What differentiates a trauma from an event that is merely upsetting? Wildly different definitions of trauma have been used in both formal (psychiatric) and informal (cultural, colloquial) settings. Yet there is a dearth of empirical work examining the features of events that individuals use to define an event as a ‘trauma’. First, a group of qualitative coders classified features (e.g., actual physical injury, loss of possessions) of 600 event descriptions (e.g., “was verbally harassed by a boss”, “watched a video of an adult being shot and killed”). Next, across two studies, machine learning was used to predict whether individuals rated event descriptions as ‘trauma’ or ‘traumatic’ in over 100,000 judgment tasks. In Study 1, examining continuous ratings from ‘not at all traumatic’ to ‘extremely traumatic’, a cross-validated LASSO regression with polynomial features provided the best out-of-sample predictions (r² = 0.76), outperforming ridge regression, support vector regression, and linear regression. In Study 2, using binary judgments, a random forest model accurately predicted out-of-sample individual responses (AUC = 0.96), outperforming a neural network and an AdaBoost ensemble classifier. The most important event features across the two studies were actual death, threat of death, and the presence of a human perpetrator. The most important human features in predicting judgments were political orientation and gender.

什么区别创伤与仅仅令人不安的事件？在正式（精神病学）和非正式（文化、口语）环境中，对创伤的定义截然不同。然而，缺乏实证工作来检验个人用来将事件定义为“创伤”的事件特征。首先，一组定性编码员对 600 个事件描述的特征（例如，实际身体伤害、财产损失）进行了分类（例如，“被老板口头骚扰”、“观看了一个成年人被枪杀的视频”）。接下来，在两项研究中，机器学习用于预测个人在超过 100,000 个判断任务中将事件描述评为“创伤”还是“创伤”。在研究 1 中，检查从“完全没有创伤”到“非常有创伤”的连续评级，r ²  = 0.76)，优于岭回归、支持向量回归和线性回归。在研究 2 中，使用二元判断，随机森林模型准确预测样本外个体响应 ( AUC  = 0.96)，优于神经网络和 AdaBoost 集成分类器。两项研究中最重要的事件特征是实际死亡、死亡威胁和人类肇事者的存在。预测判断的最重要的人类特征是政治取向和性别。