Depression is a highly prevalent and seriously debilitating disorder. Current assessment and treatment approaches alike are beset with substantial limitations. While empirical studies emphasize the importance of implementing standardized assessment tools in clinical practice, many practitioners refrain from using them due to various barriers, such as time constraints. Therefore, it is crucial to develop tools that enable efficient, convenient, and accurate depression assessments in clinical practice. Furthermore, although first-line treatments of depression, such as cognitive behavioral therapy or pharmacotherapy, have been found to be effective, a substantial number of non-responders and/or patients who experience symptom recurrence after treatment persists. To address the underlying mechanisms that impede effective treatment, it is necessary to develop more effective intervention targets. Previous research has focused on identifying markers of depression that could be used for precise assessment and targeted treatment of the disorder. Speech is a particularly compelling indicator of depression, as it encompasses cognitive planning, motor action, and sensory processing, each of which is hypothesized to be involved in the perpetuation of depressive symptomatology. Previous studies demonstrated that certain speech patterns are indicative of depression and can be automatically assessed using machine learning (ML). Furthermore, there is empirical evidence for the impact of bodily state modulations on cognitive and affective processes related to depression. This dissertation pursues two primary objectives: First, to advance depression assessment, and second, to enhance depression treatment. In terms of assessment, goals encompass the development, validation, and optimization of speech-based ML models for depression assessment (Studies I-V), along with the evaluation of linguistic speech content – specifically, absolutist word usage – as a potential indicator of depression (Study VI). In terms of treatment, the goal is to evaluate whether targeted voice modulations enhance the efficacy of depression treatment (Study VII). Study I utilized n = 550 speech recordings from N = 267 participants with current or previous depression to develop an ML-based depression classification model with a state-of-the-art ML approach. While the classification accuracy of the ML model was significantly better than chance, it was lower than that based on cut-off scores from established depression scales. Studies II and III compared several different types of speech recordings from n = 48 individuals with clinical depression, n = 48 individuals with subclinical depression, and n = 48 individuals with no depression to train ML-based depression classification models. Results from Study II indicated that speech with depression-associated content did not improve ML-based classifications compared to neutral content. Study III demonstrated that ML models achieved higher depression classification performances when trained on speech recordings that incorporated instructions to maximize conviction in the voice, compared to speech recordings without such voice modulation instructions. Using longitudinal speech data from the same sample as in Studies II and III, Studies IV and V evaluated the efficacy of personalizing ML-based monitoring of depressed mood. In Study IV, personalization was achieved through speaker-dependent data modeling, which requires collecting a speaker’s sample data when applying the ML model. In contrast, Study V introduced an efficient zero-shot personalization approach that enabled model personalization without the need for speaker data during the application of the ML model. Results from both studies demonstrated that personalized ML models outperformed non-personalized ML models in monitoring depressed mood. Based on the same speech recordings, Study VI evaluated participants’ absolutist word usage as an indicator of depression. The findings suggest that individuals with clinical or subclinical depression used more absolutist words in their speech compared to non-depressed individuals. Absolutist word usage also served as an indicator of maladaptive emotion regulation, as it was associated with lower momentary depressed mood but more severe overall depressive symptomatology. Using the same sample used in Studies II-VI, Study VII examined whether instructions to maximize conviction in the voice could enhance the efficacy of reading aloud anti-depressive self-statements to reduce depressed mood. The findings suggest that participants with depression who were instructed to apply such voice modulations experienced greater reductions in depressed mood compared to those who read the self-statements without following such instructions. This difference was not found for participants with subclinical depression or those with no depression. In conclusion, the findings from the studies reported on in this dissertation confirmed theoretical assumptions regarding the relationship between speech and depression, as well as how this relationship contributes to the perpetuation of depressive symptomatology. That said, the findings also indicate that current ML-based depression assessment methods using speech require further optimization and validation before clinical implementation. In particular, depression assessments may be optimized by training ML models on speech recordings that incorporate targeted voice modulations, by personalizing ML models, and/or by analyzing both paralinguistic and linguistic features of speech. In terms of enhancing depression treatment, the findings suggest that interventions should extend their focus beyond cognitive factors to also target paralinguistic speech characteristics. Future research should continue to optimize and validate ML-based depression assessments with a specific focus on clinical implementation. To that end, interdisciplinary collaboration is needed to balance the urgent requirements of clinical care with the potential risks and limitations associated with this approach. Additionally, future studies should explore the feasibility of integrating speech analysis into depression treatment to better inform, personalize, and enhance interventions.
Jonathan Felix Bauer (Thu,) studied this question.