Manifestation of depression in speech overlaps with characteristics used to represent and recognize speaker identity

Automated Speech Analysis for Risk Detection of Depression, Anxiety, Insomnia, and Fatigue: Algorithm Development and Validation Study

BACKGROUND

While speech analysis holds promise for mental health assessment, research often focuses on single symptoms, despite symptom co-occurrences and interactions. In addition, predictive models in mental health do not properly assess the limitations of speech-based systems, such as uncertainty, or fairness for a safe clinical deployment.

OBJECTIVE

We investigated the predictive potential of mobile-collected speech data for detecting and estimating depression, anxiety, fatigue, and insomnia, focusing on other factors than mere accuracy, in the general population.

METHODS

We included 865 healthy adults and recorded their answers regarding their perceived mental and sleep states. We asked how they felt and if they had slept well lately. Clinically validated questionnaires measuring depression, anxiety, insomnia, and fatigue severity were also used. We developed a novel speech and machine learning pipeline involving voice activity detection, feature extraction, and model training. We automatically modeled speech with pretrained deep learning models that were pretrained on a large, open, and free database, and we selected the best one on the validation set. Based on the best speech modeling approach, clinical threshold detection, individual score prediction, model uncertainty estimation, and performance fairness across demographics (age, sex, and education) were evaluated. We used a train-validation-test split for all evaluations: to develop our models, select the best ones, and assess the generalizability of held-out data.

RESULTS

The best model was Whisper M with a max pooling and oversampling method. Our methods achieved good detection performance for all symptoms, depression (Patient Health Questionnaire-9: area under the curve [AUC]=0.76; F1-score=0.49 and Beck Depression Inventory: AUC=0.78; F1-score=0.65), anxiety (Generalized Anxiety Disorder 7-item scale: AUC=0.77; F1-score=0.50), insomnia (Athens Insomnia Scale: AUC=0.73; F1-score=0.62), and fatigue (Multidimensional Fatigue Inventory total score: AUC=0.68; F1-score=0.88). The system performed well when it needed to abstain from making predictions, as demonstrated by low abstention rates in depression detection with the Beck Depression Inventory and fatigue, with risk-coverage AUCs below 0.4. Individual symptom scores were accurately predicted (correlations were all significant with Pearson strengths between 0.31 and 0.49). Fairness analysis revealed that models were consistent for sex (average disparity ratio [DR] 0.86, SD 0.13), to a lesser extent for education level (average DR 0.47, SD 0.30), and worse for age groups (average DR 0.33, SD 0.30).

CONCLUSIONS

This study demonstrates the potential of speech-based systems for multifaceted mental health assessment in the general population, not only for detecting clinical thresholds but also for estimating their severity. Addressing fairness and incorporating uncertainty estimation with selective classification are key contributions that can enhance the clinical utility and responsible implementation of such systems.

Deconstructing demographic bias in speech-based machine learning models for digital health

Introduction

Machine learning (ML) algorithms have been heralded as promising solutions to the realization of assistive systems in digital healthcare, due to their ability to detect fine-grain patterns that are not easily perceived by humans. Yet, ML algorithms have also been critiqued for treating individuals differently based on their demography, thus propagating existing disparities. This paper explores gender and race bias in speech-based ML algorithms that detect behavioral and mental health outcomes.

Methods

This paper examines potential sources of bias in the data used to train the ML, encompassing acoustic features extracted from speech signals and associated labels, as well as in the ML decisions. The paper further examines approaches to reduce existing bias via using the features that are the least informative of one's demographic information as the ML input, and transforming the feature space in an adversarial manner to diminish the evidence of the demographic information while retaining information about the focal behavioral and mental health state.

Results

Results are presented in two domains, the first pertaining to gender and race bias when estimating levels of anxiety, and the second pertaining to gender bias in detecting depression. Findings indicate the presence of statistically significant differences in both acoustic features and labels among demographic groups, as well as differential ML performance among groups. The statistically significant differences present in the label space are partially preserved in the ML decisions. Although variations in ML performance across demographic groups were noted, results are mixed regarding the models' ability to accurately estimate healthcare outcomes for the sensitive groups.

Discussion

These findings underscore the necessity for careful and thoughtful design in developing ML models that are capable of maintaining crucial aspects of the data and perform effectively across all populations in digital healthcare applications.

Enhancing accuracy and privacy in speech-based depression detection through speaker disentanglement

Speech signals are valuable biomarkers for assessing an individual's mental health, including identifying Major Depressive Disorder (MDD) automatically. A frequently used approach in this regard is to employ features related to speaker identity, such as speaker-embeddings. However, over-reliance on speaker identity features in mental health screening systems can compromise patient privacy. Moreover, some aspects of speaker identity may not be relevant for depression detection and could serve as a bias factor that hampers system performance. To overcome these limitations, we propose disentangling speaker-identity information from depression-related information. Specifically, we present four distinct disentanglement methods to achieve this - adversarial speaker identification (SID)-loss maximization (ADV), SID-loss equalization with variance (LEV), SID-loss equalization using Cross-Entropy (LECE) and SID-loss equalization using KL divergence (LEKLD). Our experiments, which incorporated diverse input features and model architectures, have yielded improved F1 scores for MDD detection and voice-privacy attributes, as quantified by Gain in Voice Distinctiveness G V   D and De-Identification Scores (DeID). On the DAIC-WOZ dataset (English), LECE using ComparE16 features results in the best F1-Scores of 80% which represents the audio-only SOTA depression detection F1-Score along with a G V   D of -1.1 dB and a DeID of 85%. On the EATD dataset (Mandarin), ADV using raw-audio signal achieves an F1-Score of 72.38% surpassing multi-modal SOTA along with a G V   D of -0.89 dB dB and a DeID of 51.21%. By reducing the dependence on speaker-identity-related features, our method offers a promising direction for speech-based depression detection that preserves patient privacy.

Improving speech depression detection using transfer learning with wav2vec 2.0 in low-resource environments

Depression, a pervasive global mental disorder, profoundly impacts daily lives. Despite numerous deep learning studies focused on depression detection through speech analysis, the shortage of annotated bulk samples hampers the development of effective models. In response to this challenge, our research introduces a transfer learning approach for detecting depression in speech, aiming to overcome constraints imposed by limited resources. In the context of feature representation, we obtain depression-related features by fine-tuning wav2vec 2.0. By integrating 1D-CNN and attention pooling structures, we generate advanced features at the segment level, thereby enhancing the model's capability to capture temporal relationships within audio frames. In the realm of prediction results, we integrate LSTM and self-attention mechanisms. This incorporation assigns greater weights to segments associated with depression, thereby augmenting the model's discernment of depression-related information. The experimental results indicate that our model has achieved impressive F1 scores, reaching 79% on the DAIC-WOZ dataset and 90.53% on the CMDC dataset. It outperforms recent baseline models in the field of speech-based depression detection. This provides a promising solution for effective depression detection in low-resource environments.