Abstract
Predicting suicidal ideation is a critical challenge in mental health, often complicated by individuals concealing their emotional distress. Traditional deep learning models struggle with the "natural noise" inherent in such weakly annotated physiological datasets. This paper introduces TNANet (Temporal-Noise-Aware Neural Network), which combines self-supervised feature encoding via Deep Belief Networks (DBNs) with a two-stage confidence learning (CL) process. TNANet achieves a significant performance boost, including a 10% accuracy improvement over state-of-the-art baselines on specialized prisoner PPG data.
Introduction
The prediction of suicidal ideation, particularly in high-risk environments like prisons, remains a difficult task. While physiological signals like photoplethysmography (PPG) provide objective data on the autonomic nervous system, they are frequently "noisy." In this context, noise refers to the discrepancy between a person's internal state and their physiological expression, often due to intentional concealment. Existing models often fail to account for this "natural noise," leading to poor generalizability.
The TNANet Architecture
TNANet addresses these challenges through a multi-layered architecture designed for robust feature extraction and noise handling.
The architecture consists of three primary modules:
- Deep Belief Network (DBN) Module: This module utilizes Restricted Boltzmann Machines (RBMs) for self-supervised feature encoding. By learning the underlying structure of PPG features without labels, the model reduces redundancy and creates a stable representation of physiological data.
- Convolutional Module: Following encoding, 2D filters are applied to learn the relative importance of different features and reduce dimensionality.
- Classification Module: A final softmax layer produces the prediction probabilities for suicidal ideation.
Two-Stage Confidence Learning
The core innovation of TNANet is its two-stage "confidence learning" (CL) framework, which specifically targets label noise.
- Stage 1: The model is initially trained on the entire noisy dataset to establish baseline predictions.
- Stage 2: Using the Pruning by Noise Rank (PBNR) method, the model identifies samples where the predicted probability and the provided label are highly discordant. These "noisy" samples—likely representing individuals concealing ideation—are removed. The model is then re-trained on this "purified" dataset, leading to sharper, more accurate classification.
Experimental Results
The model was evaluated on a unique dataset of PPG signals from 2,190 male prisoners, as well as three public EEG benchmarks with artificial label noise.
- Performance: TNANet achieved 63.33% accuracy on the prisoner dataset, outperforming SOTA models like EEGNet.
- Robustness: Across all datasets, the model showed an average accuracy improvement of 9.00% and an F1-score improvement of 9.34% over baseline methods.
- Clinical Relevance: The model identified key heart rate variability markers, such as pNN20 and SDNN, as top predictors, validating its findings against established medical literature.
Conclusion
TNANet demonstrates that identifying and pruning "natural noise" is more effective than simply increasing model complexity for mental health applications. While currently focused on male prisoner demographics, future work will explore multimodal integration and real-time monitoring to enhance the model's clinical utility across broader populations.
Summary updated