Language-based detection of depression with machine learning: systematic review and meta-analysis

Area of Science:

• Computational linguistics
• Psychiatry
• Artificial intelligence

Background:

• Early depression detection is crucial for effective intervention.
• Natural language processing (NLP) and machine learning (ML) show potential for automated depression detection from text.
• Existing evidence on the performance of NLP and ML for depression detection is limited.

Purpose of the Study:

• To systematically review and meta-analyze studies on NLP and ML for depression detection from spoken or written language.
• To quantify the performance of automated depression detection methods.
• To identify factors influencing the performance of these methods.

Main Methods:

• Systematic review and meta-analysis of studies identified through six electronic databases and additional sources.
• Quantitative synthesis of data from 123 eligible articles, with one representative result per dataset.
• Pooled analysis of accuracy, precision, recall, AUC, and balanced accuracy from 43 studies involving 40,983 text samples.

Main Results:

• Pooled accuracy was 0.80, with pooled precision of 0.78 and recall of 0.76.
• Subgroup analyses revealed significant variations based on language, text source, feature type, and classifier.
• Text source was the only significant predictor in meta-regressions, explaining 13.6% of the variance.

Conclusions:

• Automated depression detection from text demonstrates promising performance but exhibits substantial heterogeneity.
• Findings highlight both the potential and limitations of text-based depression detection.
• Methodological standardization and validation are essential before widespread clinical adoption.