Abstract

This paper represents our research to (i) propose an artificial intelligence, AI-based solution to identify depression and (ii) investigate our psychiatric knowledge. Concerning the first objective, we collected and annotated a new audio data set, and scrutinized the performance of eight regression approaches. Our studies showed that k-nearest neighbor and random forest form the group having the most acceptable results. Regarding our second objective, we determined the importance of the features of our best model using the SHapley Additive exPlanations approach: our findings showed that the fourth Mel-frequency cepstral coefficients, harmonic difference, and shimmer are the most important features.

中文翻译：

---

一种基于人工智能的解释解决方案，利用声学特征识别抑郁症严重程度症状

摘要

本文代表了我们的研究：(i) 提出一种基于人工智能的解决方案来识别抑郁症；(ii) 调查我们的精神病学知识。关于第一个目标，我们收集并注释了一个新的音频数据集，并仔细检查了八种回归方法的性能。我们的研究表明，k 最近邻和随机森林构成了具有最可接受结果的组。关于我们的第二个目标，我们使用 SHapley 加法解释方法确定了最佳模型特征的重要性：我们的研究结果表明，第四梅尔频率倒谱系数、谐波差和闪烁是最重要的特征。