TNANet: A Temporal-Noise-Aware Neural Network for Suicidal Ideation Prediction with Noisy Physiological Data

TL;DR

TNANet tackles the challenge of noisy physiological labels in suicidal ideation prediction by integrating a self-supervised DBN encoder with a supervised CNN and a confidence-learning-based noise filtration scheme. The model is evaluated on a novel PPG dataset collected from prisoners and on three public EEG datasets, showing superior accuracy and F1 across benchmarks, including a 63.33% accuracy on the PPG data. A two-stage training pipeline first identifies noisy samples via confidence learning and then retrains on a purified set, yielding robust performance gains. The approach reveals physiologically meaningful indicators such as HRV metrics and PPG amplitude, suggesting practical value for continuous mental-health monitoring in real-world, noisy-label environments.

Abstract

The robust generalization of deep learning models in the presence of inherent noise remains a significant challenge, especially when labels are subjective and noise is indiscernible in natural settings. This problem is particularly pronounced in many practical applications. In this paper, we address a special and important scenario of monitoring suicidal ideation, where time-series data, such as photoplethysmography (PPG), is susceptible to such noise. Current methods predominantly focus on image and text data or address artificially introduced noise, neglecting the complexities of natural noise in time-series analysis. To tackle this, we introduce a novel neural network model tailored for analyzing noisy physiological time-series data, named TNANet, which merges advanced encoding techniques with confidence learning, enhancing prediction accuracy. Another contribution of our work is the collection of a specialized dataset of PPG signals derived from real-world environments for suicidal ideation prediction. Employing this dataset, our TNANet achieves the prediction accuracy of 63.33% in a binary classification task, outperforming state-of-the-art models. Furthermore, comprehensive evaluations were conducted on three other well-known public datasets with artificially introduced noise to rigorously test the TNANet's capabilities. These tests consistently demonstrated TNANet's superior performance by achieving an accuracy improvement of more than 10% compared to baseline methods.

Figures (3)

• Figure 1: The architecture of our proposed TNANet model. It is composed of a DBN module, a convolution module, and a classification module. The input layer of the model is composed of $D^m$ vectors with a length $T$, where $D^m$ denotes the number of extracted features and $T$ represents the number of temporal points of each feature vector. The DBN module, which is constituted by RBM-1 and RBM-2, consists of one $T$-unit input layer ($V^{[1]}$), one 50-unit hidden layer ($H^{[1]}$), and one 25-unit output layer ($H^{[2]}$), with $H_2$ denoting the exact units amount. For each feature extracted from the preprocessed PPG data, the windows are formatted into a $T$-dim vector as the DBN input $X$. F is the number of filters (set to 16), P$=\min({H_2}//4, 8)$ is the pooling size in the second block of the convolution module. $\vec{Y^*}_1$ and $\vec{Y^*}_0$ are the predicted probabilities of being positive and negative, respectively.
• Figure 2: The training pipeline of TNANet. The purple arrows denote the first cross-validation stage and CL stage of TNANet, while the pink arrows denote the second cross-validation stage.
• Figure 3: Heartbeats detection in photoplethysmogram (PPG) signals.