, obtain key feature information concerning an online user over time from feature selection of online intelligent medical inquiry-based physician’s cognitive diagnosis big data. Secondly, obtain key feature information, including some typical heterogeneous medical data, such as demographic statistics (i.e. age and gender), disease (i.e. disease name and symptom), therapeutic approach and test (i.e. drugs, lab results and physician’s instruction) and so on (Jensen, Jensen, & Brunak, 2012); data heterogeneity requires different modeling technologies for effective analysis and multiple options for their combinations. However, direction application of the key information will be highly challenging due to the heterogeneous sparsity of such key information, so, before specific analysis of medical risks, consistency processing is required for key feature information of every online intelligent medical inquiry user’s data (Hripcsak & Albers, 2012; Pathak, Kho, & Denny, 2013); namely, fuse the key feature information after feature selection with offline EMR, and transform them into clinically relevant VEMR feature information. There have been some researches expressing offline medical big data after consistency processing and exploring medical application with the obtained data, but there is rare research considering temporal relationships between online and offline medical events in the expressing form of a map. Presentation with a map is simple, direct and understandable, and facilitates a physical’s understanding on a patient’s health condition. Temporal relationship is a key factor for medical risk prediction, provides important information concerning possible disease(s) in the future, and contributes to timely adoption of effective preventive measures. Integrate a map with temporal relationships; fusion of the key feature information after feature selection of online users’ intelligent medical inquiry-based physician’s cognitive diagnosis big data with offline EMR is expressed in a sequential feature map; the method of integrating map evolution-covered time dimension information in common map information will effectively solve the problem concerning sparsity of online physicians’ cognitive diagnosis big data and, therefore, allow the sequential feature map-based disease risk prediction research in online users.

Therefore, the key feature information after feature selection of online intelligent medical inquiry-based physician’s cognitive diagnosis big data was fused with offline EMR in the paper to form VEMR. A map was integrated with temporal relationships to form a sequential feature map, which was used to describe online intelligent medical inquiry users’ VEMR data and based on which disease risk prediction was analyzed. The key feature information was obtained via redundancy- complementarity dispersion data-driven features; based on the presentation form of map and the VEMR temporal relationships, optimized TRApriori algorithm was used to mine a frequent-feature map upon temporal profile graph; relying on graph reconstruction theories, the frequent-feature map was used to reconstruct online intelligent medical inquiry users’ portraits, to obtain their reconstruction coefficients and to realize prediction of diseases risks. In addition, a neighborhood-based collaborative prediction model of online intelligent medical inquiry user’s sequential feature map was proposed and then used to obtain a table of ranked diseases probabilities and corresponding diseases’ possible risk initiating times existing in an online intelligent medical inquiry user’s disease risk portrait.