Leveraging deep learning and clinical notes for surveillance and prediction of intentional self-harm and suicide

项目总结： 自杀是美国的主要死因之一，自杀人数超过4.7万人 每年有自杀身亡的人。识别有自杀危险的个人是一种 为全面预防自杀迈出了重要的一步。尽管进行了广泛的研究 关于故意自残和自杀的危险因素，自杀的前瞻性预测仍然是一种 预测能力差的艰巨任务。最近的研究表明，新的机器学习 应用于电子健康档案(EHR)的方法显示了良好的结果。然而，更多 深度学习等高级计算方法在 这一领域，特别是在临床病历文本分类的先进方法领域。我们的 这个项目的目的是改善自杀行为的表型，并预测 通过将死亡率数据与EHR数据和 利用最先进的自然语言计算方法。我们还将调查 模型的解释能力和可解释性的方法，以提高未来的采用率 临床医生。我们将通过检查可重复性和通用性来验证我们的模型 两个卫生系统在两个地点使用类似的数据。

PROJECT SUMMARY: Suicide is one of the leading causes of death in the United States, with more than 47,000 individuals dying by suicide each year. The identification of individuals at risk for suicide is an important step for a comprehensive approach to suicide prevention. Despite extensive research on risk factors for intentional self-harm and suicide, prospective prediction of suicide remains a difficult task with poor predictive power. Recent studies suggest that new machine learning methods applied to electronic health records (EHR) show promising results. However, more advanced computational approaches such as deep learning, have not been fully leveraged in this field, especially in the area of advanced methods for text classification of clinical notes. Our aims in this project, are to improve the phenotyping of suicidal behavior, and the prediction of future suicidal behavior and suicide deaths by integrating mortality data with EHR data and leveraging state-of-the-art natural language computational approaches. We will also investigate methods for explain ability and interpretability of the models to improve future adoption by clinicians. We will validate our models by examining reproducibility and generalizability across two health systems using similar data at both sites.