Machine Learning and Personalized Prognosis for Depression Treatment

摘要 抑郁症的治疗对大约50%-60%的接受治疗的患者有效，但 在治疗开始之前，成功应答的概率通常是未知的。其结果是，抑郁症 治疗通常以反复试验的方式进行，直到获得满意的反应。我们的 目的是通过应用集成机器学习技术来提供个性化的预测 多个弱预测因子的新颖、非线性组合，共同产生准确的 治疗结果。这种统计方法同时考虑了许多预测变量，并且 迭代地构建复杂的预测模型，该模型通常显著优于传统统计模型 方法：研究方法。目标1是应用随机梯度增强决策树来预测西酞普兰的疗效。 STAR*D临床试验的档案数据。在初步分析中，我们随机选择了1223名患者 训练该模型和另外407名患者独立测试该模型(75：25)，并进行调整 通过交叉验证选择的参数，以将原木损失降至最低。对独立测试的结果预测 样本优于无信息率(p&lt；0.001)，总体预测准确率为66%。 尽管这一水平的预测明显好于无信息模型，但我们计划改进 通过以下方式预测模型：1)添加捕捉并发(非 研究)药物使用和2)基于早期反应迹象的更新模型预测。目标2是使用一个 类似的机器学习方法来检查抑郁症患者对基于互联网的CBT的反应。基于互联网的 抑郁症的治疗越来越受欢迎，提供了有效的医疗保健途径，减少了治疗 成本，并有很好的证据证明治疗效果。重要的是，我们有一个大型数据集(N=1,013 以开发基于患者属性预测治疗反应的治疗匹配算法。 研究影响：该项目的总体目标是使用机器学习方法来开发治疗方法 匹配算法。从长远来看，我们可以设想一个系统，根据许多因素对患者进行评估 重要的预测变量，并提供个人化的治疗成功概率。这些概率 然后用于指导治疗选择或在预测到不良反应的情况下修改当前治疗。 开发算法，成功预测一种特定形式的治疗是否可能成功 对于拥有一组给定属性的患者来说，这将是朝着高效和个性化迈出的巨大一步 抑郁症治疗。

Abstract Depression treatment is effective for approximately 50-60% of patients who receive treatment, but the probability of a successful response is typically unknown before treatment begins. As a result, depression treatment is routinely delivered in a trial-and-error fashion until a satisfactory response is achieved. Our objective is to provide a personalized prognosis by applying ensemble machine learning techniques to discover novel, non-linear combinations of multiple weak predictors that collectively yield accurate predictions of treatment outcome. This statistical approach considers many prediction variables simultaneously and iteratively constructs a complex prediction model that often dramatically outperforms traditional statistical methods. Aim 1 is to apply stochastic gradient boosted decision trees to predict response to citalopram using archival data from the STAR*D clinical trial. In preliminary analyses, we randomly selected 1223 patients to train the model and another 407 patients to independently test the model (a 75-25 split), with tuning parameters selected by cross-validation to minimize log-loss. The resulting prediction on the independent test sample was superior to the no-information rate (p < 0.001), with an overall predictive accuracy of 66%. Although this level of prediction is significantly better than a no information model, we plan to improve the model's prognostication by 1) adding features that capture the “pharmacological noise” of concurrent (non- study) medication use and 2) updating model predictions based on early signs of response. Aim 2 is to use a similar machine learning approach to examine response to internet-based CBT for depression. Internet-based treatments for depression are growing in popularity, provide efficient access to health care, reduce treatment costs, and have good evidence for treatment efficacy. Importantly, we have a large dataset (N = 1,013) within which to develop treatment-matching algorithms that predict treatment response based on patient attributes. Study Impact: The overarching goal of this project is to use machine learning methods to develop treatment matching algorithms. In the long term, we can envision a system that evaluates a patient on a number of important predictor variables and provides a personalized probability of treatment success. These probabilities would then be used to guide treatment selection or modify current treatment if a poor response is predicted. Developing algorithms that successfully predict whether a particular form of treatment is likely to be successful for a patient with a given set of attributes would be a tremendous step towards efficient and personalized depression treatment.