Skip to content

Findings and Visualizations

Summary of Experimental Results

This page summarizes the key quantitative findings and qualitative observations from the dual-model framework for depression and anxiety severity prediction.

The evaluation is performed using a stratified 70 percent training split and 30 percent test split to preserve severity-level proportions.


Model Performance

  • Depression Severity Model

Test accuracy: 99.87 percent
Near-perfect precision, recall, and F1-score across severity levels.

  • Anxiety Severity Model

Test accuracy: 98.93 percent
Strong performance overall with small reductions in higher severity categories.

  • Discriminative Power

AUC results confirm strong separation across classes.
Depression remains near-perfect. Anxiety shows a minor reduction in higher classes.


Classification Report

Depression Prediction

Class Precision Recall F1-score Support
0 1.00 1.00 1.00 256
1 1.00 1.00 1.00 203
2 1.00 1.00 1.00 166
3 0.99 1.00 1.00 125
Macro Avg 1.00 1.00 1.00 750
Weighted Avg 1.00 1.00 1.00 750

Anxiety Prediction

Class Precision Recall F1-score Support
0 1.00 1.00 1.00 260
1 1.00 1.00 1.00 204
2 0.97 0.98 0.98 170
3 0.97 0.96 0.97 116
Macro Avg 1.00 1.00 1.00 750
Weighted Avg 1.00 1.00 1.00 750

AUC Comparison

Class AUC Depression AUC Anxiety
0 (None) 1.00 1.00
1 (Mild) 1.00 1.00
2 (Moderate) 1.00 0.98
3 (Severe) 0.99 0.97
Macro Avg 1.00 0.99
Weighted Avg 1.00 0.99

Embedding Space Visualizations

  • Binary Separation Views

Two-dimensional embedding projections show clustering patterns for severe versus none classes for both depression and anxiety.

  • UMAP Projections

UMAP projections display a graded continuum across all four severity levels, with partial overlap expected for continuous constructs.

Add your figures here:

  • Figure 1 Embedding space for depression, none vs severe
  • Figure 2 Embedding space for anxiety, none vs severe
  • Figure 3 UMAP projection for depression severity
  • Figure 4 UMAP projection for anxiety severity

Linguistic Pattern Observations

Word-frequency analysis reveals severity-associated language shifts:

  • Lower severity levels show more neutral, routine daily language.
  • Moderate and severe levels show increased negative and self-focused vocabulary.
  • Anxiety shows stronger patterns of fear, worry, and physiological stress language at higher severity.

Add your tables here:

  • Table 1 Depression word frequency by severity
  • Table 2 Anxiety word frequency by severity

Key Observations

  • Errors Are Usually Adjacent

Misclassifications occur primarily between neighboring severity categories such as moderate and severe.

  • Higher Severity Anxiety Is Harder

Anxiety shows slightly reduced recall in severe class, consistent with overlap in language across high severity states.

  • High-Risk Separation Is Strong

Severe cases are rarely misclassified as none, supporting reliability for high-risk identification.


Example Outputs

The following structure can be used to present example predictions from the prototype:

  • Input

Arabic text entry

  • Output

Depression score and anxiety score from 0 to 3

  • Alerts

High score, worsening trend, or sudden spike if history is provided

Include only synthetic or fully anonymized examples.


Connection to Alert Framework

The predictive alert mechanism converts severity outputs into longitudinal risk signals by detecting:

  • sustained high severity
  • gradual worsening trends
  • sudden spikes

For details, refer to the alert framework section in the paper and the deployment documentation.