Research On Personalized Hospital Readmission Prediction

Hospital readmission prediction employs patients’ historical medical records to evaluate the readmission risk of patients upon discharge.Frequent unplanned readmission will consume a large amount of medical resources and cause a great burden on patients.Therefore,it is crucial to use highly reliable predictive models to assess the readmission risk of patients in a timely manner upon discharge.Hospital readmission prediction can provide data support for the formulation and adjustment of medical insurance policies,also help reduce some unnecessary medical insurance expenses;at the same time,it can help medical institutions allocate resources in a more efficient way and improve the utilization rate of hospitalization services.In addition,it can assist doctors to provide patients with timely evaluation guidance to reduce the burden on individual patients through early intervention.Personalized hospital readmission prediction is to assess the readmission risk of patients by considering their personalized features,based on the idea of dialectical treatment.Many related studies utilize machine learning and data mining methods,taking all available patients’ historical admission records as the input,to learn a generic predictive model that can predict the probability of future rcadmission for all patients.They mainly focus on population-level healthcare,that is,large-scale statistics based on the experience or rough classification.These methods mainly pursue the generality and universality of prediction,which aim at learning a general predictive model that can be applied to multiple diseases and populations.However,due to the natural differences in geological regions,medical technologies,economic levels,individual traits,etc.,the readmission rate varies for the same disease among different patients,it is difficult to achieve population-level healthcare.Although the way of learning a generic predictive model for all patients based on all available data is able to capture the overall trend for the entire patient population,it may miss or abandon some important information of individual patients due to considerations of the whole situation.Thus,many scholars begin to propose personalized predictive methods in view of the differences between individual patients.The population-level healthcare is gradually changing to personalized healthcare.They usually build an independent model for each target patient,to predict the potential risk in patients and assist doctors in formulating the treatment plan that can produce the best effect for patients.Although this way can retain the specificity of individual patients,it is difficult to capture the common information of the entire patient population.In addition,many patients have very similar readmission patterns due to the same complications or receiving similar treatments,many independent models may be interrelated in nature.They may have similar model parameters that can be shared with each other.If each model has to be trained independently from scratch,not only the model parameters are large in scale and the learning process is slow,but repeated training will also waste a lot of computing resources.Thus,it is a difficult research problem of effectively combining and using the common information and the personalized information,to mine patients’ readmission modes in the process of prediction.Therefore,in this paper,we conduct a research on personalized hospital readmission prediction.We design a multi-cluster joint prediction approach to achieve a good balance in both the common and the personalized information.We use a patient clustering strategy to divide patients into multiple clusters based on their historical medical records from multiple angles,to preserve the relevant information about patients with highly similar clinical behaviors and avoid repeated training of similar patients’ models.In addition,we establish a specific readmission predictive model for each cluster,and adopt the multi-task learning approach to learn multiple models jointly,in order to perform multi-cluster joint prediction.Through the partital parameter sharing of multiple patient clusters,the common information among the entire patient population can be captured.At the same time,the personalized information of each cluster of similar patients can be preserved in the case where partial parameters of multiple clusters are independent of each other.In this way,the common information about the entire patient population and the personalized information of each specific cluster can be combined and utilized in a balanced manner,thereby achieving a more precise and efficient predictionIn this paper,we evaluate the performance of our proposed approach on the real-world dataset of patients with Coronary Heart Disease(CHD)from multiple hospitals in a city in Shandong province.We extract medical records of each CHD patient in the year of the most frequent admission to hospital and partition the entire one-year time interval into a nine-month observation window and a three-month prediction window.We use their admission records over the preceding nine months to predict whether the patient will be readmitted for CHD within three months upon discharge.We carried out experimental evaluation and analysis from multiple perspectives.The results of multiple evaluation indicators show that our proposed approach significantly outperforms the widely used comparisons in hospital readmission prediction.In the process of multi-cluster joint learning,some parameters can be reused between multiple clusters through the sharing of some parameters at the bottom of the network,thereby speeding up the training process and saving the training overhead Meanwhile,personalized information of specific clusters can be preserved in the case where some model parameters are independent,which help improve the overall readmission prediction performance.