Research On The Adverse Outcome Prediction Method For Complex Medical Data

The era of big data generates massive medical data which contains rich knowledge and information.Mining and utilizing the knowledge and information from the complex medical data for adverse outcome prediction is a challenge in the field of the medical networking and intelligent health management.This thesis explores how to mine and utilize the knowledge and information from the complexity of medical data in the construction process of datadriven adverse outcome prediction models from feature selection,information aggregation,and model training.The main findings and contributions are as follows:Feature selection for non-linear,multi-state,and time series medical data.To solve the problem the existing feature selection methods can not effectively mine the information from non-linear,multi-state,and time series in medical data,this thesis proposes a feature selection method for non-linear,multi-state,and time series medical data named Deep Survey+Multi+PI.Firstly,the method introduces survival analysis method for feature selection to mining and utilizing the temporal information in time series data via the survival time;secondly,it construct a deep learning-based survival analysis method for feature selection to process the nonlinear mapping relationships in complex medical data;Then,it employs a multi-task learning to drive the feature selection to learn the interaction information between different states from the medical data;Finally,it employs the concept of the permutation importance to calculate the features’ importance.The experimental results indicate that the Deep Survey+Multi+PI can effectively mine and utilize the knowledge and information from the three considered complexity to improve the adverse outcome prediction model’s performances.Multi-center collaborative information aggregation towards heterogeneous and non-I.I.D medical data in context of the privacy protection.To protect the privacy information during the process of the information aggregation,and consider the complexity from the heterogeneous and non-I.I.D of the medical data,this thesis proposes a multi-center collaborative information aggregation method towards heterogeneous and non-I.I.D medical data in context of the privacy protection named TNHDFL+ CWAVG.Firstly,the method employs the federated learning to achieve multi-center collaborative information aggregation without directly exchange the raw medical data;Then the method employs the knowledge distillation method and a contribution-based weighted averaging algorithm to process the complexity of heterogeneous and non-I.I.D from the medical data.The experimental results indicate that the TNHDFL+ CWAVG can effectively achieve information aggregation of heterogeneous and non-I.I.D medical data.Time series classification towards the right-censored and cost-sensitive.Considering the complexity from the right-censored and cost-sensitive of the adverse outcome prediction task,and the complexity from calculating the similarity between time-series samples,this these proposes a cost-sensitive learning strategy toward right-censored time series medical data.This these uses the learning strategy to train a classification to achieve early predication for adverse outcomes.Firstly,the strategy employs the PU learning which regard right-censored medical data as a PU dataset to construct the strategies main framework;Secondly,it employs a cost-sensitive method under the PU learning framework to adjust model parameters to escaping false negativity;Then it employs a property and outlier weighted time series similarity to selecting reliable negative samples from the unlabeled samples.The experimental results show that the prediction model of adverse outcomes based on the proposed learning strategy training can find about 95% of patients with potential adverse outcomes,and its performance in accuracy,recall and F1 measurement is better than the classic PU learning method,and it is practical and progressiveness.Finally,this thesis explores the challenges in mining and utilizing the knowledge contained in the complexity of medical data to improve the accuracy of adverse outcome prediction models from three processes: feature selection,information aggregation,and model training.It simultaneously considered the complexities that must be solved when constructing data-driven adverse outcome prediction models using medical data.Mining and utilizing the knowledge and information from the complex medical data effectively improved the accuracy of adverse outcome prediction models.The prediction model can improve doctors’ decisionmaking,especially in excluding the risk of adverse outcomes for some patients with high accuracy which helps a lot in reducing the work pressure of doctors,and can avoid misdiagnosis caused by doctors’ fatigue,stress,and other reasons.It also has positive significance in improving the quality and efficiency of doctors’ decision-making and improving patients’ treatment experience.