此页面是公开的。
Cost and stigma keep many people from mental health assessments, and even people in professional care are assessed only intermittently. An arXiv paper asks whether speech could help: can a machine-learning model detect an anxiety disorder from a short audio journal? Working with data from Kintsugi Mindful Wellness, Inc., the authors found that audio-embedding and multi-modal models reached an AUROC of 0.68-0.69. The random baseline scored 0.50.
The data
The dataset has 2263 self-recorded audio journals. After each recording the user took the GAD-7 anxiety questionnaire, giving a raw score from 0 to 21 and a bucket of none, mild, moderate or severe. Most journals run around 1-2 minutes. No human transcripts exist, so the team transcribed the audio with Wav2Vec2, which produced about 244K words, 18K of them unique, and 108 words per journal on average. For the task, mild, moderate and severe were grouped as "anxiety" and None as "non-anxiety".
| Class | Train | Valid | Test | Total |
|---|---|---|---|---|
| No anxiety | 840 | 149 | 173 | 1162 |
| Anxiety | 790 | 139 | 166 | 1095 |
| Total | 1630 | 288 | 339 | 2257 |
Five approaches
- Hand-crafted audio features. The 88 GeMAPS acoustic features plus emotion and sentiment from the transcripts (90 in all), fed to a logistic regression model with L1 regularization.
- Transcript embeddings. Sentence-BERT embeddings of the transcripts, classified with a gradient boosting model.
- Transcript embeddings with sample weights. The same, weighting each example by its raw GAD-7 score.
- Wav2vec audio embeddings. 512-dimensional embeddings of each recording, classified with an SVM.
- Multi-modal model. Audio and text embeddings joined into a 1792-dimensional vector, classified with an SVM.
Figure from the paper, licensed CC BY 4.0; changed here: shown without its caption.
Results
| Model | Precision | Recall | F1 | AUROC |
|---|---|---|---|---|
| Random baseline | 0.50 | 0.48 | 0.49 | 0.50 |
| Audio features | 0.63 | 0.53 | 0.58 | 0.66 |
| Transcript features | 0.64 | 0.57 | 0.60 | 0.68 |
| Transcript features with sample weights | 0.61 | 0.55 | 0.58 | 0.59 |
| Wav2Vec features | 0.66 | 0.61 | 0.64 | 0.69 |
| Multi-modal model | 0.66 | 0.60 | 0.61 | 0.68 |
Sample weights made things worse, so the authors dropped them from later models. For the audio-feature model, an analysis with SHAP put loudness, F0, emotion and sentiment among the top features, which the authors say clinical literature links to anxiety.
Figure from the paper, licensed CC BY 4.0; changed here: shown without its caption.
The authors also note the small dataset invites overfitting. Next they want to train on depression detection, a larger dataset, and fine-tune on anxiety, because the two conditions are similar.
Sources
- arXiv paper "Detecting anxiety from short clips of free-form speech", https://arxiv.org/abs/2312.15272, licensed CC BY 4.0. Changed here: rewritten and condensed, two figures shown without their captions; rewritten in hubnx's own words.
- Authors of the original: Prabhat Agarwal, Akshat Jindal, Shreya Singh.
留言
