Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Probabilistic Textual Time Series Depression Detection

Fabian Schmidt, Seyedehmoniba Ravan, Vladimir Vlassov

TL;DR

PTTSD introduces a probabilistic textual time series framework for predicting PHQ-8 depression scores from utterance-level transcripts, producing calibrated, temporally evolving uncertainty rather than single-point estimates. The model combines a BiLSTM with self-attention and residual connections and supports sequence-to-sequence and sequence-to-one predictions with Gaussian or Student's-$t$ output heads trained via negative log-likelihood. It achieves state-of-the-art performance among text-only systems on DAIC and E-DAIC, while providing reliable uncertainty estimates demonstrated through calibration analysis and case studies. The approach highlights the clinical value of uncertainty-aware forecasting in mental health NLP, offering a reproducible, prompt-free pipeline that leverages full transcripts. Future work may extend to multimodal signals and real-clinician validation to assess practical utility and trustworthiness.

Abstract

Accurate and interpretable predictions of depression severity are essential for clinical decision support, yet existing models often lack uncertainty estimates and temporal modeling. We propose PTTSD, a Probabilistic Textual Time Series Depression Detection framework that predicts PHQ-8 scores from utterance-level clinical interviews while modeling uncertainty over time. PTTSD includes sequence-to-sequence and sequence-to-one variants, both combining bidirectional LSTMs, self-attention, and residual connections with Gaussian or Student-t output heads trained via negative log-likelihood. Evaluated on E-DAIC and DAIC-WOZ, PTTSD achieves state-of-the-art performance among text-only systems (e.g., MAE = 3.85 on E-DAIC, 3.55 on DAIC) and produces well-calibrated prediction intervals. Ablations confirm the value of attention and probabilistic modeling, while comparisons with MentalBERT establish generality. A three-part calibration analysis and qualitative case studies further highlight the interpretability and clinical relevance of uncertainty-aware forecasting.

Probabilistic Textual Time Series Depression Detection

TL;DR

PTTSD introduces a probabilistic textual time series framework for predicting PHQ-8 depression scores from utterance-level transcripts, producing calibrated, temporally evolving uncertainty rather than single-point estimates. The model combines a BiLSTM with self-attention and residual connections and supports sequence-to-sequence and sequence-to-one predictions with Gaussian or Student's- output heads trained via negative log-likelihood. It achieves state-of-the-art performance among text-only systems on DAIC and E-DAIC, while providing reliable uncertainty estimates demonstrated through calibration analysis and case studies. The approach highlights the clinical value of uncertainty-aware forecasting in mental health NLP, offering a reproducible, prompt-free pipeline that leverages full transcripts. Future work may extend to multimodal signals and real-clinician validation to assess practical utility and trustworthiness.

Abstract

Accurate and interpretable predictions of depression severity are essential for clinical decision support, yet existing models often lack uncertainty estimates and temporal modeling. We propose PTTSD, a Probabilistic Textual Time Series Depression Detection framework that predicts PHQ-8 scores from utterance-level clinical interviews while modeling uncertainty over time. PTTSD includes sequence-to-sequence and sequence-to-one variants, both combining bidirectional LSTMs, self-attention, and residual connections with Gaussian or Student-t output heads trained via negative log-likelihood. Evaluated on E-DAIC and DAIC-WOZ, PTTSD achieves state-of-the-art performance among text-only systems (e.g., MAE = 3.85 on E-DAIC, 3.55 on DAIC) and produces well-calibrated prediction intervals. Ablations confirm the value of attention and probabilistic modeling, while comparisons with MentalBERT establish generality. A three-part calibration analysis and qualitative case studies further highlight the interpretability and clinical relevance of uncertainty-aware forecasting.

Paper Structure

This paper contains 44 sections, 14 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Probabilistic Textual Time Series Depression Detection
  4. Problem Formulation
  5. Data and Preprocessing
  6. Generating Utterance Embeddings
  7. Probabilistic LSTM Architecture
  8. Sequence-to-Sequence.
  9. Sequence-to-One.
  10. Sequence Modeling and Predictive Distributions
  11. Gaussian distribution.
  12. Student's $\boldsymbol{t}$-distribution.
  13. Loss Functions
  14. Training Procedure
  15. Experiments
  16. ...and 29 more sections

Figures (8)

  • Figure 1: Probabilistic Textual Time Series Depression Detection
  • Figure 2: Ablation results
  • Figure 3: Calibration analysis of PTTSD seq2seq on DAIC test set (Gaussian NLL vs. MSE)
  • Figure 4: Case studies illustrating prediction, ground truth, and model uncertainty for three representative participants. Each subplot shows the temporal trajectory of predictions (blue), ground truth (dashed green), and uncertainty intervals (light blue area) across timesteps.
  • Figure 5: Uncertainty distributions by severity group
  • ...and 3 more figures