Using Audio Data to Facilitate Depression Risk Assessment in Primary Health Care

Adam Valen Levinson, Abhay Goyal, Roger Ho Chun Man, Roy Ka-Wei Lee, Koustuv Saha, Nimay Parekh, Frederick L. Altice, Lam Yin Cheung, Munmun De Choudhury, Navin KumararXiv 2023 · DOI · arXiv

Abstract

Depression remains a significant challenge in primary health care (PHC), where diagnostic accuracy is often compromised by limited resources and high patient volumes. This paper explores a "lightweight" screening approach using only audio data from telehealth sessions. By applying an automated machine learning (AutoML) pipeline to spectral features extracted from patient interviews, we achieved an F1-score of 0.96. These results suggest that audio-only analysis can provide a high-accuracy, low-bandwidth, and privacy-preserving alternative to traditional multimodal diagnostic tools, particularly for underserved populations.

Introduction

Depression is a leading cause of disability worldwide and a major burden on primary health care systems. Despite its prevalence, approximately 25% of depression diagnoses in PHC settings are estimated to be incorrect. The shift toward telehealth during the COVID-19 pandemic introduced new opportunities for AI-driven diagnostics; however, many existing tools rely on multimodal data, including video and text. In rural or low-income areas, high bandwidth requirements for video and privacy concerns often limit the feasibility of these tools. This study investigates the efficacy of an audio-only screening tool to facilitate more equitable and accessible mental health assessments.

Methodology

The researchers conducted a pilot study involving 24 participants: 12 with a formal diagnosis of major depressive disorder (MDD) and 12 healthy controls. Data collection involved 30-minute semi-structured Zoom interviews covering demographic information, reading tasks, and standard mental health screenings (e.g., PHQ-9).

Technical Pipeline

  1. Preprocessing: Audio recordings were processed using PyAudioAnalysis to remove silences and segmented into 4-second clips, resulting in a dataset of 3,969 segments.
  2. Feature Extraction: Sixteen acoustic features were extracted per segment using the librosa library. Key features included Mel-frequency cepstral coefficients (MFCCs), Spectral Centroid, Spectral Complexity, and Zero-Crossing Rate.
  3. Model Optimization: The Tree-based Pipeline Optimization Tool (TPOT) was employed to automate the selection of the best machine learning pipeline. TPOT utilized genetic programming to identify an optimal combination of preprocessing and classification algorithms.

Results and Key Insights

The AutoML process converged on a K-Nearest Neighbors (KNN) classifier (k=3) paired with a StandardScaler. Analysis of the extracted features revealed significant differences between the depressed and control groups, particularly in spectral complexity, which serves as a proxy for the richness and variation in a person's voice.

Table 1: Comparison of acoustic feature metrics between depressed and non-depressed groups.

The model demonstrated exceptional performance on the test set, indicating that spectral features of the voice can serve as robust biological markers for depression.

Table 2: Performance metrics for the depression classification model.

As shown in Table 2, the model achieved a precision of 0.98 and a recall of 0.93, yielding an overall F1-score of 0.96. These metrics suggest the tool is highly effective at identifying depression risk while maintaining a very low false-positive rate.

Discussion

The success of this audio-only approach has significant implications for health equity. By eliminating the need for high-speed video, diagnostic support becomes accessible to patients with limited internet connectivity. Furthermore, the use of AutoML demonstrated that simpler, non-parametric models like KNN can outperform complex deep learning architectures when trained on high-quality, specialized feature sets.

Limitations and Future Directions

While promising, this proof-of-concept study is limited by its small sample size (N=24). Future research should focus on validating these findings with larger, more diverse cohorts. Additionally, moving from binary classification to predicting symptom severity and implementing privacy-preserving technologies, such as homomorphic encryption, will be critical for clinical integration.

Conclusion

This study demonstrates that audio-only data is a viable and powerful medium for depression risk assessment. The high accuracy achieved through an AutoML pipeline suggests that such tools could significantly improve diagnostic accuracy in primary care, providing a scalable and equitable solution for mental health screening.

Summary updated

Abstract

Telehealth is a valuable tool for primary health care (PHC), where depression is a common condition. PHC is the first point of contact for most people with depression, but about 25% of diagnoses made by PHC physicians are inaccurate. Many other barriers also hinder depression detection and treatment in PHC. Artificial intelligence (AI) may help reduce depression misdiagnosis in PHC and improve overall diagnosis and treatment outcomes. Telehealth consultations often have video issues, such as poor connectivity or dropped calls. Audio-only telehealth is often more practical for lower-income patients who may lack stable internet connections. Thus, our study focused on using audio data to predict depression risk. The objectives were to: 1) Collect audio data from 24 people (12 with depression and 12 without mental health or major health condition diagnoses); 2) Build a machine learning model to predict depression risk. TPOT, an autoML tool, was used to select the best machine learning algorithm, which was the K-nearest neighbors classifier. The selected model had high performance in classifying depression risk (Precision: 0.98, Recall: 0.93, F1-Score: 0.96). These findings may lead to a range of tools to help screen for and treat depression. By developing tools to detect depression risk, patients can be routed to AI-driven chatbots for initial screenings. Partnerships with a range of stakeholders are crucial to implementing these solutions. Moreover, ethical considerations, especially around data privacy and potential biases in AI models, need to be at the forefront of any AI-driven intervention in mental health care.
BibTeX
@misc{levinson2023using,
  author = {Adam Valen Levinson and Abhay Goyal and Roger Ho Chun Man and Roy Ka-Wei Lee and Koustuv Saha and Nimay Parekh and Frederick L. Altice and Lam Yin Cheung and Munmun De Choudhury and Navin Kumar},
  title = {{Using Audio Data to Facilitate Depression Risk Assessment in Primary Health Care}},
  year = {2023},
  eprint = {2310.10928},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.10928},
}
Figure