Multiverse analysis of machine learning: classification between groups defined by suicidal ideation screening status...

Area of Science:

• Computational psychiatry
• Speech signal processing
• Machine learning in mental health

Background:

• Suicide is a critical public health issue among university students.
• Current assessment methods for suicidal ideation are subjective and lack continuous monitoring capabilities.
• Objective, non-invasive correlates of suicidal ideation are needed, with speech signals showing promise.

Purpose of the Study:

• To explore objective acoustic correlates of suicidal ideation using a comprehensive multiverse analysis of speech data.
• To identify robust and generalizable acoustic features for distinguishing students with and without suicidal ideation.
• To assess the impact of analytical choices on the predictive performance of speech-based models.

Main Methods:

• A multiverse analysis of 1,764 distinct analytical pipelines was performed on speech data from 96 Chinese university students.
• Participants included individuals screening positive and negative for suicidal ideation, matched by controls.
• Pipelines varied in preprocessing, feature extraction (e.g., F0, MFCCs), dimensionality reduction, and machine learning models.

Main Results:

• Predictive performance varied significantly across analytical pipelines (AUC 0.508–0.856).
• A core set of acoustic features, including fundamental frequency (F0) and Mel-frequency cepstral coefficients (MFCCs), demonstrated robust differences between groups.
• These core features were significant in 98.8% of eligible specifications, indicating stability.

Conclusions:

• Speech-based prediction of suicidal ideation status is sensitive to analytical decisions.
• Discriminative acoustic features derived from machine learning are remarkably stable.
• Observed acoustic differences likely reflect a combination of suicidal ideation, depression, anxiety, and general distress.