A review of automated clinical depression diagnosis tools

By Dr. Priyom Bose, Ph.D.Reviewed by Benedette Cuffari, M.Sc.Nov 23 2023

A recent npj Mental Health Research article reviewed current research on how the latest computational methods that utilize facial, acoustic, and semantic features are being deployed to predict major depressive disorders.

Study: A systematic review on automated clinical depression diagnosis. Image Credit: meeboonstudio / Shutterstock.com

Background

Approximately 280 million people throughout the world are current estimated to suffer from depression. Traditionally, semi-structured interviews have been used to assess depression; however, this method is subjective and susceptible to bias and social stigma.

Additionally, the social stigma and shortage of trained professionals make it difficult for patients to seek help, especially in low- and middle-income countries. Thus, an automated depression assessment tool could offer the objective diagnosis of depression.

Improving the quality of data related to mental health is one way of improving the objectivity in assessing depression. Analysis of this data using sensors and machine learning methods could also add immense value.

For example, biomarkers of depression could be detected using sensors that measure other biological signals such as heart rate. Machine learning can allow practitioners to make correct diagnoses, detect high risk individuals, and monitor symptoms over time. 

About the study

The current review article surveys research on how computational methods are being used to identify depressive disorders. The Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines were used to ensure a thorough and rigorous evaluation of the findings documented over the last ten years. Google Scholar was used as the search engine. 

Articles that lacked detailed results or comprehensive methodology were excluded. Additionally, papers related to autism, Parkinson’s disease (PD), and substance overdose were excluded.

Specific information was synthesized from the articles including age range, mental disorders, best metrics, number of subjects, predictive features, and type of validation.

The key aims of the present study are to review and summarize results from the latest research, identify notable differences in semantic, acoustic, and visual features, correlate these to depression symptoms, and identify challenges associated with automated depression assessment.

Key findings

A total of 264 articles were included in this review. The automated speech feature extraction was used by most studies to assess major depressive disorder.

Most models predicted that performance may not be generalizable, as this is dependent on various factors including sample size and feature engineering. With regard to acoustic features, patients with mental disorders often exhibited monotonous speech. Shimmer and jitter were associated with depression severity. 

A challenge while developing automated models is the lack of accounting for comorbidities, which was an issue not reported by several previous studies. In the future, models should be trained to be inclusive and exclusive of comorbidities and compared to provide a better understanding of the model’s accuracy. 

If a model is trained well, it can accurately detect mental health issues in a randomly selected individual, irrespective of the environment in which the individual is interviewed, age, and the use of a different accent or language. However, most previous studies have identified disorders among new subjects in a similar environment. 

Reproducibility of findings is a key issue to progress research in this area. While clinical data may not always be shared, the code and transcripts for training should be available to foster collaboration.

In the future, models should be made more robust and generalizable, as cross-cultural generalization is a key area for future research. Furthermore, to ensure that findings of the automated system are ethically used, researchers should provide thorough documentation on how collected data will be used. 

Conclusions

The need for reproducibility of findings was stressed, as comparing results aids researchers in gauging model performance and estimating overfitting. Openness should also be demonstrated in sharing data and code. More research on multiple datasets could improve the robustness and generalizability of these models.

The future of automated mental health evaluations and treatment is promising as more multimodality features are used to discipline machine-learning models. As this is in line with the principles of personalized and preventive diagnosis, it could entail favorable outcomes for patients with mental health disorders.