JMIR Publications

ABSTRACT

Background:

Early detection of mental disorders symptoms can lead to prompt and correct diagnosis and reduce the recurrence of these symptoms and associated disabilities. Creating a tool to detect early symptoms is crucial for taking the necessary measures to prevent major onsets of mental diseases. Early indicators of mental health disorders can be detected through changes in daily activity patterns, which activity trackers and speech data can capture.

Objective:

We aim to compare the accuracy of personalized machine-learning models with population-level models and evaluate the robustness of these models across various languages. Additionally, investigate the significance of speech data when the user reads a neutral text versus reflecting on their daily life experiences while predicting mental disorders.

Methods:

Our research is based on longitudinal data from each participant. Hence, we designed the collection process to capture several data points in time that could aid machine learning algorithms to capture patterns of mental disorder symptoms better. This research uses machine learning models to predict the levels of anxiety, stress, and depression in participants based on data collected from wearable devices and voice recordings. The data includes daily activity from smartwatches and voice data collected through text reading and free-form speech.

Results:

The study is ongoing, and data are collected from at least 50 participants attending two major universities, and the data collection complies with ethical and personal data privacy requirements.

Conclusions:

The study aims to advance personalized machine learning for mental health, generate a dataset to predict DASS21 results, and deploy a framework to detect onsets of depression, anxiety, and stress, with the final goal of developing a non-invasive and objective method for collecting mental health data and prompt detection of mental disorder symptoms.