Abstract

Mental disorders like depression represent 28% of global disability. It affects around 7.5% percent of global disability. Depression is a common disorder that affects the state of mind, normal activities, emotions, and produces sleep disorders. It is estimated that approximately 50% of depressive patients suffering from sleep disturbances. In this paper, a data mining process to classify depressive and not depressive episodes during nighttime is carried out based on a formal method of data mining called Knowledge Discovery in Databases (KDD). KDD guides the process of data mining with stages well established: Pre-KDD, Selection, Pre-processing, Transformation, Data Mining, Evaluation, and Post-KDD. The dataset used for the classification is the DEPRESJON dataset, which contains the motor activity of 23 unipolar and bipolar depressed patients and 32 healthy controls. The classification is carried out with two different approaches; a multivariate and univariate analysis to classify depressive and non-depressive episodes. For the multivariate analysis, the Random Forest algorithm is implemented with a model construct of 8 features and, results of the classification are; specificity equal to 0.9927 and sensitivity equal to 0.9991. The univariate analysis shows that the maximum in time of the activity is the most descriptive characteristic of the model with 0.908 in accuracy for the classification of depressive episodes.

中文翻译：

---

使用夜间数据对抑郁发作进行分类；多元和单变量分析

摘要

抑郁症等精神疾病占全球残疾的28％。它影响了全球约7.5％的残疾。抑郁症是一种常见的疾病，会影响心理状态，正常活动，情绪并产生睡眠障碍。据估计，约有50％的抑郁症患者患有睡眠障碍。本文基于一种称为“数据库中的知识发现”（KDD）的正式数据挖掘方法，对夜间的抑郁和非抑郁发作进行了分类的数据挖掘过程。KDD指导数据挖掘的过程具有完善的阶段：KDD之前，选择，预处理，转换，数据挖掘，评估和KDD后。用于分类的数据集是DEPRESJON数据集，其中包含23名单相和双相抑郁症患者的运动活动以及32位健康对照。分类是通过两种不同的方法进行的：多变量和单变量分析，对抑郁和非抑郁发作进行分类。对于多变量分析，采用具有8个特征的模型构造来实现随机森林算法，分类结果为：特异性等于0.9927，灵敏度等于0.9991。单变量分析表明，活动时间的最大值是模型的最具描述性的特征，对于抑郁发作的分类准确度为0.908。对于多变量分析，采用具有8个特征的模型构造来实现随机森林算法，分类结果为：特异性等于0.9927，灵敏度等于0.9991。单变量分析表明，活动时间的最大值是模型的最具描述性的特征，对于抑郁发作的分类准确度为0.908。对于多变量分析，采用具有8个特征的模型构造来实现随机森林算法，分类结果为：特异性等于0.9927，灵敏度等于0.9991。单变量分析表明，活动时间的最大值是模型的最具描述性的特征，对于抑郁发作的分类准确度为0.908。