Research Of Modeling Theory And Method In Computational Psychophysiology

The field of psychophysiology is concerned with the manipulation of psychological variables and their corresponding observed effects on physiological processes. Thus, psychophysiology is studies in which psychological factors serve as independent variables and physiological responses serve as dependent variables. Further more, the studies in psychophysiology want to implement psychophysiological inference based on strong psychophysiological relationships. However, traditional methods for analyzing complex psychophysiological relationships do not apply in physiological big data. In the era of "big data", how to analyze complex psychophysiological relationships and find reliable, replicable and general one-to-one relationship, has become a challenge in psychophysiology. In order to break through these limitations, we put forward a new research concept named "Computational Psychophysiology" which is the marrying of psychophysiology with advanced computational and analysis techniques. This has allowed psychophysiology researchers to perform far more complex analysis of multimodal physiological measurement data, may lead to a better grasp of the complex psychophysiological mapping relationships, and make the interpretation, evaluation and reasoning of different mental states more objectively and quantitatively. Finally, this may help to identify new phenotypes for normal and abnormal psychological functions and provide an engineering methodology for clinical diagnosis of mental disorder.As a new research concept, there is a lack of a systematic theory and research framework in "Computational Psychophysiology". Simultaneously, the multimodal physiological measurement data usually has characteristics of poor stability, small sample size, and may be affected significantly by individual differences. This has brought new challenges to instantaneous, reliable, effective computing and reasoning. In this dissertation, we focuses on construction of research framework and innovation of prediction algorithms in "Computational Psychophysiology ", the main work and contributions are as follows:(1) Theory and research framework of "Computational Psychophysiology " : In view of lack of a systematic theory and research framework, we systematically expounded the necessity of studying psychophysiological problems using the "Computational Thinking", of quantitatively perceiving and computing psychological states from the perspective of engineering for the first time. We also analyzed the feasibility of psychophysiological reasoning based on probability and eventually constructed a preliminary theory and research framework of "Computational Psychophysiology" by combining the data-driven approaches, theory-driven approaches and ontology-based context modeling. Simultaneously, we considered three major dimensions that may influence the psychophysiological relations adequately, such as multi-modal data, temporal variation and individual context. The above theories have an important reference value to the systematic research and clinical application of "Computational Psychophysiology".(2) NLC kernel function for “Computational Psychophysiology”: Because the multimodal physiological measurement data is usually with poor stability, small sample size and affected significantly by individual differences, we constructed an infinite geometric series based on the normalization of first-order polynomial kernel function under the inspiration of multiple kernel learning(MKL) methods. When the sum of this geometric series converged, we got a new normalized linear combination(NLC) kernel function. Theoretical analysis result shows that NLC kernel function has no arguments, and can be self-adaptive to data distribution. These advantages would ease the influence of data stability and individual differences, reduce the time complexity of model selection and parameter optimization, and eventually make the model own better classification performance and stronger generalization ability. The experiment result demonstrates that, on different types of physiological measurement data, classification performance of NLC kernel function is statistically significant equal to or even better than Polynomial and Gaussian kernel function whose parameters have been optimized using grid search method. With the significant reduction in time complexity, NLC kernel function would lay the foundation for instantaneous, reliable and effective psychophysiological reasoning.(3) Multi-task feature selection method for “Computational Psychophysiology”: Since the high-dimensional and small-sampling physiological measurement data could lead to "Curse of Dimensionality", we proposed a new feature selection method under the inspiration of multi-task learning theory. Because there is no “gold standard” to validate the psychophysiological inference, we need to establish concurrent validity with reference to other types of dependent measures, such as subjective self-report or observable behavioural markers, and validate the inference by measuring the relationship between psychophysiological states and other measures. Therefore, in order to increase reference information for feature selection, we introduced the relationship between psychophysiological states and other measures as auxiliary information using Multi-task Joint Classi?cation and Regression(MTJCL) algorithm, and selected feature subsets with stronger discriminant ability and smaller redundancy. Besides, these feature subsets may ensure the concurrency of different types of validation conclusions, and enhance the reliability and accuracy of psychophysiological reasoning further. The experiment result shows that we can obtain better classification and prediction performance on feature subsets selected by our method in comparison with other four feature selection algorithms. Simultaneously, main features selected by MTJCL algorithm are consistent with previous physiological or pathological research conclusions.