Modeling Temporality with Natural Language Processing to Predict Readmission Risk of Patients with Psychosis

点击翻译按钮获取中文摘要

Project Summary A substantial proportion of psychiatric inpatients are readmitted within 30 days of discharge. Readmissions not only are disruptive but also cause enormous economic burden for patients and families, and are a key driver of rising healthcare costs. Reducing and predicting unplanned readmission are therefore major unmet needs of psychiatric care. Developing machine learning (ML)-based natural language processing (NLP) prediction tools using electronic health records (EHRs) is a key priority as such tools could not only be used to help target the delivery of resource-intensive interventions to those patients at greatest risk, but also reduce psychiatric health- care costs. A key aspect in building effective risk predictive models is the modeling of temporal structure in the narratives. Information about the historical and present health states and timing of events (e.g., substance use start/stop timing, recent fluctuations in suicidality or symptoms), may play a key role in predicting readmission risk. Natural language annotation (i.e., tagging text such as events, symptoms, and anchoring them on a timeline) is a key step for training ML classifiers. No psychiatry-specific resources or guidelines exist for the modeling of temporality in clinical text, and as a result no robust scalable and explainable ML predictive models incorporating temporal information have been developed. We propose to deliver a psychiatric specific temporal relation annotation scheme, build open-source tools for extracting temporal information, and develop readmission prediction models for psychiatric patients. Aim 1 is a data resource creation aim in which we create a large repository of psychiatric text for building our readmission classifier, de-identify a subset of that data to allow for sharing with the research community, and create a layer of temporal annotations for that subset. In Aim 2, we extract temporal information from the data in the repository to create temporal graphs, and apply graph neural networks to these graphs to extract features for predicting 30-day readmission risk. In Aim 3 we build and evaluate multiple versions of 30-day readmission risk classifiers, and feedback performance to Aim 2 to improve temporal modeling. We develop unsupervised clustering on top of our classifiers to discover patient sub-groups. We include practical evaluations including a comparison to human experts and an evaluation of model performance on simulated future data. The study brings together a team experienced in psychiatric phenotyping and application of EHRs, and a team active in developing cutting- edge methods in ML for natural language data. This work will serve as the foundation for future translational studies, including implementing readmission classifiers into clinical workflows and clinical trials of interventions to reduce readmission risk.