An unsupervised machine learning approach using passive movement data to understand depression and schizophrenia.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Digital Health

Background:

• Schizophrenia and Major Depressive Disorder (MDD) are significant mental health burdens with overlapping symptoms.
• Understanding longitudinal behavioral changes, particularly movement and sleep abnormalities, is crucial but challenging with traditional methods.
• Wearable sensors offer a way to collect naturalistic, passively-obtained data that may capture transdiagnostic features.

Purpose of the Study:

• To investigate naturalistic behavioral differences between individuals with schizophrenia, MDD, and controls using wearable sensor data.
• To apply unsupervised machine learning to identify distinct behavioral phenotypes without relying on diagnostic labels.
• To explore the temporal dynamics of these behavioral markers.

Main Methods:

• Utilized minute-level actigraphy data from 23 individuals with schizophrenia, 22 with MDD, and 32 controls.
• Applied unsupervised machine learning clustering to week-long, unlabeled actigraphy data to identify behavioral patterns.
• Analyzed actigraphic data over time to pinpoint influential time points in the models.

Main Results:

• Identified distinct actigraphic phenotypes that align with existing diagnostic constructs.
• Found statistically significant differences between groups, with individuals experiencing depression exhibiting the highest behavioral variability.
• Unsupervised clustering successfully differentiated behavioral patterns across diagnostic groups.

Conclusions:

• Passively collected movement data, analyzed with unsupervised deep learning, shows potential for identifying naturalistic phenotypes in mental health disorders.
• This approach can aid in discriminating between conditions like MDD and schizophrenia.
• Highlights the utility of wearable sensors in characterizing complex psychopathology.