Probabilistic Textual Time Series Depression Detection

ACL ARR 2025 May Submission2165 Authors

Abstract: We propose PTTSD, a Probabilistic Textual Time Series Depression Detection framework for predicting PHQ-8 depression severity scores from utterance-level clinical interviews. PTTSD models both predictive means and calibrated uncertainty over time using Gaussian and Student's-t distributions, trained via negative log-likelihood losses. Our architecture combines bidirectional LSTMs with self-attention and residual connections to model textual sequences, and employs uncertainty-aware output heads for calibrated probabilistic predictions. On the E-DAIC dataset, PTTSD achieves state-of-the-art performance among text-only systems (MAE = 3.85, RMSE = 4.52), outperforming recent baselines. Extensive ablation and sensitivity studies underscore the value of self-attention, probabilistic modeling, and calibrated uncertainty, establishing PTTSD as a robust and interpretable framework for uncertainty-aware depression forecasting in clinical NLP.

Paper Type: Long

Research Area: Machine Learning for NLP

Research Area Keywords: multi-task learning, representation learning, structured prediction, transfer learning / domain adaptation, generalization

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data analysis

Languages Studied: English

Submission Number: 2165